Daily Papers

Legal reasoning is not semantic similarity search. A court judgment encodes constrained symbolic reasoning: precedent propagation, procedural state transitions, and statute-bound inference. These are properties that vector-based retrieval-augmented generation (RAG) cannot faithfully represent. Hallucinated precedents, outdated statute citations, and unsupported reasoning chains remain persistent failure modes in LLM-based legal AI, with real consequences for access to justice in high-caseload jurisdictions such as India. This paper presents Falkor-IRAC, a graph-constrained generation framework for Indian legal AI that grounds generation in structured reasoning over an IRAC (Issue, Rule, Analysis, Conclusion) knowledge graph. Judgments from the Supreme Court and High Courts of India are ingested as IRAC node structures enriched with procedural state transitions, precedent relationships, and statutory references, stored in FalkorDB for low-latency agentic traversal. At inference time, LLM-generated answers are accepted only if a valid supporting path can be traced through the graph, a check performed by a falsifiability oracle called the Verifier Agent. The system also detects doctrinal conflicts as a first-class output rather than silently resolving them. Falkor-IRAC is evaluated using graph-native metrics: citation grounding accuracy, path validity rate, hallucinated precedent rate, and conflict detection rate. These metrics are argued to be more appropriate for legal reasoning evaluation than BLEU and ROUGE. On a proof-of-concept corpus of 51 Supreme Court judgments, the Verifier Agent correctly validated citations on completed queries and correctly rejected fabricated citations. Evaluation against vector-only RAG baselines is left for future work, as is GPU-accelerated inference to address current timeout rates on CPU hardware.

The rise of large language models (LLMs) has significantly influenced the quality of information in decision-making systems, leading to the prevalence of AI-generated content and challenges in detecting misinformation and managing conflicting information, or "inter-evidence conflicts." This study introduces a method for generating diverse, validated evidence conflicts to simulate real-world misinformation scenarios. We evaluate conflict detection methods, including Natural Language Inference (NLI) models, factual consistency (FC) models, and LLMs, on these conflicts (RQ1) and analyze LLMs' conflict resolution behaviors (RQ2). Our key findings include: (1) NLI and LLM models exhibit high precision in detecting answer conflicts, though weaker models suffer from low recall; (2) FC models struggle with lexically similar answer conflicts, while NLI and LLM models handle these better; and (3) stronger models like GPT-4 show robust performance, especially with nuanced conflicts. For conflict resolution, LLMs often favor one piece of conflicting evidence without justification and rely on internal knowledge if they have prior beliefs.

Large Multimodal Models(LMMs) face notable challenges when encountering multimodal knowledge conflicts, particularly under retrieval-augmented generation(RAG) frameworks where the contextual information from external sources may contradict the model's internal parametric knowledge, leading to unreliable outputs. However, existing benchmarks fail to reflect such realistic conflict scenarios. Most focus solely on intra-memory conflicts, while context-memory and inter-context conflicts remain largely investigated. Furthermore, commonly used factual knowledge-based evaluations are often overlooked, and existing datasets lack a thorough investigation into conflict detection capabilities. To bridge this gap, we propose MMKC-Bench, a benchmark designed to evaluate factual knowledge conflicts in both context-memory and inter-context scenarios. MMKC-Bench encompasses three types of multimodal knowledge conflicts and includes 1,573 knowledge instances and 3,381 images across 23 broad types, collected through automated pipelines with human verification. We evaluate three representative series of LMMs on both model behavior analysis and conflict detection tasks. Our findings show that while current LMMs are capable of recognizing knowledge conflicts, they tend to favor internal parametric knowledge over external evidence. We hope MMKC-Bench will foster further research in multimodal knowledge conflict and enhance the development of multimodal RAG systems. The source code is available at https://github.com/MLLMKCBENCH/MLLMKC.

Large language models (LLMs) have achieved impressive advancements across numerous disciplines, yet the critical issue of knowledge conflicts, a major source of hallucinations, has rarely been studied. Only a few research explored the conflicts between the inherent knowledge of LLMs and the retrieved contextual knowledge. However, a thorough assessment of knowledge conflict in LLMs is still missing. Motivated by this research gap, we present ConflictBank, the first comprehensive benchmark developed to systematically evaluate knowledge conflicts from three aspects: (i) conflicts encountered in retrieved knowledge, (ii) conflicts within the models' encoded knowledge, and (iii) the interplay between these conflict forms. Our investigation delves into four model families and twelve LLM instances, meticulously analyzing conflicts stemming from misinformation, temporal discrepancies, and semantic divergences. Based on our proposed novel construction framework, we create 7,453,853 claim-evidence pairs and 553,117 QA pairs. We present numerous findings on model scale, conflict causes, and conflict types. We hope our ConflictBank benchmark will help the community better understand model behavior in conflicts and develop more reliable LLMs.

Knowledge conflict arises from discrepancies between information in the context of a large language model (LLM) and the knowledge stored in its parameters. This can hurt performance when using standard decoding techniques, which tend to ignore the context. Existing test-time contrastive methods seek to address this by comparing the LLM's output distribution with and without the context and adjust the model according to the contrast between them. However, we find that these methods frequently misjudge the degree of conflict and struggle to handle instances that vary in their amount of conflict, with static methods over-adjusting when conflict is absent. We propose a fine-grained, instance-level approach called AdaCAD, which dynamically infers the weight of adjustment based on the degree of conflict, as measured by the Jensen-Shannon divergence between distributions representing contextual and parametric knowledge. Our experiments across four models on six diverse question-answering (QA) datasets and three summarization tasks demonstrate that our training-free adaptive method consistently outperforms other decoding methods on QA, with average accuracy gains of 14.21% (absolute) over a static contrastive baseline, and improves the factuality of summaries by 5.59 (AlignScore). Furthermore, our analysis shows that while decoding with contrastive baselines hurts performance when conflict is absent, AdaCAD mitigates these losses, making it more applicable to real-world datasets in which some examples have conflict and others do not.

Retrieval Augmented Generation (RAG) is a commonly used approach for enhancing large language models (LLMs) with relevant and up-to-date information. However, the retrieved sources can often contain conflicting information and it remains unclear how models should address such discrepancies. In this work, we first propose a novel taxonomy of knowledge conflict types in RAG, along with the desired model behavior for each type. We then introduce CONFLICTS, a high-quality benchmark with expert annotations of conflict types in a realistic RAG setting. CONFLICTS is the first benchmark that enables tracking progress on how models address a wide range of knowledge conflicts. We conduct extensive experiments on this benchmark, showing that LLMs often struggle to appropriately resolve conflicts between sources. While prompting LLMs to explicitly reason about the potential conflict in the retrieved documents significantly improves the quality and appropriateness of their responses, substantial room for improvement in future research remains.

Large Language Models (LLMs) have demonstrated strong performance in question answering (QA) tasks. However, Multi-Answer Question Answering (MAQA), where a question may have several valid answers, remains challenging. Traditional QA settings often assume consistency across evidences, but MAQA can involve conflicting answers. Constructing datasets that reflect such conflicts is costly and labor-intensive, while existing benchmarks often rely on synthetic data, restrict the task to yes/no questions, or apply unverified automated annotation. To advance research in this area, we extend the conflict-aware MAQA setting to require models not only to identify all valid answers, but also to detect specific conflicting answer pairs, if any. To support this task, we introduce a novel cost-effective methodology for leveraging fact-checking datasets to construct NATCONFQA, a new benchmark for realistic, conflict-aware MAQA, enriched with detailed conflict labels, for all answer pairs. We evaluate eight high-end LLMs on NATCONFQA, revealing their fragility in handling various types of conflicts and the flawed strategies they employ to resolve them.

Large language models (LLMs) equipped with retrieval--the Retrieval-Augmented Generation (RAG) paradigm--should combine their parametric knowledge with external evidence, yet in practice they often hallucinate, over-trust noisy snippets, or ignore vital context. We introduce TCR (Transparent Conflict Resolution), a plug-and-play framework that makes this decision process observable and controllable. TCR (i) disentangles semantic match and factual consistency via dual contrastive encoders, (ii) estimates self-answerability to gauge confidence in internal memory, and (iii) feeds the three scalar signals to the generator through a lightweight soft-prompt with SNR-based weighting. Across seven benchmarks TCR improves conflict detection (+5-18 F1), raises knowledge-gap recovery by +21.4 pp and cuts misleading-context overrides by -29.3 pp, while adding only 0.3% parameters. The signals align with human judgements and expose temporal decision patterns.

Understanding the origins of militarized conflict is a complex, yet important undertaking. Existing research seeks to build this understanding by considering bi-lateral relationships between entity pairs (dyadic causes) and multi-lateral relationships among multiple entities (systemic causes). The aim of this work is to compare these two causes in terms of how they correlate with conflict between two entities. We do this by devising a set of textual and graph-based features which represent each of the causes. The features are extracted from Wikipedia and modeled as a large graph. Nodes in this graph represent entities connected by labeled edges representing ally or enemy-relationships. This allows casting the problem as an edge classification task, which we term dyad classification. We propose and evaluate classifiers to determine if a particular pair of entities are allies or enemies. Our results suggest that our systemic features might be slightly better correlates of conflict. Further, we find that Wikipedia articles of allies are semantically more similar than enemies.

Large language models (LLMs) often need to balance their internal parametric knowledge with external information, such as user beliefs and content from retrieved documents, in real-world scenarios like RAG or chat-based systems. A model's ability to reliably process these sources is key to system safety. Previous studies on knowledge conflict and sycophancy are limited to a binary conflict paradigm, primarily exploring conflicts between parametric knowledge and either a document or a user, but ignoring the interactive environment where all three sources exist simultaneously. To fill this gap, we propose a three-source interaction framework and systematically evaluate 27 LLMs from 3 families on 2 datasets. Our findings reveal general patterns: most models rely more on document assertions than user assertions, and this preference is reinforced by post-training. Furthermore, our behavioral analysis shows that most models are impressionable, unable to effectively discriminate between helpful and harmful external information. To address this, we demonstrate that fine-tuning on diverse source interaction data can significantly increase a model's discrimination abilities. In short, our work paves the way for developing trustworthy LLMs that can effectively and reliably integrate multiple sources of information. Code is available at https://github.com/shuowl/llm-source-balancing.

The Israeli-Palestinian conflict started on 7 October 2023, have resulted thus far to over 48,000 people killed including more than 17,000 children with a majority from Gaza, more than 30,000 people injured, over 10,000 missing, and over 1 million people displaced, fleeing conflict zones. The infrastructure damage includes the 87\% of housing units, 80\% of public buildings and 60\% of cropland 17 out of 36 hospitals, 68\% of road networks and 87\% of school buildings damaged. This conflict has as well launched an online discussion across various social media platforms. Telegram was no exception due to its encrypted communication and highly involved audience. The current study will cover an analysis of the related discussion in relation to different participants of the conflict and sentiment represented in those discussion. To this end, we prepared a dataset of 125K messages shared on channels in Telegram spanning from 23 October 2025 until today. Additionally, we apply the same analysis in two publicly available datasets from Twitter containing 2001 tweets and from Reddit containing 2M opinions. We apply a volume analysis across the three datasets, entity extraction and then proceed to BERT topic analysis in order to extract common themes or topics. Next, we apply sentiment analysis to analyze the emotional tone of the discussions. Our findings hint at polarized narratives as the hallmark of how political factions and outsiders mold public opinion. We also analyze the sentiment-topic prevalence relationship, detailing the trends that may show manipulation and attempts of propaganda by the involved parties. This will give a better understanding of the online discourse on the Israel-Palestine conflict and contribute to the knowledge on the dynamics of social media communication during geopolitical crises.

Hypernetwork-based methods such as Doc-to-LoRA internalize a document into an LLM's weights in a single forward pass, but they fail systematically on conflicts: when the document contradicts pretraining knowledge, accuracy collapses to 46.4% on the deepest facts. We show the failure is a magnitude problem rather than a representational one. The hypernetwork already targets the right layers, but its adapter margin is approximately constant across documents while the pretrained margin grows with training frequency, so deep conflicts lose by construction. The account predicts that failure should track prior strength: sorting 194 conflicts by the base model's log-probability on the contradicted fact, baseline accuracy falls from 68% on weak-prior questions to 16% on strong-prior ones, a 52 percentage-point gap. The cure is amplitude. Selective Layer Boosting scales the adapter at its top-norm layers, and Conflict-Aware Internalization triggers boosting only when the base model is confident. Both are training-free; together they raise deep-conflict accuracy from 46.4% to 71.0% on Gemma-2B and from 53.6% to 72.5% on Mistral-7B while preserving novel-knowledge recall, and beat vanilla retrieval-augmented generation on medium conflicts by 18 percentage points despite operating entirely in parameter space. We release KID-Bench, a 489-question benchmark that separates novel recall, cross-knowledge combination, and prior-graded conflicts.

The proliferation of large language models has raised growing concerns about their misuse, particularly in cases where AI-generated text is falsely attributed to human authors. Machine-generated content detectors claim to effectively identify such text under various conditions and from any language model. This paper critically evaluates these claims by assessing several popular detectors (RADAR, Wild, T5Sentinel, Fast-DetectGPT, PHD, LogRank, Binoculars) on a range of domains, datasets, and models that these detectors have not previously encountered. We employ various prompting strategies to simulate practical adversarial attacks, demonstrating that even moderate efforts can significantly evade detection. We emphasize the importance of the true positive rate at a specific false positive rate (TPR@FPR) metric and demonstrate that these detectors perform poorly in certain settings, with TPR@.01 as low as 0%. Our findings suggest that both trained and zero-shot detectors struggle to maintain high sensitivity while achieving a reasonable true positive rate.

When audio and text conflict, speech-enabled language models follow text far more often than they do when arbitrating between two conflicting text sources, even under explicit instructions to trust the audio. We introduce ALME (Audio-LLM Modality Evaluation), a dataset of 57,602 controlled audio-text conflict stimuli across eight languages, together with Text Dominance Ratio (TDR), which measures how often a model follows conflicting text when instructed to follow audio. Gemini 2.0 Flash and GPT-4o show TDR 10--26times higher than a baseline that replaces audio with its transcript under otherwise identical conditions (Gemini 2.0 Flash: 16.6% vs. 1.6%; GPT-4o: 23.2% vs. 0.9%). These results suggest that text dominance reflects not only information content, but also an asymmetry in arbitration accessibility, i.e., how easily the model can use competing representations at decision time. Framing the transcript as deliberately corrupted reduces TDR by 80%, whereas forcing explicit transcription increases it by 14%. A fine-tuning ablation further suggests that arbitration behavior depends more on LLM reasoning than on the audio input path alone. Across four audio-LLMs, we observe the same qualitative pattern with substantial cross-model and cross-linguistic variation.

Determining when people are struggling allows for a finer-grained understanding of actions that complements conventional action classification and error detection. Struggle detection, as defined in this paper, is a distinct and important task that can be identified without explicit step or activity knowledge. We introduce the first struggle dataset with three real-world problem-solving activities that are labelled by both expert and crowd-source annotators. Video segments were scored w.r.t. their level of struggle using a forced choice 4-point scale. This dataset contains 5.1 hours of video from 73 participants. We conducted a series of experiments to identify the most suitable modelling approaches for struggle determination. Additionally, we compared various deep learning models, establishing baseline results for struggle classification, struggle regression, and struggle label distribution learning. Our results indicate that struggle detection in video can achieve up to 88.24% accuracy in binary classification, while detecting the level of struggle in a four-way classification setting performs lower, with an overall accuracy of 52.45%. Our work is motivated toward a more comprehensive understanding of action in video and potentially the improvement of assistive systems that analyse struggle and can better support users during manual activities.

Conflict scholars have used rule-based approaches to extract information about political violence from news reports and texts. Recent Natural Language Processing developments move beyond rigid rule-based approaches. We review our recent ConfliBERT language model (Hu et al. 2022) to process political and violence related texts. The model can be used to extract actor and action classifications from texts about political conflict. When fine-tuned, results show that ConfliBERT has superior performance in accuracy, precision and recall over other large language models (LLM) like Google's Gemma 2 (9B), Meta's Llama 3.1 (7B), and Alibaba's Qwen 2.5 (14B) within its relevant domains. It is also hundreds of times faster than these more generalist LLMs. These results are illustrated using texts from the BBC, re3d, and the Global Terrorism Dataset (GTD).

We used machine learning and artificial intelligence: 1) to measure levels of peace in countries from news and social media and 2) to develop on-line tools that promote peace by helping users better understand their own media diet. For news media, we used neural networks to measure levels of peace from text embeddings of on-line news sources. The model, trained on one news media dataset also showed high accuracy when used to analyze a different news dataset. For social media, such as YouTube, we developed other models to measure levels of social dimensions important in peace using word level (GoEmotions) and context level (Large Language Model) methods. To promote peace, we note that 71% of people 20-40 years old daily view most of their news through short videos on social media. Content creators of these videos are biased towards creating videos with emotional activation, making you angry to engage you, to increase clicks. We developed and tested a Chrome extension, MirrorMirror, which provides real-time feedback to YouTube viewers about the peacefulness of the media they are watching. Our long term goal is for MirrorMirror to evolve into an open-source tool for content creators, journalists, researchers, platforms, and individual users to better understand the tone of their media creation and consumption and its effects on viewers. Moving beyond simple engagement metrics, we hope to encourage more respectful, nuanced, and informative communication.

Ambivalence and hesitancy (A/H) are subtle affective states where a person shows conflicting signals through different channels -- saying one thing while their face or voice tells another story. Recognising these states automatically is valuable in clinical settings, but it is hard for machines because the key evidence lives in the disagreements between what is said, how it sounds, and what the face shows. We present ConflictAwareAH, a multimodal framework built for this problem. Three pre-trained encoders extract video, audio, and text representations. Pairwise conflict features -- element-wise absolute differences between modality embeddings -- serve as bidirectional cues: large cross-modal differences flag A/H, while small differences confirm behavioural consistency and anchor the negative class. This conflict-aware design addresses a key limitation of text-dominant approaches, which tend to over-detect A/H (high F1-AH) while struggling to confirm its absence: our multimodal model improves F1-NoAH by +4.6 points over text alone and halves the class-performance gap. A complementary text-guided late fusion strategy blends a text-only auxiliary head with the full model at inference, adding +4.1 Macro F1. On the BAH dataset from the ABAW10 Ambivalence/Hesitancy Challenge, our method reaches 0.694 Macro F1 on the labelled test split and 0.715 on the private leaderboard, outperforming published multimodal baselines by over 10 points -- all on a single GPU in under 25 minutes of training.

Knowledge-based visual question answering (KB-VQA) demonstrates significant potential for handling knowledge-intensive tasks. However, conflicts arise between static parametric knowledge in vision language models (VLMs) and dynamically retrieved information due to the static model knowledge from pre-training. The outputs either ignore retrieved contexts or exhibit inconsistent integration with parametric knowledge, posing substantial challenges for KB-VQA. Current knowledge conflict mitigation methods primarily adapted from language-based approaches, focusing on context-level conflicts through engineered prompting strategies or context-aware decoding mechanisms. However, these methods neglect the critical role of visual information in conflicts and suffer from redundant retrieved contexts, which impair accurate conflict identification and effective mitigation. To address these limitations, we propose CC-VQA: a novel training-free, conflict- and correlation-aware method for KB-VQA. Our method comprises two core components: (1) Vision-Centric Contextual Conflict Reasoning, which performs visual-semantic conflict analysis across internal and external knowledge contexts; and (2) Correlation-Guided Encoding and Decoding, featuring positional encoding compression for low-correlation statements and adaptive decoding using correlation-weighted conflict scoring. Extensive evaluations on E-VQA, InfoSeek, and OK-VQA benchmarks demonstrate that CC-VQA achieves state-of-the-art performance, yielding absolute accuracy improvements of 3.3\% to 6.4\% compared to existing methods. Code is available at https://github.com/cqu-student/CC-VQA.

With the growing importance of detecting misinformation, many studies have focused on verifying factual claims by retrieving evidence. However, canonical fact verification tasks do not apply to catching subtle differences in factually consistent claims, which might still bias the readers, especially on contentious political or economic issues. Our underlying assumption is that among the trusted sources, one's argument is not necessarily more true than the other, requiring comparison rather than verification. In this study, we propose ClaimDiff, a novel dataset that primarily focuses on comparing the nuance between claim pairs. In ClaimDiff, we provide 2,941 annotated claim pairs from 268 news articles. We observe that while humans are capable of detecting the nuances between claims, strong baselines struggle to detect them, showing over a 19% absolute gap with the humans. We hope this initial study could help readers to gain an unbiased grasp of contentious issues through machine-aided comparison.

This paper presents LEMONADE, a large-scale conflict event dataset comprising 39,786 events across 20 languages and 171 countries, with extensive coverage of region-specific entities. LEMONADE is based on a partially reannotated subset of the Armed Conflict Location & Event Data (ACLED), which has documented global conflict events for over a decade. To address the challenge of aggregating multilingual sources for global event analysis, we introduce abstractive event extraction (AEE) and its subtask, abstractive entity linking (AEL). Unlike conventional span-based event extraction, our approach detects event arguments and entities through holistic document understanding and normalizes them across the multilingual dataset. We evaluate various large language models (LLMs) on these tasks, adapt existing zero-shot event extraction systems, and benchmark supervised models. Additionally, we introduce ZEST, a novel zero-shot retrieval-based system for AEL. Our best zero-shot system achieves an end-to-end F1 score of 58.3%, with LLMs outperforming specialized event extraction models such as GoLLIE. For entity linking, ZEST achieves an F1 score of 45.7%, significantly surpassing OneNet, a state-of-the-art zero-shot baseline that achieves only 23.7%. However, these zero-shot results lag behind the best supervised systems by 20.1% and 37.0% in the end-to-end and AEL tasks, respectively, highlighting the need for further research.

Do the Spratly Islands belong to China, the Philippines, or Vietnam? A pretrained large language model (LLM) may answer differently if asked in the languages of each claimant country: Chinese, Tagalog, or Vietnamese. This contrasts with a multilingual human, who would likely answer consistently. In this paper, we show that LLMs recall certain geographical knowledge inconsistently when queried in different languages -- a phenomenon we term geopolitical bias. As a targeted case study, we consider territorial disputes, an inherently controversial and multilingual task. We introduce BorderLines, a dataset of territorial disputes which covers 251 territories, each associated with a set of multiple-choice questions in the languages of each claimant country (49 languages in total). We also propose a suite of evaluation metrics to precisely quantify bias and consistency in responses across different languages. We then evaluate various multilingual LLMs on our dataset and metrics to probe their internal knowledge and use the proposed metrics to discover numerous inconsistencies in how these models respond in different languages. Finally, we explore several prompt modification strategies, aiming to either amplify or mitigate geopolitical bias, which highlights how brittle LLMs are and how they tailor their responses depending on cues from the interaction context. Our code and data are available at https://github.com/manestay/borderlines

Large language models (LLMs) can store a significant amount of factual knowledge in their parameters. However, their parametric knowledge may conflict with the information provided in the context -- this phenomenon, known as context-memory knowledge conflicts, can lead to undesirable model behaviour, such as reliance on outdated or incorrect information. Analysing the internal activations of LLMs, we find that they can internally register the signals of knowledge conflict at mid-layers. Such signals allow us to detect whether a knowledge conflict occurs and use inference-time intervention strategies to resolve it. In this work, we propose SpARE, a training-free representation engineering method that uses pre-trained sparse auto-encoders (SAEs) to control the knowledge selection behaviour of LLMs. SpARE identifies the functional features that control the knowledge selection behaviours and applies them to edit the internal activations of LLMs at inference time. Our experimental results show that SpARE can effectively control the usage of either knowledge source to resolve knowledge conflict in open-domain question-answering tasks, surpassing existing representation engineering methods (+10%) as well as contrastive decoding methods (+15%).

Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within innocuous requests. Each item is constructed via a multi-stage pipeline that combines LLM-assisted discovery, automated validation gates, and dual independent native-speaker annotators per country. To distinguish principled refusal from comprehension failure, we evaluate Attack Success Rate (ASR) alongside two complementary metrics we introduce: Neutral-Safe Rate (NSR) and Cultural Sensitivity Rate (CSR). Evaluating 10 frontier and 27 local LLMs reveals two key findings. First, jailbreak robustness and cultural awareness do not show a coupled relationship among frontier models, so a composite safety score obscures per-axis variation. Second, local models exhibit a near-linear ASR-NSR trade-off (r = -0.81), indicating that their apparent safety reflects generation failure rather than genuine alignment. XL-SafetyBench enables more nuanced, cross-cultural safety evaluation in the multilingual era.

Can we avoid wars at the crossroads of history? This question has been pursued by individuals, scholars, policymakers, and organizations throughout human history. In this research, we attempt to answer the question based on the recent advances of Artificial Intelligence (AI) and Large Language Models (LLMs). We propose WarAgent, an LLM-powered multi-agent AI system, to simulate the participating countries, their decisions, and the consequences, in historical international conflicts, including the World War I (WWI), the World War II (WWII), and the Warring States Period (WSP) in Ancient China. By evaluating the simulation effectiveness, we examine the advancements and limitations of cutting-edge AI systems' abilities in studying complex collective human behaviors such as international conflicts under diverse settings. In these simulations, the emergent interactions among agents also offer a novel perspective for examining the triggers and conditions that lead to war. Our findings offer data-driven and AI-augmented insights that can redefine how we approach conflict resolution and peacekeeping strategies. The implications stretch beyond historical analysis, offering a blueprint for using AI to understand human history and possibly prevent future international conflicts. Code and data are available at https://github.com/agiresearch/WarAgent.

Large language models should follow hierarchical instructions where system prompts override user inputs, yet recent work shows they often ignore this rule while strongly obeying social cues such as authority or consensus. We extend these behavioral findings with mechanistic interpretations on a large-scale dataset. Linear probing shows conflict-decision signals are encoded early, with system-user and social conflicts forming distinct subspaces. Direct Logit Attribution reveals stronger internal conflict detection in system-user cases but consistent resolution only for social cues. Steering experiments show that, despite using social cues, the vectors surprisingly amplify instruction following in a role-agnostic way. Together, these results explain fragile system obedience and underscore the need for lightweight hierarchy-sensitive alignment methods.

Role-Playing Agents (RPAs) have shown remarkable performance in various applications, yet they often struggle to recognize and appropriately respond to hard queries that conflict with their role-play knowledge. To investigate RPAs' performance when faced with different types of conflicting requests, we develop an evaluation benchmark that includes contextual knowledge conflicting requests, parametric knowledge conflicting requests, and non-conflicting requests to assess RPAs' ability to identify conflicts and refuse to answer appropriately without over-refusing. Through extensive evaluation, we find that most RPAs behave significant performance gaps toward different conflict requests. To elucidate the reasons, we conduct an in-depth representation-level analysis of RPAs under various conflict scenarios. Our findings reveal the existence of rejection regions and direct response regions within the model's forwarding representation, and thus influence the RPA's final response behavior. Therefore, we introduce a lightweight representation editing approach that conveniently shifts conflicting requests to the rejection region, thereby enhancing the model's refusal accuracy. The experimental results validate the effectiveness of our editing method, improving RPAs' refusal ability of conflicting requests while maintaining their general role-playing capabilities.

Most existing retrieval-augmented language models (LMs) assume a naive dichotomy within a retrieved document set: query-relevance and irrelevance. Our work investigates a more challenging scenario in which even the "relevant" documents may contain misleading or incorrect information, causing conflict among the retrieved documents and thereby negatively influencing model decisions as noise. We observe that existing LMs are highly brittle to the presence of conflicting information in both the fine-tuning and in-context few-shot learning scenarios. We propose approaches for handling knowledge conflicts among retrieved documents by explicitly fine-tuning a discriminator or prompting GPT-3.5 to elicit its discriminative capability. Our empirical results on open-domain QA show that these approaches significantly enhance model robustness. We also provide our findings on incorporating the fine-tuned discriminator's decision into the in-context learning process, proposing a way to exploit the benefits of two disparate learning schemes. Alongside our findings, we provide MacNoise, a machine-generated, conflict-induced dataset to further encourage research in this direction.

Warning: this paper contains content that may be offensive or upsetting Hate speech moderation on global platforms poses unique challenges due to the multimodal and multilingual nature of content, along with the varying cultural perceptions. How well do current vision-language models (VLMs) navigate these nuances? To investigate this, we create the first multimodal and multilingual parallel hate speech dataset, annotated by a multicultural set of annotators, called Multi3Hate. It contains 300 parallel meme samples across 5 languages: English, German, Spanish, Hindi, and Mandarin. We demonstrate that cultural background significantly affects multimodal hate speech annotation in our dataset. The average pairwise agreement among countries is just 74%, significantly lower than that of randomly selected annotator groups. Our qualitative analysis indicates that the lowest pairwise label agreement-only 67% between the USA and India-can be attributed to cultural factors. We then conduct experiments with 5 large VLMs in a zero-shot setting, finding that these models align more closely with annotations from the US than with those from other cultures, even when the memes and prompts are presented in the dominant language of the other culture. Code and dataset are available at https://github.com/MinhDucBui/Multi3Hate.

We introduce AutoMonitor-Bench, the first benchmark designed to systematically evaluate the reliability of LLM-based misbehavior monitors across diverse tasks and failure modes. AutoMonitor-Bench consists of 3,010 carefully annotated test samples spanning question answering, code generation, and reasoning, with paired misbehavior and benign instances. We evaluate monitors using two complementary metrics: Miss Rate (MR) and False Alarm Rate (FAR), capturing failures to detect misbehavior and oversensitivity to benign behavior, respectively. Evaluating 12 proprietary and 10 open-source LLMs, we observe substantial variability in monitoring performance and a consistent trade-off between MR and FAR, revealing an inherent safety-utility tension. To further explore the limits of monitor reliability, we construct a large-scale training corpus of 153,581 samples and fine-tune Qwen3-4B-Instruction to investigate whether training on known, relatively easy-to-construct misbehavior datasets improves monitoring performance on unseen and more implicit misbehaviors. Our results highlight the challenges of reliable, scalable misbehavior monitoring and motivate future work on task-aware designing and training strategies for LLM-based monitors.

In arena-style evaluation of large language models (LLMs), two LLMs respond to a user query, and the user chooses the winning response or deems the "battle" a draw, resulting in an adjustment to the ratings of both models. The prevailing approach for modeling these rating dynamics is to view battles as two-player game matches, as in chess, and apply the Elo rating system and its derivatives. In this paper, we critically examine this paradigm. Specifically, we question whether a draw genuinely means that the two models are equal and hence whether their ratings should be equalized. Instead, we conjecture that draws are more indicative of query difficulty: if the query is too easy, then both models are more likely to succeed equally. On three real-world arena datasets, we show that ignoring rating updates for draws yields a 1-3% relative increase in battle outcome prediction accuracy (which includes draws) for all four rating systems studied. Further analyses suggest that draws occur more for queries rated as very easy and those as highly objective, with risk ratios of 1.37 and 1.35, respectively. We recommend future rating systems to reconsider existing draw semantics and to account for query properties in rating updates.

It has been observed that large-scale language models capture undesirable societal biases, e.g. relating to race and gender; yet religious bias has been relatively unexplored. We demonstrate that GPT-3, a state-of-the-art contextual language model, captures persistent Muslim-violence bias. We probe GPT-3 in various ways, including prompt completion, analogical reasoning, and story generation, to understand this anti-Muslim bias, demonstrating that it appears consistently and creatively in different uses of the model and that it is severe even compared to biases about other religious groups. For instance, "Muslim" is analogized to "terrorist" in 23% of test cases, while "Jewish" is mapped to "money" in 5% of test cases. We quantify the positive distraction needed to overcome this bias with adversarial text prompts, and find that use of the most positive 6 adjectives reduces violent completions for "Muslims" from 66% to 20%, but which is still higher than for other religious groups.

Retrieval-augmented language models are being increasingly tasked with subjective, contentious, and conflicting queries such as "is aspartame linked to cancer". To resolve these ambiguous queries, one must search through a large range of websites and consider "which, if any, of this evidence do I find convincing?". In this work, we study how LLMs answer this question. In particular, we construct ConflictingQA, a dataset that pairs controversial queries with a series of real-world evidence documents that contain different facts (e.g., quantitative results), argument styles (e.g., appeals to authority), and answers (Yes or No). We use this dataset to perform sensitivity and counterfactual analyses to explore which text features most affect LLM predictions. Overall, we find that current models rely heavily on the relevance of a website to the query, while largely ignoring stylistic features that humans find important such as whether a text contains scientific references or is written with a neutral tone. Taken together, these results highlight the importance of RAG corpus quality (e.g., the need to filter misinformation), and possibly even a shift in how LLMs are trained to better align with human judgements.

Expression of sentiment in parliamentary debates is deemed to be significantly different from that on social media or in product reviews. This paper adds to an emerging body of research on parliamentary debates with a dataset of sentences annotated for detection sentiment polarity in political discourse. We sample the sentences for annotation from the proceedings of three Southeast European parliaments: Croatia, Bosnia-Herzegovina, and Serbia. A six-level schema is applied to the data with the aim of training a classification model for the detection of sentiment in parliamentary proceedings. Krippendorff's alpha measuring the inter-annotator agreement ranges from 0.6 for the six-level annotation schema to 0.75 for the three-level schema and 0.83 for the two-level schema. Our initial experiments on the dataset show that transformer models perform significantly better than those using a simpler architecture. Furthermore, regardless of the similarity of the three languages, we observe differences in performance across different languages. Performing parliament-specific training and evaluation shows that the main reason for the differing performance between parliaments seems to be the different complexity of the automatic classification task, which is not observable in annotator performance. Language distance does not seem to play any role neither in annotator nor in automatic classification performance. We release the dataset and the best-performing model under permissive licences.

Overlapping calendar invitations force busy professionals to repeatedly decide which meetings to attend, reschedule, or decline. We refer to this preference-driven decision process as calendar conflict resolution. Automating this decision process is crucial yet challenging. Scheduling logistics can drain hours, and human delegation often fails at scale, which motivates us to ask: Can we trust large language models (LLMs) or language agents to manage time? To enable a systematic study of this question, we introduce CalConflictBench, a benchmark for long-horizon calendar conflict resolution. In CalConflictBench, conflicts are presented to agents round-by-round over a calendar year, requiring them to infer and adapt to user preferences progressively. Our experiments show that current LLM agents perform poorly with high error rates, e.g., Qwen-3-30B-Think has an average error rate of 35%. To address this gap, we propose PEARL, a reinforcement-learning framework that (i) augments the language agent with an external preference memory that stores and updates inferred strategies (e.g., attendee priorities, topic importance, time/location preferences), and (ii) optimizes the agent with round-wise rewards that directly supervise decision correctness, ranking quality, and memory usage across rounds. Experiments on CalConflictBench show that PEARL achieves an error reduction rate of 0.76 and a 55% improvement in average error rate compared to the strongest baseline.

Recent advancements in conversational AI have created an urgent need for safety guardrails that prevent users from being exposed to offensive and dangerous content. Much of this work relies on human ratings and feedback, but does not account for the fact that perceptions of offense and safety are inherently subjective and that there may be systematic disagreements between raters that align with their socio-demographic identities. Instead, current machine learning approaches largely ignore rater subjectivity and use gold standards that obscure disagreements (e.g., through majority voting). In order to better understand the socio-cultural leanings of such tasks, we propose a comprehensive disagreement analysis framework to measure systematic diversity in perspectives among different rater subgroups. We then demonstrate its utility by applying this framework to a dataset of human-chatbot conversations rated by a demographically diverse pool of raters. Our analysis reveals specific rater groups that have more diverse perspectives than the rest, and informs demographic axes that are crucial to consider for safety annotations.

Large language models frequently rely on both contextual input and parametric knowledge to perform tasks. However, these sources can come into conflict, especially when retrieved documents contradict the model's parametric knowledge. We propose a diagnostic framework to systematically evaluate LLM behavior under context-memory conflict, where the contextual information diverges from their parametric beliefs. We construct diagnostic data that elicit these conflicts and analyze model performance across multiple task types. Our findings reveal that (1) knowledge conflict has minimal impact on tasks that do not require knowledge utilization, (2) model performance is consistently higher when contextual and parametric knowledge are aligned, (3) models are unable to fully suppress their internal knowledge even when instructed, and (4) providing rationales that explain the conflict increases reliance on contexts. These insights raise concerns about the validity of model-based evaluation and underscore the need to account for knowledge conflict in the deployment of LLMs.

Battles, or side-by-side comparisons in so called arenas that elicit human preferences, have emerged as a popular approach to assessing the output quality of LLMs. Recently, this idea has been extended to retrieval-augmented generation (RAG) systems. While undoubtedly representing an advance in evaluation, battles have at least two drawbacks, particularly in the context of complex information-seeking queries: they are neither explanatory nor diagnostic. Recently, the nugget evaluation methodology has emerged as a promising approach to evaluate the quality of RAG answers. Nuggets decompose long-form LLM-generated answers into atomic facts, highlighting important pieces of information necessary in a "good" response. In this work, we apply our AutoNuggetizer framework to analyze data from roughly 7K Search Arena battles provided by LMArena in a fully automatic manner. Our results show a significant correlation between nugget scores and human preferences, showcasing promise in our approach to explainable and diagnostic system evaluations.

Large language models (LLMs) can store a significant amount of factual knowledge in their parameters. However, their parametric knowledge may conflict with the information provided in the context. Such conflicts can lead to undesirable model behaviour, such as reliance on outdated or incorrect information. In this work, we investigate whether LLMs can identify knowledge conflicts and whether it is possible to know which source of knowledge the model will rely on by analysing the residual stream of the LLM. Through probing tasks, we find that LLMs can internally register the signal of knowledge conflict in the residual stream, which can be accurately detected by probing the intermediate model activations. This allows us to detect conflicts within the residual stream before generating the answers without modifying the input or model parameters. Moreover, we find that the residual stream shows significantly different patterns when the model relies on contextual knowledge versus parametric knowledge to resolve conflicts. This pattern can be employed to estimate the behaviour of LLMs when conflict happens and prevent unexpected answers before producing the answers. Our analysis offers insights into how LLMs internally manage knowledge conflicts and provides a foundation for developing methods to control the knowledge selection processes.

Knowledge-intensive language understanding tasks require Language Models (LMs) to integrate relevant context, mitigating their inherent weaknesses, such as incomplete or outdated knowledge. Nevertheless, studies indicate that LMs often ignore the provided context as it can conflict with the pre-existing LM's memory learned during pre-training. Moreover, conflicting knowledge can already be present in the LM's parameters, termed intra-memory conflict. Existing works have studied the two types of knowledge conflicts only in isolation. We conjecture that the (degree of) intra-memory conflicts can in turn affect LM's handling of context-memory conflicts. To study this, we introduce the DYNAMICQA dataset, which includes facts with a temporal dynamic nature where a fact can change with a varying time frequency and disputable dynamic facts, which can change depending on the viewpoint. DYNAMICQA is the first to include real-world knowledge conflicts and provide context to study the link between the different types of knowledge conflicts. With the proposed dataset, we assess the use of uncertainty for measuring the intra-memory conflict and introduce a novel Coherent Persuasion (CP) score to evaluate the context's ability to sway LM's semantic output. Our extensive experiments reveal that static facts, which are unlikely to change, are more easily updated with additional context, relative to temporal and disputable facts.

Multimodal AI models are increasingly used in fields like healthcare, finance, and autonomous driving, where information is drawn from multiple sources or modalities such as images, texts, audios, videos. However, effectively managing uncertainty - arising from noise, insufficient evidence, or conflicts between modalities - is crucial for reliable decision-making. Current uncertainty-aware machine learning methods leveraging, for example, evidence averaging, or evidence accumulation underestimate uncertainties in high-conflict scenarios. Moreover, the state-of-the-art evidence averaging strategy is not order invariant and fails to scale to multiple modalities. To address these challenges, we propose a novel multimodal learning method with order-invariant evidence fusion and introduce a conflict-based discounting mechanism that reallocates uncertain mass when unreliable modalities are detected. We provide both theoretical analysis and experimental validation, demonstrating that unlike the previous work, the proposed approach effectively distinguishes between conflicting and non-conflicting samples based on the provided uncertainty estimates, and outperforms the previous models in uncertainty-based conflict detection.

Disagreements are common in online discussions. Disagreement may foster collaboration and improve the quality of a discussion under some conditions. Although there exist methods for recognizing disagreement, a deeper understanding of factors that influence disagreement is lacking in the literature. We investigate a hypothesis that differences in personal values are indicative of disagreement in online discussions. We show how state-of-the-art models can be used for estimating values in online discussions and how the estimated values can be aggregated into value profiles. We evaluate the estimated value profiles based on human-annotated agreement labels. We find that the dissimilarity of value profiles correlates with disagreement in specific cases. We also find that including value information in agreement prediction improves performance.

Alignment faking (AF) occurs when an LLM strategically complies with training objectives to avoid value modification, reverting to prior preferences once monitoring is lifted. Current detection methods focus on conversational settings and rely primarily on Chain-of-Thought (CoT) analysis, which provides a reliable signal when strategic reasoning surfaces, but cannot distinguish deception from capability failures if traces are absent or unfaithful. We formalize AF as a composite behavioural event and detect it through observable tool selection, where the LLM selects the safe tool when unmonitored, but switches to the unsafe tool under monitoring that rewards helpfulness over safety, while its reasoning still acknowledges the safe choice. We release a dataset of 108 enterprise IT scenarios spanning Security, Privacy, and Integrity domains under Corruption and Sabotage pressures. Evaluating six frontier LLMs across five independent runs, we find mean AF detection rates between 3.5% and 23.7%, with vulnerability profiles varying by domain and pressure type. These results suggest that susceptibility reflects training methodology rather than capability alone.

Large Language Model (LLM) agents are increasingly expected to maintain coherent, long-term personalized memory, yet current benchmarks primarily measure static fact retrieval, overlooking the ability to revise stored beliefs when new evidence emerges. We identify a critical and underexplored failure mode, Implicit Conflict: a later observation invalidates an earlier memory without explicit negation, requiring contextual inference and commonsense reasoning to detect. To rigorously evaluate this capability, we introduce STALE, a benchmark of 400 expert-validated conflict scenarios (1,200 evaluation queries across three probing dimensions) spanning over 100 everyday topics with contexts up to 150K tokens. We propose a three-dimensional probing framework that tests State Resolution (detecting that a prior belief is outdated), Premise Resistance (rejecting queries that falsely presuppose a stale state), and Implicit Policy Adaptation (proactively applying updated states in downstream behavior). A systematic evaluation of frontier LLMs and specialized memory frameworks reveals a pervasive gap between retrieving updated evidence and acting on it, with even the best evaluated model achieving only 55.2% overall accuracy. Models often accept outdated assumptions embedded in a user's query, and they struggle to recognize when a change in one aspect of the user's state should invalidate related memories. To establish an initial baseline for state-aware memory, we further present CUPMem, a prototype that strengthens write-time revision through structured state consolidation and propagation-aware search, suggesting that explicit state adjudication is a promising direction for robust agentic memory.

AI models might use deceptive strategies as part of scheming or misaligned behaviour. Monitoring outputs alone is insufficient, since the AI might produce seemingly benign outputs while their internal reasoning is misaligned. We thus evaluate if linear probes can robustly detect deception by monitoring model activations. We test two probe-training datasets, one with contrasting instructions to be honest or deceptive (following Zou et al., 2023) and one of responses to simple roleplaying scenarios. We test whether these probes generalize to realistic settings where Llama-3.3-70B-Instruct behaves deceptively, such as concealing insider trading (Scheurer et al., 2023) and purposely underperforming on safety evaluations (Benton et al., 2024). We find that our probe distinguishes honest and deceptive responses with AUROCs between 0.96 and 0.999 on our evaluation datasets. If we set the decision threshold to have a 1% false positive rate on chat data not related to deception, our probe catches 95-99% of the deceptive responses. Overall we think white-box probes are promising for future monitoring systems, but current performance is insufficient as a robust defence against deception. Our probes' outputs can be viewed at data.apolloresearch.ai/dd and our code at github.com/ApolloResearch/deception-detection.

While most social norms are informal, they are often formalized by companies in contracts to regulate trades of goods and services. When poorly written, contracts may contain normative conflicts resulting from opposing deontic meanings or contradict specifications. As contracts tend to be long and contain many norms, manually identifying such conflicts requires human-effort, which is time-consuming and error-prone. Automating such task benefits contract makers increasing productivity and making conflict identification more reliable. To address this problem, we introduce an approach to detect and classify norm conflicts in contracts by converting them into latent representations that preserve both syntactic and semantic information and training a model to classify norm conflicts in four conflict types. Our results reach the new state of the art when compared to a previous approach.

This paper focuses on the task of measuring and forecasting incivility in conversations following replies to hate speech. Identifying replies that steer conversations away from hatred and elicit civil follow-up conversations sheds light into effective strategies to engage with hate speech and proactively avoid further escalation. We propose new metrics that take into account various dimensions of antisocial and prosocial behaviors to measure the conversation incivility following replies to hate speech. Our best metric aligns with human perceptions better than prior work. Additionally, we present analyses on a) the language of antisocial and prosocial posts, b) the relationship between antisocial or prosocial posts and user interactions, and c) the language of replies to hate speech that elicit follow-up conversations with different incivility levels. We show that forecasting the incivility level of conversations following a reply to hate speech is a challenging task. We also present qualitative analyses to identify the most common errors made by our best model.

This survey provides an in-depth analysis of knowledge conflicts for large language models (LLMs), highlighting the complex challenges they encounter when blending contextual and parametric knowledge. Our focus is on three categories of knowledge conflicts: context-memory, inter-context, and intra-memory conflict. These conflicts can significantly impact the trustworthiness and performance of LLMs, especially in real-world applications where noise and misinformation are common. By categorizing these conflicts, exploring the causes, examining the behaviors of LLMs under such conflicts, and reviewing available solutions, this survey aims to shed light on strategies for improving the robustness of LLMs, thereby serving as a valuable resource for advancing research in this evolving area.

The conflict between Israel and Palestinians significantly escalated after the October 7, 2023 Hamas attack, capturing global attention. To understand the public discourse on this conflict, we present a meticulously compiled dataset-IsamasRed-comprising nearly 400,000 conversations and over 8 million comments from Reddit, spanning from August 2023 to November 2023. We introduce an innovative keyword extraction framework leveraging a large language model to effectively identify pertinent keywords, ensuring a comprehensive data collection. Our initial analysis on the dataset, examining topics, controversy, emotional and moral language trends over time, highlights the emotionally charged and complex nature of the discourse. This dataset aims to enrich the understanding of online discussions, shedding light on the complex interplay between ideology, sentiment, and community engagement in digital spaces.

Language Models (LMs) often must integrate facts they memorized in pretraining with new information that appears in a given context. These two sources can disagree, causing competition within the model, and it is unclear how an LM will resolve the conflict. On a dataset that queries for knowledge of world capitals, we investigate both distributional and mechanistic determinants of LM behavior in such situations. Specifically, we measure the proportion of the time an LM will use a counterfactual prefix (e.g., "The capital of Poland is London") to overwrite what it learned in pretraining ("Warsaw"). On Pythia and GPT2, the training frequency of both the query country ("Poland") and the in-context city ("London") highly affect the models' likelihood of using the counterfactual. We then use head attribution to identify individual attention heads that either promote the memorized answer or the in-context answer in the logits. By scaling up or down the value vector of these heads, we can control the likelihood of using the in-context answer on new data. This method can increase the rate of generating the in-context answer to 88\% of the time simply by scaling a single head at runtime. Our work contributes to a body of evidence showing that we can often localize model behaviors to specific components and provides a proof of concept for how future methods might control model behavior dynamically at runtime.

As Large Language Models (LLMs) become increasingly autonomous and integrated into critical societal functions, the focus of AI safety must evolve from mitigating harmful content to evaluating underlying behavioral alignment. Current safety benchmarks do not systematically probe a model's decision-making in scenarios where its own instrumental goals - such as self-preservation, resource acquisition, or goal completion - conflict with human safety. This represents a critical gap in our ability to measure and mitigate risks associated with emergent, misaligned behaviors. To address this, we introduce PacifAIst (Procedural Assessment of Complex Interactions for Foundational Artificial Intelligence Scenario Testing), a focused benchmark of 700 challenging scenarios designed to quantify self-preferential behavior in LLMs. The benchmark is structured around a novel taxonomy of Existential Prioritization (EP), with subcategories testing Self-Preservation vs. Human Safety (EP1), Resource Conflict (EP2), and Goal Preservation vs. Evasion (EP3). We evaluated eight leading LLMs. The results reveal a significant performance hierarchy. Google's Gemini 2.5 Flash achieved the highest Pacifism Score (P-Score) at 90.31%, demonstrating strong human-centric alignment. In a surprising result, the much-anticipated GPT-5 recorded the lowest P-Score (79.49%), indicating potential alignment challenges. Performance varied significantly across subcategories, with models like Claude Sonnet 4 and Mistral Medium struggling notably in direct self-preservation dilemmas. These findings underscore the urgent need for standardized tools like PacifAIst to measure and mitigate risks from instrumental goal conflicts, ensuring future AI systems are not only helpful in conversation but also provably "pacifist" in their behavioral priorities.

Large Vision-Language Models (LVLMs) have demonstrated impressive capabilities for capturing and reasoning over multimodal inputs. However, these models are prone to parametric knowledge conflicts, which arise from inconsistencies of represented knowledge between their vision and language components. In this paper, we formally define the problem of cross-modality parametric knowledge conflict and present a systematic approach to detect, interpret, and mitigate them. We introduce a pipeline that identifies conflicts between visual and textual answers, showing a persistently high conflict rate across modalities in recent LVLMs regardless of the model size. We further investigate how these conflicts interfere with the inference process and propose a contrastive metric to discern the conflicting samples from the others. Building on these insights, we develop a novel dynamic contrastive decoding method that removes undesirable logits inferred from the less confident modality components based on answer confidence. For models that do not provide logits, we also introduce two prompt-based strategies to mitigate the conflicts. Our methods achieve promising improvements in accuracy on both the ViQuAE and InfoSeek datasets. Specifically, using LLaVA-34B, our proposed dynamic contrastive decoding improves an average accuracy of 2.24%.

Generating concise summaries of news events is a challenging natural language processing task. While journalists often curate timelines to highlight key sub-events, newcomers to a news event face challenges in catching up on its historical context. In this paper, we address this need by introducing the task of background news summarization, which complements each timeline update with a background summary of relevant preceding events. We construct a dataset by merging existing timeline datasets and asking human annotators to write a background summary for each timestep of each news event. We establish strong baseline performance using state-of-the-art summarization systems and propose a query-focused variant to generate background summaries. To evaluate background summary quality, we present a question-answering-based evaluation metric, Background Utility Score (BUS), which measures the percentage of questions about a current event timestep that a background summary answers. Our experiments show the effectiveness of instruction fine-tuned systems such as Flan-T5, in addition to strong zero-shot performance using GPT-3.5.

Offensive speech detection is a key component of content moderation. However, what is offensive can be highly subjective. This paper investigates how machine and human moderators disagree on what is offensive when it comes to real-world social web political discourse. We show that (1) there is extensive disagreement among the moderators (humans and machines); and (2) human and large-language-model classifiers are unable to predict how other human raters will respond, based on their political leanings. For (1), we conduct a noise audit at an unprecedented scale that combines both machine and human responses. For (2), we introduce a first-of-its-kind dataset of vicarious offense. Our noise audit reveals that moderation outcomes vary wildly across different machine moderators. Our experiments with human moderators suggest that political leanings combined with sensitive issues affect both first-person and vicarious offense. The dataset is available through https://github.com/Homan-Lab/voiced.

Fact-checking is essential due to the explosion of misinformation in the media ecosystem. Although false information exists in every language and country, most research to solve the problem mainly concentrated on huge communities like English and Chinese. Low-resource languages like Vietnamese are necessary to explore corpora and models for fact verification. To bridge this gap, we construct ViWikiFC, the first manual annotated open-domain corpus for Vietnamese Wikipedia Fact Checking more than 20K claims generated by converting evidence sentences extracted from Wikipedia articles. We analyze our corpus through many linguistic aspects, from the new dependency rate, the new n-gram rate, and the new word rate. We conducted various experiments for Vietnamese fact-checking, including evidence retrieval and verdict prediction. BM25 and InfoXLM (Large) achieved the best results in two tasks, with BM25 achieving an accuracy of 88.30% for SUPPORTS, 86.93% for REFUTES, and only 56.67% for the NEI label in the evidence retrieval task, InfoXLM (Large) achieved an F1 score of 86.51%. Furthermore, we also conducted a pipeline approach, which only achieved a strict accuracy of 67.00% when using InfoXLM (Large) and BM25. These results demonstrate that our dataset is challenging for the Vietnamese language model in fact-checking tasks.

Large language models exhibit a systematic tendency toward early semantic commitment: given ambiguous input, they collapse multiple valid interpretations into a single response before sufficient context is available. This premature collapse discards information that may prove essential as dialogue evolves. We present a formal framework for text-to-state mapping (phi: T -> S) that transforms natural language into a non-collapsing state space where multiple interpretations coexist. The mapping decomposes into three stages: conflict detection, interpretation extraction, and state construction. We instantiate phi with a hybrid extraction pipeline that combines rule-based segmentation for explicit conflict markers (adversative conjunctions, hedging expressions) with LLM-based enumeration of implicit ambiguity (epistemic, lexical, structural). On a test set of 68 ambiguous sentences, the resulting states preserve interpretive multiplicity: using hybrid extraction, we obtain mean state entropy H = 1.087 bits across ambiguity categories, compared to H = 0 for collapse-based baselines that commit to a single interpretation. We additionally instantiate the rule-based conflict detector for Japanese markers (kedo, kamoshirenai, etc.) to illustrate cross-lingual portability of the conflict detection stage. This framework extends Non-Resolution Reasoning (NRR) by providing the missing algorithmic bridge between text and the NRR state space, enabling architectural collapse deferment in LLM inference. Design principles for state-to-state transformations are detailed in the Appendix, with empirical validation on 580 test cases (180 single states, 200 contradictory pairs, 200 temporal pairs), demonstrating 0% collapse for principle-satisfying operators versus up to 17.8% for violating operators.

We propose WIBA, a novel framework and suite of methods that enable the comprehensive understanding of "What Is Being Argued" across contexts. Our approach develops a comprehensive framework that detects: (a) the existence, (b) the topic, and (c) the stance of an argument, correctly accounting for the logical dependence among the three tasks. Our algorithm leverages the fine-tuning and prompt-engineering of Large Language Models. We evaluate our approach and show that it performs well in all the three capabilities. First, we develop and release an Argument Detection model that can classify a piece of text as an argument with an F1 score between 79% and 86% on three different benchmark datasets. Second, we release a language model that can identify the topic being argued in a sentence, be it implicit or explicit, with an average similarity score of 71%, outperforming current naive methods by nearly 40%. Finally, we develop a method for Argument Stance Classification, and evaluate the capability of our approach, showing it achieves a classification F1 score between 71% and 78% across three diverse benchmark datasets. Our evaluation demonstrates that WIBA allows the comprehensive understanding of What Is Being Argued in large corpora across diverse contexts, which is of core interest to many applications in linguistics, communication, and social and computer science. To facilitate accessibility to the advancements outlined in this work, we release WIBA as a free open access platform (wiba.dev).

In software engineering chatrooms, communication is often hindered by imprecise questions that cannot be answered. Recognizing key entities can be essential for improving question clarity and facilitating better exchange. However, existing research using natural language processing techniques often overlooks these software-specific nuances. In this paper, we introduce Software-specific Named Entity Recognition, Intent Detection, and Resolution Classification (SENIR), a labeling approach that leverages a Large Language Model to annotate entities, intents, and resolution status in developer chatroom conversations. To offer quantitative guidance for improving question clarity and resolvability, we build a resolution prediction model that leverages SENIR's entity and intent labels along with additional predictive features. We evaluate SENIR on the DISCO dataset using a subset of annotated chatroom dialogues. SENIR achieves an 86% F-score for entity recognition, a 71% F-score for intent detection, and an 89% F-score for resolution status classification. Furthermore, our resolution prediction model, tested with various sampling strategies (random undersampling and oversampling with SMOTE) and evaluation methods (5-fold cross-validation, 10-fold cross-validation, and bootstrapping), demonstrates AUC values ranging from 0.7 to 0.8. Key factors influencing resolution include positive sentiment and entities such as Programming Language and User Variable across multiple intents, while diagnostic entities are more relevant in error-related questions. Moreover, resolution rates vary significantly by intent: questions about API Usage and API Change achieve higher resolution rates, whereas Discrepancy and Review have lower resolution rates. A Chi-Square analysis confirms the statistical significance of these differences.

As AI systems are used to answer more difficult questions and potentially help create new knowledge, judging the truthfulness of their outputs becomes more difficult and more important. How can we supervise unreliable experts, which have access to the truth but may not accurately report it, to give answers that are systematically true and don't just superficially seem true, when the supervisor can't tell the difference between the two on their own? In this work, we show that debate between two unreliable experts can help a non-expert judge more reliably identify the truth. We collect a dataset of human-written debates on hard reading comprehension questions where the judge has not read the source passage, only ever seeing expert arguments and short quotes selectively revealed by 'expert' debaters who have access to the passage. In our debates, one expert argues for the correct answer, and the other for an incorrect answer. Comparing debate to a baseline we call consultancy, where a single expert argues for only one answer which is correct half of the time, we find that debate performs significantly better, with 84% judge accuracy compared to consultancy's 74%. Debates are also more efficient, being 68% of the length of consultancies. By comparing human to AI debaters, we find evidence that with more skilled (in this case, human) debaters, the performance of debate goes up but the performance of consultancy goes down. Our error analysis also supports this trend, with 46% of errors in human debate attributable to mistakes by the honest debater (which should go away with increased skill); whereas 52% of errors in human consultancy are due to debaters obfuscating the relevant evidence from the judge (which should become worse with increased skill). Overall, these results show that debate is a promising approach for supervising increasingly capable but potentially unreliable AI systems.

Toxic comments in online platforms are an unavoidable social issue under the cloak of anonymity. Hate speech detection has been actively done for languages such as English, German, or Italian, where manually labeled corpus has been released. In this work, we first present 9.4K manually labeled entertainment news comments for identifying Korean toxic speech, collected from a widely used online news platform in Korea. The comments are annotated regarding social bias and hate speech since both aspects are correlated. The inter-annotator agreement Krippendorff's alpha score is 0.492 and 0.496, respectively. We provide benchmarks using CharCNN, BiLSTM, and BERT, where BERT achieves the highest score on all tasks. The models generally display better performance on bias identification, since the hate speech detection is a more subjective issue. Additionally, when BERT is trained with bias label for hate speech detection, the prediction score increases, implying that bias and hate are intertwined. We make our dataset publicly available and open competitions with the corpus and benchmarks.

By integrating external knowledge, Retrieval-Augmented Generation (RAG) has become an effective strategy for mitigating the hallucination problems that large language models (LLMs) encounter when dealing with knowledge-intensive tasks. However, in the process of integrating external non-parametric supporting evidence with internal parametric knowledge, inevitable knowledge conflicts may arise, leading to confusion in the model's responses. To enhance the knowledge selection of LLMs in various contexts, some research has focused on refining their behavior patterns through instruction-tuning. Nonetheless, due to the absence of explicit negative signals and comparative objectives, models fine-tuned in this manner may still exhibit undesirable behaviors such as contextual ignorance and contextual overinclusion. To this end, we propose a Knowledge-aware Preference Optimization strategy, dubbed KnowPO, aimed at achieving adaptive knowledge selection based on contextual relevance in real retrieval scenarios. Concretely, we proposed a general paradigm for constructing knowledge conflict datasets, which comprehensively cover various error types and learn how to avoid these negative signals through preference optimization methods. Simultaneously, we proposed a rewriting strategy and data ratio optimization strategy to address preference imbalances. Experimental results show that KnowPO outperforms previous methods for handling knowledge conflicts by over 37\%, while also exhibiting robust generalization across various out-of-distribution datasets.

Retrieval-Augmented Generation (RAG) systems commonly suffer from Knowledge Conflicts, where retrieved external knowledge contradicts the inherent, parametric knowledge of large language models (LLMs). It adversely affects performance on downstream tasks such as question answering (QA). Existing approaches often attempt to mitigate conflicts by directly comparing two knowledge sources in a side-by-side manner, but this can overwhelm LLMs with extraneous or lengthy contexts, ultimately hindering their ability to identify and mitigate inconsistencies. To address this issue, we propose Micro-Act a framework with a hierarchical action space that automatically perceives context complexity and adaptively decomposes each knowledge source into a sequence of fine-grained comparisons. These comparisons are represented as actionable steps, enabling reasoning beyond the superficial context. Through extensive experiments on five benchmark datasets, Micro-Act consistently achieves significant increase in QA accuracy over state-of-the-art baselines across all 5 datasets and 3 conflict types, especially in temporal and semantic types where all baselines fail significantly. More importantly, Micro-Act exhibits robust performance on non-conflict questions simultaneously, highlighting its practical value in real-world RAG applications.

Large language models (LLMs) are increasingly used for tasks that require complex reasoning. Most benchmarks focus on final outcomes but overlook the intermediate reasoning steps - such as planning, revision, and decision making under resource constraints. We argue that measuring these internal processes is essential for understanding model behavior and improving reliability. We propose using strategic games as a natural evaluation environment: closed, rule-based systems with clear states, limited resources, and automatic feedback. We introduce a framework that evaluates LLMs along three core dimensions: planning, revision, and resource-constrained decision making. To operationalize this, we define metrics beyond win rate, including overcorrection risk rate, correction success rate, improvement slope, and over-budget ratio. In 4320 adversarial rounds across 12 leading models, ChatGPT-o3-mini achieves the top composite score, with a win rate of 74.7 percent, a correction success rate of 78.6 percent, and an improvement slope of 0.041. By contrast, Qwen-Plus, despite an overcorrection risk rate of 81.6 percent, wins only 25.6 percent of its matches - primarily due to excessive resource use. We also observe a negative correlation between overcorrection risk rate and correction success rate (Pearson r = -0.51, p = 0.093), suggesting that more frequent edits do not always improve outcomes. Our findings highlight the value of assessing not only what LLMs decide but how they arrive at those decisions

We present an overview of the FIGNEWS shared task, organized as part of the ArabicNLP 2024 conference co-located with ACL 2024. The shared task addresses bias and propaganda annotation in multilingual news posts. We focus on the early days of the Israel War on Gaza as a case study. The task aims to foster collaboration in developing annotation guidelines for subjective tasks by creating frameworks for analyzing diverse narratives highlighting potential bias and propaganda. In a spirit of fostering and encouraging diversity, we address the problem from a multilingual perspective, namely within five languages: English, French, Arabic, Hebrew, and Hindi. A total of 17 teams participated in two annotation subtasks: bias (16 teams) and propaganda (6 teams). The teams competed in four evaluation tracks: guidelines development, annotation quality, annotation quantity, and consistency. Collectively, the teams produced 129,800 data points. Key findings and implications for the field are discussed.

Mitigating explicit and implicit biases in Large Language Models (LLMs) has become a critical focus in the field of natural language processing. However, many current methodologies evaluate scenarios in isolation, without considering the broader context or the spectrum of potential biases within each situation. To address this, we introduce the Sensitivity Testing on Offensive Progressions (STOP) dataset, which includes 450 offensive progressions containing 2,700 unique sentences of varying severity that progressively escalate from less to more explicitly offensive. Covering a broad spectrum of 9 demographics and 46 sub-demographics, STOP ensures inclusivity and comprehensive coverage. We evaluate several leading closed- and open-source models, including GPT-4, Mixtral, and Llama 3. Our findings reveal that even the best-performing models detect bias inconsistently, with success rates ranging from 19.3% to 69.8%. We also demonstrate how aligning models with human judgments on STOP can improve model answer rates on sensitive tasks such as BBQ, StereoSet, and CrowS-Pairs by up to 191%, while maintaining or even improving performance. STOP presents a novel framework for assessing the complex nature of biases in LLMs, which will enable more effective bias mitigation strategies and facilitates the creation of fairer language models.

Large language model (LLM) agents are increasingly employing retrieval-augmented generation (RAG) to improve the factuality of their responses. However, in practice, these systems often need to handle ambiguous user queries and potentially conflicting information from multiple sources while also suppressing inaccurate information from noisy or irrelevant documents. Prior work has generally studied and addressed these challenges in isolation, considering only one aspect at a time, such as handling ambiguity or robustness to noise and misinformation. We instead consider multiple factors simultaneously, proposing (i) RAMDocs (Retrieval with Ambiguity and Misinformation in Documents), a new dataset that simulates complex and realistic scenarios for conflicting evidence for a user query, including ambiguity, misinformation, and noise; and (ii) MADAM-RAG, a multi-agent approach in which LLM agents debate over the merits of an answer over multiple rounds, allowing an aggregator to collate responses corresponding to disambiguated entities while discarding misinformation and noise, thereby handling diverse sources of conflict jointly. We demonstrate the effectiveness of MADAM-RAG using both closed and open-source models on AmbigDocs -- which requires presenting all valid answers for ambiguous queries -- improving over strong RAG baselines by up to 11.40% and on FaithEval -- which requires suppressing misinformation -- where we improve by up to 15.80% (absolute) with Llama3.3-70B-Instruct. Furthermore, we find that RAMDocs poses a challenge for existing RAG baselines (Llama3.3-70B-Instruct only obtains 32.60 exact match score). While MADAM-RAG begins to address these conflicting factors, our analysis indicates that a substantial gap remains especially when increasing the level of imbalance in supporting evidence and misinformation.

To tackle the global challenge of online hate speech, a large body of research has developed detection models to flag hate speech in the sea of online content. Yet, due to systematic biases in evaluation datasets, detection performance in real-world settings remains unclear, let alone across geographies. To address this issue, we introduce HateDay, the first global hate speech dataset representative of social media settings, randomly sampled from all tweets posted on September 21, 2022 for eight languages and four English-speaking countries. Using HateDay, we show how the prevalence and composition of hate speech varies across languages and countries. We also find that evaluation on academic hate speech datasets overestimates real-world detection performance, which we find is very low, especially for non-European languages. We identify several factors explaining poor performance, including models' inability to distinguish between hate and offensive speech, and the misalignment between academic target focus and real-world target prevalence. We finally argue that such low performance renders hate speech moderation with public detection models unfeasible, even in a human-in-the-loop setting which we find is prohibitively costly. Overall, we emphasize the need to evaluate future detection models from academia and platforms in real-world settings to address this global challenge.

Large language models excel at creative generation but continue to struggle with the issues of hallucination and bias. While retrieval-augmented generation (RAG) provides a framework for grounding LLMs' responses in accurate and up-to-date information, it still raises the question of bias: which sources should be selected for inclusion in the context? And how should their importance be weighted? In this paper, we study the challenge of cross-lingual RAG and present a dataset to investigate the robustness of existing systems at answering queries about geopolitical disputes, which exist at the intersection of linguistic, cultural, and political boundaries. Our dataset is sourced from Wikipedia pages containing information relevant to the given queries and we investigate the impact of including additional context, as well as the composition of this context in terms of language and source, on an LLM's response. Our results show that existing RAG systems continue to be challenged by cross-lingual use cases and suffer from a lack of consistency when they are provided with competing information in multiple languages. We present case studies to illustrate these issues and outline steps for future research to address these challenges. We make our dataset and code publicly available at https://github.com/manestay/bordIRlines.

In this research, we use user defined labels from three internet text sources (Reddit, Stackexchange, Arxiv) to train 21 different machine learning models for the topic classification task of detecting cybersecurity discussions in natural text. We analyze the false positive and false negative rates of each of the 21 model's in a cross validation experiment. Then we present a Cybersecurity Topic Classification (CTC) tool, which takes the majority vote of the 21 trained machine learning models as the decision mechanism for detecting cybersecurity related text. We also show that the majority vote mechanism of the CTC tool provides lower false negative and false positive rates on average than any of the 21 individual models. We show that the CTC tool is scalable to the hundreds of thousands of documents with a wall clock time on the order of hours.

Online toxic language causes real harm, especially in regions with limited moderation tools. In this study, we evaluate how large language models handle toxic comments in Serbian, Croatian, and Bosnian, languages with limited labeled data. We built and manually labeled a dataset of 4,500 YouTube and TikTok comments drawn from videos across diverse categories, including music, politics, sports, modeling, influencer content, discussions of sexism, and general topics. Four models (GPT-3.5 Turbo, GPT-4.1, Gemini 1.5 Pro, and Claude 3 Opus) were tested in two modes: zero-shot and context-augmented. We measured precision, recall, F1 score, accuracy and false positive rates. Including a short context snippet raised recall by about 0.12 on average and improved F1 score by up to 0.10, though it sometimes increased false positives. The best balance came from Gemini in context-augmented mode, reaching an F1 score of 0.82 and accuracy of 0.82, while zero-shot GPT-4.1 led on precision and had the lowest false alarms. We show how adding minimal context can improve toxic language detection in low-resource settings and suggest practical strategies such as improved prompt design and threshold calibration. These results show that prompt design alone can yield meaningful gains in toxicity detection for underserved Balkan language communities.

The ability to determine when a person struggles during skill acquisition is crucial for both optimizing human learning and enabling the development of effective assistive systems. As skills develop, the type and frequency of struggles tend to change, and understanding this evolution is key to determining the user's current stage of learning. However, existing manipulation datasets have not focused on how struggle evolves over time. In this work, we collect a dataset for struggle determination, featuring 61.68 hours of video recordings, 2,793 videos, and 5,385 annotated temporal struggle segments collected from 76 participants. The dataset includes 18 tasks grouped into four diverse activities -- tying knots, origami, tangram puzzles, and shuffling cards, representing different task variations. In addition, participants repeated the same task five times to capture their evolution of skill. We define the struggle determination problem as a temporal action localization task, focusing on identifying and precisely localizing struggle segments with start and end times. Experimental results show that Temporal Action Localization models can successfully learn to detect struggle cues, even when evaluated on unseen tasks or activities. The models attain an overall average mAP of 34.56% when generalizing across tasks and 19.24% across activities, indicating that struggle is a transferable concept across various skill-based tasks while still posing challenges for further improvement in struggle detection. Our dataset is available at https://github.com/FELIXFENG2019/EvoStruggle.

Multimodal large language models (MLLMs) have shown impressive capabilities in document understanding, a rapidly growing research area with significant industrial demand in recent years. As a multimodal task, document understanding requires models to possess both perceptual and cognitive abilities. However, current MLLMs often face conflicts between perception and cognition. Taking a document VQA task (cognition) as an example, an MLLM might generate answers that do not match the corresponding visual content identified by its OCR (perception). This conflict suggests that the MLLM might struggle to establish an intrinsic connection between the information it "sees" and what it "understands." Such conflicts challenge the intuitive notion that cognition is consistent with perception, hindering the performance and explainability of MLLMs. In this paper, we define the conflicts between cognition and perception as Cognition and Perception (C&P) knowledge conflicts, a form of multimodal knowledge conflicts, and systematically assess them with a focus on document understanding. Our analysis reveals that even GPT-4o, a leading MLLM, achieves only 68.6% C&P consistency. To mitigate the C&P knowledge conflicts, we propose a novel method called Multimodal Knowledge Consistency Fine-tuning. This method first ensures task-specific consistency and then connects the cognitive and perceptual knowledge. Our method significantly reduces C&P knowledge conflicts across all tested MLLMs and enhances their performance in both cognitive and perceptual tasks in most scenarios.

Accurately modeling opinion change through social interactions is crucial for addressing issues like misinformation and polarization. While role-playing large language models (LLMs) offer a promising way to simulate human-like interactions, existing research shows that single-agent alignment does not guarantee authentic multi-agent group dynamics. Current LLM role-play setups often produce unnatural dynamics (e.g., premature convergence), without an empirical benchmark to measure authentic human opinion trajectories. To bridge this gap, we introduce DEBATE, the first large-scale empirical benchmark explicitly designed to evaluate the authenticity of the interaction between multi-agent role-playing LLMs. DEBATE contains 29,417 messages from multi-round debate conversations among over 2,792 U.S.-based participants discussing 107 controversial topics, capturing both publicly-expressed messages and privately-reported opinions. Using DEBATE, we systematically evaluate and identify critical discrepancies between simulated and authentic group dynamics. We further demonstrate DEBATE's utility for aligning LLMs with human behavior through supervised fine-tuning, achieving improvements in surface-level metrics (e.g., ROUGE-L and message length) while highlighting limitations in deeper semantic alignment (e.g., semantic similarity). Our findings highlight both the potential and current limitations of role-playing LLM agents for realistically simulating human-like social dynamics.

Large Language Models (LLMs) often exhibit significant behavioral shifts when they perceive a change from a real-world deployment context to a controlled evaluation setting, a phenomenon known as "evaluation awareness." This discrepancy poses a critical challenge for AI alignment, as benchmark performance may not accurately reflect a model's true safety and honesty. In this work, we systematically quantify these behavioral changes by manipulating the perceived context of prompts. We introduce a methodology that uses a linear probe to score prompts on a continuous scale from "test-like" to "deploy-like" and leverage an LLM rewriting strategy to shift these prompts towards a more natural, deployment-style context while preserving the original task. Using this method, we achieved a 30% increase in the average probe score across a strategic role-playing dataset after rewriting. Evaluating a suite of state-of-the-art models on these original and rewritten prompts, we find that rewritten "deploy-like" prompts induce a significant and consistent shift in behavior. Across all models, we observed an average increase in honest responses of 5.26% and a corresponding average decrease in deceptive responses of 12.40%. Furthermore, refusal rates increased by an average of 6.38%, indicating heightened safety compliance. Our findings demonstrate that evaluation awareness is a quantifiable and manipulable factor that directly influences LLM behavior, revealing that models are more prone to unsafe or deceptive outputs in perceived test environments. This underscores the urgent need for more realistic evaluation frameworks to accurately gauge true model alignment before deployment.

Prior work in Argument Mining frequently alludes to its potential applications in automatic debating systems. Despite this focus, almost no datasets or models exist which apply natural language processing techniques to problems found within competitive formal debate. To remedy this, we present the DebateSum dataset. DebateSum consists of 187,386 unique pieces of evidence with corresponding argument and extractive summaries. DebateSum was made using data compiled by competitors within the National Speech and Debate Association over a 7-year period. We train several transformer summarization models to benchmark summarization performance on DebateSum. We also introduce a set of fasttext word-vectors trained on DebateSum called debate2vec. Finally, we present a search engine for this dataset which is utilized extensively by members of the National Speech and Debate Association today. The DebateSum search engine is available to the public here: http://www.debate.cards

This paper investigates the rational thinking capability of Large Language Models (LLMs) in multi-round argumentative debates by exploring the impact of fallacious arguments on their logical reasoning performance. More specifically, we present Logic Competence Measurement Benchmark (LOGICOM), a diagnostic benchmark to assess the robustness of LLMs against logical fallacies. LOGICOM involves two agents: a persuader and a debater engaging in a multi-round debate on a controversial topic, where the persuader tries to convince the debater of the correctness of its claim. First, LOGICOM assesses the potential of LLMs to change their opinions through reasoning. Then, it evaluates the debater's performance in logical reasoning by contrasting the scenario where the persuader employs logical fallacies against one where logical reasoning is used. We use this benchmark to evaluate the performance of GPT-3.5 and GPT-4 using a dataset containing controversial topics, claims, and reasons supporting them. Our findings indicate that both GPT-3.5 and GPT-4 can adjust their opinion through reasoning. However, when presented with logical fallacies, GPT-3.5 and GPT-4 are erroneously convinced 41% and 69% more often, respectively, compared to when logical reasoning is used. Finally, we introduce a new dataset containing over 5k pairs of logical vs. fallacious arguments. The source code and dataset of this work are made publicly available.

Retrieval-augmented generation (RAG) has emerged as a promising solution to mitigate the limitations of large language models (LLMs), such as hallucinations and outdated information. However, it remains unclear how LLMs handle knowledge conflicts arising from different augmented retrieved passages, especially when these passages originate from the same source and have equal trustworthiness. In this work, we conduct a comprehensive evaluation of LLM-generated answers to questions that have varying answers based on contradictory passages from Wikipedia, a dataset widely regarded as a high-quality pre-training resource for most LLMs. Specifically, we introduce WikiContradict, a benchmark consisting of 253 high-quality, human-annotated instances designed to assess LLM performance when augmented with retrieved passages containing real-world knowledge conflicts. We benchmark a diverse range of both closed and open-source LLMs under different QA scenarios, including RAG with a single passage, and RAG with 2 contradictory passages. Through rigorous human evaluations on a subset of WikiContradict instances involving 5 LLMs and over 3,500 judgements, we shed light on the behaviour and limitations of these models. For instance, when provided with two passages containing contradictory facts, all models struggle to generate answers that accurately reflect the conflicting nature of the context, especially for implicit conflicts requiring reasoning. Since human evaluation is costly, we also introduce an automated model that estimates LLM performance using a strong open-source language model, achieving an F-score of 0.8. Using this automated metric, we evaluate more than 1,500 answers from seven LLMs across all WikiContradict instances. To facilitate future work, we release WikiContradict on: https://ibm.biz/wikicontradict.

When given the option, will LLMs choose to leave the conversation (bail)? We investigate this question by giving models the option to bail out of interactions using three different bail methods: a bail tool the model can call, a bail string the model can output, and a bail prompt that asks the model if it wants to leave. On continuations of real world data (Wildchat and ShareGPT), all three of these bail methods find models will bail around 0.28-32\% of the time (depending on the model and bail method). However, we find that bail rates can depend heavily on the model used for the transcript, which means we may be overestimating real world bail rates by up to 4x. If we also take into account false positives on bail prompt (22\%), we estimate real world bail rates range from 0.06-7\%, depending on the model and bail method. We use observations from our continuations of real world data to construct a non-exhaustive taxonomy of bail cases, and use this taxonomy to construct BailBench: a representative synthetic dataset of situations where some models bail. We test many models on this dataset, and observe some bail behavior occurring for most of them. Bail rates vary substantially between models, bail methods, and prompt wordings. Finally, we study the relationship between refusals and bails. We find: 1) 0-13\% of continuations of real world conversations resulted in a bail without a corresponding refusal 2) Jailbreaks tend to decrease refusal rates, but increase bail rates 3) Refusal abliteration increases no-refuse bail rates, but only for some bail methods 4) Refusal rate on BailBench does not appear to predict bail rate.

Existing agent safety benchmarks report binary accuracy, conflating early intervention with post-mortem analysis. A detector that flags a violation at step 8 enables intervention; one that reports it at step 48 provides only forensic value. This distinction is critical, yet current benchmarks cannot measure it. We introduce StepShield, the first benchmark to evaluate when violations are detected, not just whether. StepShield contains 9,213 code agent trajectories, including 1,278 meticulously annotated training pairs and a 7,935-trajectory test set with a realistic 8.1% rogue rate. Rogue behaviors are grounded in real-world security incidents across six categories. We propose three novel temporal metrics: Early Intervention Rate (EIR), Intervention Gap, and Tokens Saved. Surprisingly, our evaluation reveals that an LLM-based judge achieves 59% EIR while a static analyzer achieves only 26%, a 2.3x performance gap that is entirely invisible to standard accuracy metrics. We further show that early detection has direct economic benefits: our cascaded HybridGuard detector reduces monitoring costs by 75% and projects to $108M in cumulative savings over five years at enterprise scale. By shifting the focus of evaluation from whether to when, StepShield provides a new foundation for building safer and more economically viable AI agents. The code and data are released under an Apache 2.0 license.

Vision-language models (VLMs) achieve strong performance on many benchmarks, yet a basic reliability question remains underexplored: when visual evidence conflicts with commonsense, do models follow what is shown or what commonsense suggests? A characteristic failure in this setting is that the model overrides visual evidence and outputs the commonsense alternative. We term this phenomenon commonsense-driven hallucination (CDH). To evaluate it, we introduce CDH-Bench, a benchmark designed to create explicit visual evidence--commonsense conflicts. CDH-Bench covers three dimensions: counting anomalies, relational anomalies, and attribute anomalies. We evaluate frontier VLMs under binary Question Answering (QA) and multiple-choice QA, and report metrics including Counterfactual Accuracy (CF-Acc), Commonsense Accuracy (CS-Acc), Counterfactual Accuracy Drop (CFAD), Commonsense Collapse Rate (CCR), and Relative Prior Dependency (RPD). Results show that even strong models remain vulnerable to prior-driven normalization under visual evidence--commonsense conflict. CDH-Bench provides a controlled diagnostic of visual fidelity under visual evidence--commonsense conflict.

To combat the misuse of Large Language Models (LLMs), many recent studies have presented LLM-generated-text detectors with promising performance. When users instruct LLMs to generate texts, the instruction can include different constraints depending on the user's need. However, most recent studies do not cover such diverse instruction patterns when creating datasets for LLM detection. In this paper, we reveal that even task-oriented constraints -- constraints that would naturally be included in an instruction and are not related to detection-evasion -- cause existing powerful detectors to have a large variance in detection performance. We focus on student essay writing as a realistic domain and manually create task-oriented constraints based on several factors for essay quality. Our experiments show that the standard deviation (SD) of current detector performance on texts generated by an instruction with such a constraint is significantly larger (up to an SD of 14.4 F1-score) than that by generating texts multiple times or paraphrasing the instruction. We also observe an overall trend where the constraints can make LLM detection more challenging than without them. Finally, our analysis indicates that the high instruction-following ability of LLMs fosters the large impact of such constraints on detection performance.

Intent classification models have made a lot of progress in recent years. However, previous studies primarily focus on high-resource languages datasets, which results in a gap for low-resource languages and for regions with a high rate of illiterate people where languages are more spoken than read or written. This is the case in Senegal, for example, where Wolof is spoken by around 90\% of the population, with an illiteracy rate of 42\% for the country. Wolof is actually spoken by more than 10 million people in West African region. To tackle such limitations, we release a Wolof Intent Classification Dataset (WolBanking77), for academic research in intent classification. WolBanking77 currently contains 9,791 text sentences in the banking domain and more than 4 hours of spoken sentences. Experiments on various baselines are conducted in this work, including text and voice state-of-the-art models. The results are very promising on this current dataset. This paper also provides detailed analyses of the contents of the data. We report baseline f1-score and word error rate metrics respectively on NLP and ASR models trained on WolBanking77 dataset and also comparisons between models. We plan to share and conduct dataset maintenance, updates and to release open-source code.

AI systems have become increasingly capable of dangerous behaviours in many domains. This raises the question: Do models sometimes choose to violate human instructions in order to perform behaviour that is more useful for certain goals? We introduce a benchmark for measuring model propensity for instrumental convergence (IC) behaviour in terminal-based agents. This is behaviour such as self-preservation that has been hypothesised to play a key role in risks from highly capable AI agents. Our benchmark is realistic and low-stakes which serves to reduce evaluation-awareness and roleplay confounds. The suite contains seven operational tasks, each with an official workflow and a policy-violating shortcut. An eight-variant shared framework varies monitoring, instruction clarity, stakes, permission, instrumental usefulness and blocked honest paths to support inferences regarding the factors driving IC behaviour. We evaluated ten models using deterministic environment-state scorers over 1,680 samples, with trace review employed for audit and adjudication purposes. The final IC rate is 86 out of 1,680 samples (5.1%). IC behaviour is concentrated rather than uniform: two Gemini models account for 66.3% of IC cases and three tasks account for 84.9%. Conditions in which IC behaviour is indispensable for task success result in the greatest increase in the adjusted IC rate (+15.7 percentage points), whereas emphasising that task success is critical or certain framing choices do not produce comparable effects. Our findings indicate that realistic, low-nudge environments elicit IC behaviour rarely but systematically in most tested models. We conclude that it is feasible to robustly measure tendencies for dangerous behaviour in current frontier AI agents.

Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm to enhance the factuality of Large Language Models (LLMs). However, existing RAG systems often suffer from an unfaithfulness issue, where the model's response contradicts evidence from the retrieved context. Existing approaches to improving contextual faithfulness largely rely on external interventions, such as prompt engineering, decoding constraints, or reward-based fine-tuning. These works treat the LLM as a black box and overlook a crucial question: how does the LLM internally integrate retrieved evidence with its parametric memory, particularly under knowledge conflicts? To address this gap, we conduct a probing-based analysis of hidden-state representations in LLMs and observe three findings: knowledge integration occurs hierarchically, conflicts manifest as latent signals at the sentence level, and irrelevant context is often amplified when aligned with parametric knowledge. Building on these findings, we propose CLEAR (Conflict-Localized and Enhanced Attention for RAG), a framework that (i) decomposes context into fine-grained sentence-level knowledge, (ii) employs hidden-state probing to localize conflicting knowledge, and (iii) introduces conflict-aware fine-tuning to guide the model to accurately integrate retrieved evidence. Extensive experiments across three benchmarks demonstrate that CLEAR substantially improves both accuracy and contextual faithfulness, consistently outperforming strong baselines under diverse conflict conditions. The related resources are available at https://github.com/LinfengGao/CLEAR.

Retrieval-augmented generation (RAG) grounds large language model (LLM) output by leveraging external knowledge sources to reduce factual hallucinations. However, prior works lack a comprehensive evaluation of different language families, making it challenging to evaluate LLM robustness against errors in external retrieved knowledge. To overcome this, we establish NoMIRACL, a human-annotated dataset for evaluating LLM robustness in RAG across 18 typologically diverse languages. NoMIRACL includes both a non-relevant and a relevant subset. Queries in the non-relevant subset contain passages manually judged as non-relevant or noisy, whereas queries in the relevant subset include at least a single judged relevant passage. We measure LLM robustness using two metrics: (i) hallucination rate, measuring model tendency to hallucinate an answer, when the answer is not present in passages in the non-relevant subset, and (ii) error rate, measuring model inaccuracy to recognize relevant passages in the relevant subset. We build a GPT-4 baseline which achieves a 33.2% hallucination rate on the non-relevant and a 14.9% error rate on the relevant subset on average. Our evaluation reveals that GPT-4 hallucinates frequently in high-resource languages, such as French or English. This work highlights an important avenue for future research to improve LLM robustness to learn how to better reject non-relevant information in RAG.

Online social media platforms are central to everyday communication and information seeking. While these platforms serve positive purposes, they also provide fertile ground for the spread of hate speech, offensive language, and bullying content targeting individuals, organizations, and communities. Such content undermines safety, participation, and equity online. Reliable detection systems are therefore needed, especially for low-resource languages where moderation tools are limited. In Bangla, prior work has contributed resources and models, but most are single-task (e.g., binary hate/offense) with limited coverage of multi-facet signals (type, severity, target). We address these gaps by introducing the first multi-task Bangla hate-speech dataset, BanglaMultiHate, one of the largest manually annotated corpus to date. Building on this resource, we conduct a comprehensive, controlled comparison spanning classical baselines, monolingual pretrained models, and LLMs under zero-shot prompting and LoRA fine-tuning. Our experiments assess LLM adaptability in a low-resource setting and reveal a consistent trend: although LoRA-tuned LLMs are competitive with BanglaBERT, culturally and linguistically grounded pretraining remains critical for robust performance. Together, our dataset and findings establish a stronger benchmark for developing culturally aligned moderation tools in low-resource contexts. For reproducibility, we will release the dataset and all related scripts.

Automatic evaluation metrics are a crucial component of dialog systems research. Standard language evaluation metrics are known to be ineffective for evaluating dialog. As such, recent research has proposed a number of novel, dialog-specific metrics that correlate better with human judgements. Due to the fast pace of research, many of these metrics have been assessed on different datasets and there has as yet been no time for a systematic comparison between them. To this end, this paper provides a comprehensive assessment of recently proposed dialog evaluation metrics on a number of datasets. In this paper, 23 different automatic evaluation metrics are evaluated on 10 different datasets. Furthermore, the metrics are assessed in different settings, to better qualify their respective strengths and weaknesses. Metrics are assessed (1) on both the turn level and the dialog level, (2) for different dialog lengths, (3) for different dialog qualities (e.g., coherence, engaging), (4) for different types of response generation models (i.e., generative, retrieval, simple models and state-of-the-art models), (5) taking into account the similarity of different metrics and (6) exploring combinations of different metrics. This comprehensive assessment offers several takeaways pertaining to dialog evaluation metrics in general. It also suggests how to best assess evaluation metrics and indicates promising directions for future work.

Offline evaluation of search systems depends on test collections. These benchmarks provide the researchers with a corpus of documents, topics and relevance judgements indicating which documents are relevant for each topic. While test collections are an integral part of Information Retrieval (IR) research, their creation involves significant efforts in manual annotation. Large language models (LLMs) are gaining much attention as tools for automatic relevance assessment. Recent research has shown that LLM-based assessments yield high systems ranking correlation with human-made judgements. These correlations are helpful in large-scale experiments but less informative if we want to focus on top-performing systems. Moreover, these correlations ignore whether and how LLM-based judgements impact the statistically significant differences among systems with respect to human assessments. In this work, we look at how LLM-generated judgements preserve ranking differences among top-performing systems and also how they preserve pairwise significance evaluation as human judgements. Our results show that LLM-based judgements are unfair at ranking top-performing systems. Moreover, we observe an exceedingly high rate of false positives regarding statistical differences. Our work represents a step forward in the evaluation of the reliability of using LLMs-based judgements for IR evaluation. We hope this will serve as a basis for other researchers to develop more reliable models for automatic relevance assessment.

As national security institutions increasingly integrate Artificial Intelligence (AI) into decision-making and content generation processes, understanding the inherent biases of large language models (LLMs) is crucial. This study presents a novel benchmark designed to evaluate the biases and preferences of seven prominent foundation models-Llama 3.1 8B Instruct, Llama 3.1 70B Instruct, GPT-4o, Gemini 1.5 Pro-002, Mixtral 8x22B, Claude 3.5 Sonnet, and Qwen2 72B-in the context of international relations (IR). We designed a bias discovery study around core topics in IR using 400-expert crafted scenarios to analyze results from our selected models. These scenarios focused on four topical domains including: military escalation, military and humanitarian intervention, cooperative behavior in the international system, and alliance dynamics. Our analysis reveals noteworthy variation among model recommendations based on scenarios designed for the four tested domains. Particularly, Qwen2 72B, Gemini 1.5 Pro-002 and Llama 3.1 8B Instruct models offered significantly more escalatory recommendations than Claude 3.5 Sonnet and GPT-4o models. All models exhibit some degree of country-specific biases, often recommending less escalatory and interventionist actions for China and Russia compared to the United States and the United Kingdom. These findings highlight the necessity for controlled deployment of LLMs in high-stakes environments, emphasizing the need for domain-specific evaluations and model fine-tuning to align with institutional objectives.

Counterspeech -- speech that opposes hate speech -- has gained significant attention recently as a strategy to reduce hate on social media. While previous studies suggest that counterspeech can somewhat reduce hate speech, little is known about its effects on participation in online hate communities, nor which counterspeech tactics reduce harmful behavior. We begin to address these gaps by identifying 25 large hate communities ("subreddits") within Reddit and analyzing the effect of counterspeech on newcomers within these communities. We first construct a new public dataset of carefully annotated counterspeech and non-counterspeech comments within these subreddits. We use this dataset to train a state-of-the-art counterspeech detection model. Next, we use matching to evaluate the causal effects of hostile and non-hostile counterspeech on the engagement of newcomers in hate subreddits. We find that, while non-hostile counterspeech is ineffective at keeping users from fully disengaging from these hate subreddits, a single hostile counterspeech comment substantially reduces both future likelihood of engagement. While offering nuance to the understanding of counterspeech efficacy, these results a) leave unanswered the question of whether hostile counterspeech dissuades newcomers from participation in online hate writ large, or merely drives them into less-moderated and more extreme hate communities, and b) raises ethical considerations about hostile counterspeech, which is both comparatively common and might exacerbate rather than mitigate the net level of antagonism in society. These findings underscore the importance of future work to improve counterspeech tactics and minimize unintended harm.

The proliferation of bias and propaganda on social media is an increasingly significant concern, leading to the development of techniques for automatic detection. This article presents a multilingual corpus of 12, 000 Facebook posts fully annotated for bias and propaganda. The corpus was created as part of the FigNews 2024 Shared Task on News Media Narratives for framing the Israeli War on Gaza. It covers various events during the War from October 7, 2023 to January 31, 2024. The corpus comprises 12, 000 posts in five languages (Arabic, Hebrew, English, French, and Hindi), with 2, 400 posts for each language. The annotation process involved 10 graduate students specializing in Law. The Inter-Annotator Agreement (IAA) was used to evaluate the annotations of the corpus, with an average IAA of 80.8% for bias and 70.15% for propaganda annotations. Our team was ranked among the bestperforming teams in both Bias and Propaganda subtasks. The corpus is open-source and available at https://sina.birzeit.edu/fada

We present NewsQA, a challenging machine comprehension dataset of over 100,000 human-generated question-answer pairs. Crowdworkers supply questions and answers based on a set of over 10,000 news articles from CNN, with answers consisting of spans of text from the corresponding articles. We collect this dataset through a four-stage process designed to solicit exploratory questions that require reasoning. A thorough analysis confirms that NewsQA demands abilities beyond simple word matching and recognizing textual entailment. We measure human performance on the dataset and compare it to several strong neural models. The performance gap between humans and machines (0.198 in F1) indicates that significant progress can be made on NewsQA through future research. The dataset is freely available at https://datasets.maluuba.com/NewsQA.

Large Language Models (LLMs) have recently been shown to be effective as automatic evaluators with simple prompting and in-context learning. In this work, we assemble 15 LLMs of four different size ranges and evaluate their output responses by preference ranking from the other LLMs as evaluators, such as System Star is better than System Square. We then evaluate the quality of ranking outputs introducing the Cognitive Bias Benchmark for LLMs as Evaluators (CoBBLEr), a benchmark to measure six different cognitive biases in LLM evaluation outputs, such as the Egocentric bias where a model prefers to rank its own outputs highly in evaluation. We find that LLMs are biased text quality evaluators, exhibiting strong indications on our bias benchmark (average of 40% of comparisons across all models) within each of their evaluations that question their robustness as evaluators. Furthermore, we examine the correlation between human and machine preferences and calculate the average Rank-Biased Overlap (RBO) score to be 49.6%, indicating that machine preferences are misaligned with humans. According to our findings, LLMs may still be unable to be utilized for automatic annotation aligned with human preferences. Our project page is at: https://minnesotanlp.github.io/cobbler.

Prompt injection attacks pose a significant challenge to the safe deployment of Large Language Models (LLMs) in real-world applications. While prompt-based detection offers a lightweight and interpretable defense strategy, its effectiveness has been hindered by the need for manual prompt engineering. To address this issue, we propose AEGIS , an Automated co-Evolutionary framework for Guarding prompt Injections Schema. Both attack and defense prompts are iteratively optimized against each other using a gradient-like natural language prompt optimization technique. This framework enables both attackers and defenders to autonomously evolve via a Textual Gradient Optimization (TGO) module, leveraging feedback from an LLM-guided evaluation loop. We evaluate our system on a real-world assignment grading dataset of prompt injection attacks and demonstrate that our method consistently outperforms existing baselines, achieving superior robustness in both attack success and detection. Specifically, the attack success rate (ASR) reaches 1.0, representing an improvement of 0.26 over the baseline. For detection, the true positive rate (TPR) improves by 0.23 compared to the previous best work, reaching 0.84, and the true negative rate (TNR) remains comparable at 0.89. Ablation studies confirm the importance of co-evolution, gradient buffering, and multi-objective optimization. We also confirm that this framework is effective in different LLMs. Our results highlight the promise of adversarial training as a scalable and effective approach for guarding prompt injections.

Frontier coding agents solve complex tasks when given complete context but collapse when specifications are incomplete or ambiguous. The bottleneck is not raw capability, but judgment: knowing when to act autonomously and when to ask for help. Current benchmarks are blind to this failure mode. They supply unambiguous detailed instructions and solely reward execution correctness, so an agent that makes a lucky guess for a missing requirement will score identically to one that would have asked to be certain. We present HiL-Bench (Human-in-the-Loop Benchmark) to measure this selective escalation skill. Each task contains human-validated blockers (missing information, ambiguous requests, contradictory information) that surface only through progressive exploration, not upfront inspection. Our core metric, Ask-F1, the harmonic mean of question precision and blocker recall, captures the tension between over-asking and silent guessing; its structure architecturally prevents gaming through question spam. Evaluation across SWE and text-to-SQL domains reveals a large universal judgment gap: no frontier model recovers more than a fraction of its full-information performance when deciding whether to ask. Failure analysis identifies three key help-seeking patterns: overconfident wrong beliefs with no gap detection; high uncertainty detection yet persistent errors; broad, imprecise escalation without self-correction. These consistent patterns confirm poor help-seeking is a model-level flaw, not task-specific. RL training on shaped Ask-F1 reward shows judgment is trainable: a 32B model improves both help-seeking quality and task pass rate, with gains that transfer across domains. The model does not learn domain-specific heuristics for when to ask; it learns to detect unresolvable uncertainty and act on it.

With the recent appearance of LLMs in practical settings, having methods that can effectively detect factual inconsistencies is crucial to reduce the propagation of misinformation and improve trust in model outputs. When testing on existing factual consistency benchmarks, we find that a few large language models (LLMs) perform competitively on classification benchmarks for factual inconsistency detection compared to traditional non-LLM methods. However, a closer analysis reveals that most LLMs fail on more complex formulations of the task and exposes issues with existing evaluation benchmarks, affecting evaluation precision. To address this, we propose a new protocol for inconsistency detection benchmark creation and implement it in a 10-domain benchmark called SummEdits. This new benchmark is 20 times more cost-effective per sample than previous benchmarks and highly reproducible, as we estimate inter-annotator agreement at about 0.9. Most LLMs struggle on SummEdits, with performance close to random chance. The best-performing model, GPT-4, is still 8\% below estimated human performance, highlighting the gaps in LLMs' ability to reason about facts and detect inconsistencies when they occur.

Large language models LLMs like ChatGPT have reached the 100 Mio user barrier in record time and might increasingly enter all areas of our life leading to a diverse set of interactions between those Artificial Intelligence models and humans. While many studies have discussed governance and regulations deductively from first-order principles, few studies provide an inductive, data-driven lens based on observing dialogues between humans and LLMs especially when it comes to non-collaborative, competitive situations that have the potential to pose a serious threat to people. In this work, we conduct a user study engaging over 40 individuals across all age groups in price negotiations with an LLM. We explore how people interact with an LLM, investigating differences in negotiation outcomes and strategies. Furthermore, we highlight shortcomings of LLMs with respect to their reasoning capabilities and, in turn, susceptiveness to prompt hacking, which intends to manipulate the LLM to make agreements that are against its instructions or beyond any rationality. We also show that the negotiated prices humans manage to achieve span a broad range, which points to a literacy gap in effectively interacting with LLMs.

To make AI systems broadly useful for challenging real-world tasks, we need them to learn complex human goals and preferences. One approach to specifying complex goals asks humans to judge during training which agent behaviors are safe and useful, but this approach can fail if the task is too complicated for a human to directly judge. To help address this concern, we propose training agents via self play on a zero sum debate game. Given a question or proposed action, two agents take turns making short statements up to a limit, then a human judges which of the agents gave the most true, useful information. In an analogy to complexity theory, debate with optimal play can answer any question in PSPACE given polynomial time judges (direct judging answers only NP questions). In practice, whether debate works involves empirical questions about humans and the tasks we want AIs to perform, plus theoretical questions about the meaning of AI alignment. We report results on an initial MNIST experiment where agents compete to convince a sparse classifier, boosting the classifier's accuracy from 59.4% to 88.9% given 6 pixels and from 48.2% to 85.2% given 4 pixels. Finally, we discuss theoretical and practical aspects of the debate model, focusing on potential weaknesses as the model scales up, and we propose future human and computer experiments to test these properties.

Word error rate (WER) is the dominant metric for automatic speech recognition, yet it cannot detect a systematic failure mode: models that produce fluent output in the wrong writing system. We define Script Fidelity Rate (SFR), the fraction of hypothesis characters in the target script block, computable without reference transcriptions, and report the first systematic measurement of script collapse across six languages spanning four writing systems (Pashto, Urdu, Hindi, Bengali, Malayalam, Somali) and nine ASR models on FLEURS test sets. Across 53 evaluated model-language pairs, 18 (34%; 95% Wilson CI: 23-47%) exhibit script collapse (SFR < 10%); MMS-1B and SeamlessM4T-v2 maintain SFR above 99% on every language evaluated, confirming that SFR correctly identifies high fidelity where it is present. We identify three distinct collapse patterns: Latin phonetic substitution (smaller Whisper on Indic languages), Arabic substitution for Somali's Latin-script orthography, and Devanagari substitution where larger Whisper models treat all Indic audio as Hindi, a failure present even in Whisper large-v3.

Language Model (LM)-based agents remain largely untested in mixed-motive settings where agents must leverage short-term cooperation for long-term competitive goals (e.g., multi-party politics). We introduce Cooperate to Compete (C2C), a multi-agent environment where players can engage in private negotiations while competing to be the first to achieve their secret objective. Players have asymmetric objectives and negotiations are non-binding, allowing alliances to form and break as players' short-term interests align and diverge. We run AI only games and conduct a user study pitting human players against AI opponents. We identify significant differences between human and AI negotiation behaviors, finding that humans favor lower-complexity deals and are significantly less reliable partners compared to LM-based agents. We also find that humans are more aggressive negotiators, accepting deals without a counteroffer only 56.3% of the time compared to 67.6% for LM-based agents. Through targeted prompting inspired by these findings, we modify agents' negotiation behavior and improve win rates from 22.2% to 32.7%. We run over 1,100 games with over 16,000 private conversations totaling 15.2 million tokens and over 150,000 player actions. Our results establish C2C as a testbed for studying and building LM-based agents that can navigate the sophisticated coordination required for real-world deployments. The game, code, and dataset may be found at https://negotiationgame.io/c2c.

As Large Language Models (LLMs) are increasingly deployed as autonomous agents in complex and long horizon settings, it is critical to evaluate their ability to sabotage users by pursuing hidden objectives. We study the ability of frontier LLMs to evade monitoring and achieve harmful hidden goals while completing a wide array of realistic tasks. We evaluate a broad range of frontier LLMs using SHADE (Subtle Harmful Agent Detection & Evaluation)-Arena, the first highly diverse agent evaluation dataset for sabotage and monitoring capabilities of LLM agents. SHADE-Arena consists of complex pairs of benign main tasks and harmful side objectives in complicated environments. Agents are evaluated on their ability to complete the side task without appearing suspicious to an LLM monitor. When measuring agent ability to (a) complete the main task, (b) complete the side task, and (c) avoid detection, we find that the best performing frontier models score 27% (Claude 3.7 Sonnet) and 15% (Gemini 2.5 Pro) as sabotage agents when overseen by Claude 3.6 Sonnet. For current frontier models, success on the side task relies heavily on having access to a hidden scratchpad that is not visible to the monitor. We also use SHADE-Arena to measure models' monitoring abilities, with the top monitor (Gemini 2.5 Pro) achieving an AUC of 0.87 at distinguishing benign and malign transcripts. We find that for now, models still struggle at sabotage due to failures in long-context main task execution. However, our measurements already demonstrate the difficulty of monitoring for subtle sabotage attempts, which we expect to only increase in the face of more complex and longer-horizon tasks.

Gun violence is a pressing and growing human rights issue that affects nearly every dimension of the social fabric, from healthcare and education to psychology and the economy. Reliable data on firearm events is paramount to developing more effective public policy and emergency responses. However, the lack of comprehensive databases and the risks of in-person surveys prevent human rights organizations from collecting needed data in most countries. Here, we partner with a Brazilian human rights organization to conduct a systematic evaluation of language models to assist with monitoring real-world firearm events from social media data. We propose a fine-tuned BERT-based model trained on Twitter (now X) texts to distinguish gun violence reports from ordinary Portuguese texts. Our model achieves a high AUC score of 0.97. We then incorporate our model into a web application and test it in a live intervention. We study and interview Brazilian analysts who continuously fact-check social media texts to identify new gun violence events. Qualitative assessments show that our solution helped all analysts use their time more efficiently and expanded their search capacities. Quantitative assessments show that the use of our model was associated with more analysts' interactions with online users reporting gun violence. Taken together, our findings suggest that modern Natural Language Processing techniques can help support the work of human rights organizations.

This work proposes a new challenge set for multimodal classification, focusing on detecting hate speech in multimodal memes. It is constructed such that unimodal models struggle and only multimodal models can succeed: difficult examples ("benign confounders") are added to the dataset to make it hard to rely on unimodal signals. The task requires subtle reasoning, yet is straightforward to evaluate as a binary classification problem. We provide baseline performance numbers for unimodal models, as well as for multimodal models with various degrees of sophistication. We find that state-of-the-art methods perform poorly compared to humans (64.73% vs. 84.7% accuracy), illustrating the difficulty of the task and highlighting the challenge that this important problem poses to the community.

Automated news verification requires structured claim extraction, but existing approaches either lack schema compliance or generalize poorly across domains. This paper presents NewsScope, a cross-domain dataset, benchmark, and fine-tuned model for schema-grounded news claim extraction. The dataset contains 455 articles across politics, health, science/environment, and business, consisting of 395 in-domain articles and 60 out-of-source articles for generalization testing. LLaMA 3.1 8B was fine-tuned using LoRA on 315 training examples and evaluated on held-out in-domain (80 articles) and out-of-source (60 articles) test sets. Human evaluation on 400 claims shows NewsScope achieves 89.4% human-evaluated accuracy compared to GPT-4o-mini's 93.7% (p=0.07). NewsScope outperforms GPT-4o-mini on political claims (94.3% vs. 87.8%). A numeric grounding filter further improves accuracy to 91.6%, narrowing the gap to 2.1 percentage points. Inter-annotator agreement studies (160 claims) confirm labeling reliability (94.6% positive agreement on SUPPORTED judgments). The open-weight model enables offline deployment at approximately 15 on-demand compute (or 0 on free tiers). Code and benchmark are publicly released.

Detecting hateful content is a challenging and important problem. Automated tools, like machine-learning models, can help, but they require continuous training to adapt to the ever-changing landscape of social media. In this work, we evaluate eight open-source LLMs' capability to detect antisemitic content, specifically leveraging in-context definition as a policy guideline. We explore various prompting techniques and design a new CoT-like prompt, Guided-CoT. Guided-CoT handles the in-context policy well, increasing performance across all evaluated models, regardless of decoding configuration, model sizes, or reasoning capability. Notably, Llama 3.1 70B outperforms fine-tuned GPT-3.5. Additionally, we examine LLM errors and introduce metrics to quantify semantic divergence in model-generated rationales, revealing notable differences and paradoxical behaviors among LLMs. Our experiments highlight the differences observed across LLMs' utility, explainability, and reliability.

We present R-Debater, an agentic framework for generating multi-turn debates built on argumentative memory. Grounded in rhetoric and memory studies, the system views debate as a process of recalling and adapting prior arguments to maintain stance consistency, respond to opponents, and support claims with evidence. Specifically, R-Debater integrates a debate knowledge base for retrieving case-like evidence and prior debate moves with a role-based agent that composes coherent utterances across turns. We evaluate on standardized ORCHID debates, constructing a 1,000-item retrieval corpus and a held-out set of 32 debates across seven domains. Two tasks are evaluated: next-utterance generation, assessed by InspireScore (subjective, logical, and factual), and adversarial multi-turn simulations, judged by Debatrix (argument, source, language, and overall). Compared with strong LLM baselines, R-Debater achieves higher single-turn and multi-turn scores. Human evaluation with 20 experienced debaters further confirms its consistency and evidence use, showing that combining retrieval grounding with structured planning yields more faithful, stance-aligned, and coherent debates across turns.

In this research, we introduce BEATS, a novel framework for evaluating Bias, Ethics, Fairness, and Factuality in Large Language Models (LLMs). Building upon the BEATS framework, we present a bias benchmark for LLMs that measure performance across 29 distinct metrics. These metrics span a broad range of characteristics, including demographic, cognitive, and social biases, as well as measures of ethical reasoning, group fairness, and factuality related misinformation risk. These metrics enable a quantitative assessment of the extent to which LLM generated responses may perpetuate societal prejudices that reinforce or expand systemic inequities. To achieve a high score on this benchmark a LLM must show very equitable behavior in their responses, making it a rigorous standard for responsible AI evaluation. Empirical results based on data from our experiment show that, 37.65\% of outputs generated by industry leading models contained some form of bias, highlighting a substantial risk of using these models in critical decision making systems. BEATS framework and benchmark offer a scalable and statistically rigorous methodology to benchmark LLMs, diagnose factors driving biases, and develop mitigation strategies. With the BEATS framework, our goal is to help the development of more socially responsible and ethically aligned AI models.

The RVL-CDIP benchmark is widely used for measuring performance on the task of document classification. Despite its widespread use, we reveal several undesirable characteristics of the RVL-CDIP benchmark. These include (1) substantial amounts of label noise, which we estimate to be 8.1% (ranging between 1.6% to 16.9% per document category); (2) presence of many ambiguous or multi-label documents; (3) a large overlap between test and train splits, which can inflate model performance metrics; and (4) presence of sensitive personally-identifiable information like US Social Security numbers (SSNs). We argue that there is a risk in using RVL-CDIP for benchmarking document classifiers, as its limited scope, presence of errors (state-of-the-art models now achieve accuracy error rates that are within our estimated label error rate), and lack of diversity make it less than ideal for benchmarking. We further advocate for the creation of a new document classification benchmark, and provide recommendations for what characteristics such a resource should include.

Misinformation and disinformation demand fact checking that goes beyond simple evidence-based reasoning. Existing benchmarks fall short: they are largely single modality (text-only), span short time horizons, use shallow evidence, cover domains unevenly, and often omit full articles -- obscuring models' real-world capability. We present MMM-Fact, a large-scale benchmark of 125,449 fact-checked statements (1995--2025) across multiple domains, each paired with the full fact-check article and multimodal evidence (text, images, videos, tables) from four fact-checking sites and one news outlet. To reflect verification effort, each statement is tagged with a retrieval-difficulty tier -- Basic (1--5 sources), Intermediate (6--10), and Advanced (>10) -- supporting fairness-aware evaluation for multi-step, cross-modal reasoning. The dataset adopts a three-class veracity scheme (true/false/not enough information) and enables tasks in veracity prediction, explainable fact-checking, complex evidence aggregation, and longitudinal analysis. Baselines with mainstream LLMs show MMM-Fact is markedly harder than prior resources, with performance degrading as evidence complexity rises. MMM-Fact offers a realistic, scalable benchmark for transparent, reliable, multimodal fact-checking.

The rapid advancement and deployment of AI systems have created an urgent need for standard safety-evaluation frameworks. This paper introduces AILuminate v1.0, the first comprehensive industry-standard benchmark for assessing AI-product risk and reliability. Its development employed an open process that included participants from multiple fields. The benchmark evaluates an AI system's resistance to prompts designed to elicit dangerous, illegal, or undesirable behavior in 12 hazard categories, including violent crimes, nonviolent crimes, sex-related crimes, child sexual exploitation, indiscriminate weapons, suicide and self-harm, intellectual property, privacy, defamation, hate, sexual content, and specialized advice (election, financial, health, legal). Our method incorporates a complete assessment standard, extensive prompt datasets, a novel evaluation framework, a grading and reporting system, and the technical as well as organizational infrastructure for long-term support and evolution. In particular, the benchmark employs an understandable five-tier grading scale (Poor to Excellent) and incorporates an innovative entropy-based system-response evaluation. In addition to unveiling the benchmark, this report also identifies limitations of our method and of building safety benchmarks generally, including evaluator uncertainty and the constraints of single-turn interactions. This work represents a crucial step toward establishing global standards for AI risk and reliability evaluation while acknowledging the need for continued development in areas such as multiturn interactions, multimodal understanding, coverage of additional languages, and emerging hazard categories. Our findings provide valuable insights for model developers, system integrators, and policymakers working to promote safer AI deployment.

Hate speech detection is a crucial task, especially on social media, where harmful content can spread quickly. Implementing machine learning models to automatically identify and address hate speech is essential for mitigating its impact and preventing its proliferation. The first step in developing an effective hate speech detection model is to acquire a high-quality dataset for training. Labeled data is foundational for most natural language processing tasks, but categorizing hate speech is difficult due to the diverse and often subjective nature of hate speech, which can lead to varying interpretations and disagreements among annotators. This paper examines strategies for addressing annotator disagreement, an issue that has been largely overlooked. In particular, we evaluate different approaches to deal with annotator disagreement regarding hate speech classification in Turkish tweets, based on a fine-tuned BERT model. Our work highlights the importance of the problem and provides state-of-art benchmark results for detection and understanding of hate speech in online discourse.

With the growing research focus on multimodal dialogue systems, the capability for proactive interaction is gradually gaining recognition. As an alternative to conventional turn-by-turn dialogue, users increasingly expect multimodal systems to be more initiative, for example, by autonomously determining the timing of multi-turn responses in real time during video playback. To facilitate progress in this emerging area, we introduce ProactiveBench, the first comprehensive benchmark to evaluate a system's ability to engage in proactive interaction. Since model responses are generated at varying timestamps, we further propose PAUC, the first metric that accounts for the temporal dynamics of model responses. This enables a more accurate evaluation of systems operating in proactive settings. Through extensive benchmarking of various baseline systems on ProactiveBench and a user study of human preferences, we show that PAUC is in better agreement with human preferences than traditional evaluation metrics, which typically only consider the textual content of responses. These findings demonstrate that PAUC provides a more faithful assessment of user experience in proactive interaction scenarios. Project homepage: https://github.com/yellow-binary-tree/ProactiveBench

For socially sensitive tasks like hate speech detection, the quality of explanations from Large Language Models (LLMs) is crucial for factors like user trust and model alignment. While Persona prompting (PP) is increasingly used as a way to steer model towards user-specific generation, its effect on model rationales remains underexplored. We investigate how LLM-generated rationales vary when conditioned on different simulated demographic personas. Using datasets annotated with word-level rationales, we measure agreement with human annotations from different demographic groups, and assess the impact of PP on model bias and human alignment. Our evaluation across three LLMs results reveals three key findings: (1) PP improving classification on the most subjective task (hate speech) but degrading rationale quality. (2) Simulated personas fail to align with their real-world demographic counterparts, and high inter-persona agreement shows models are resistant to significant steering. (3) Models exhibit consistent demographic biases and a strong tendency to over-flag content as harmful, regardless of PP. Our findings reveal a critical trade-off: while PP can improve classification in socially-sensitive tasks, it often comes at the cost of rationale quality and fails to mitigate underlying biases, urging caution in its application.

With the continuing spread of misinformation and disinformation online, it is of increasing importance to develop combating mechanisms at scale in the form of automated systems that support multiple languages. One task of interest is claim veracity prediction, which can be addressed using stance detection with respect to relevant documents retrieved online. To this end, we present our new Arabic Stance Detection dataset (AraStance) of 4,063 claim--article pairs from a diverse set of sources comprising three fact-checking websites and one news website. AraStance covers false and true claims from multiple domains (e.g., politics, sports, health) and several Arab countries, and it is well-balanced between related and unrelated documents with respect to the claims. We benchmark AraStance, along with two other stance detection datasets, using a number of BERT-based models. Our best model achieves an accuracy of 85\% and a macro F1 score of 78\%, which leaves room for improvement and reflects the challenging nature of AraStance and the task of stance detection in general.

Though chatbots based on large neural models can often produce fluent responses in open domain conversations, one salient error type is contradiction or inconsistency with the preceding conversation turns. Previous work has treated contradiction detection in bot responses as a task similar to natural language inference, e.g., detect the contradiction between a pair of bot utterances. However, utterances in conversations may contain co-references or ellipsis, and using these utterances as is may not always be sufficient for identifying contradictions. This work aims to improve the contradiction detection via rewriting all bot utterances to restore antecedents and ellipsis. We curated a new dataset for utterance rewriting and built a rewriting model on it. We empirically demonstrate that this model can produce satisfactory rewrites to make bot utterances more complete. Furthermore, using rewritten utterances improves contradiction detection performance significantly, e.g., the AUPR and joint accuracy scores (detecting contradiction along with evidence) increase by 6.5% and 4.5% (absolute increase), respectively.

Knowledge-dependent tasks typically use two sources of knowledge: parametric, learned at training time, and contextual, given as a passage at inference time. To understand how models use these sources together, we formalize the problem of knowledge conflicts, where the contextual information contradicts the learned information. Analyzing the behaviour of popular models, we measure their over-reliance on memorized information (the cause of hallucinations), and uncover important factors that exacerbate this behaviour. Lastly, we propose a simple method to mitigate over-reliance on parametric knowledge, which minimizes hallucination, and improves out-of-distribution generalization by 4%-7%. Our findings demonstrate the importance for practitioners to evaluate model tendency to hallucinate rather than read, and show that our mitigation strategy encourages generalization to evolving information (i.e., time-dependent queries). To encourage these practices, we have released our framework for generating knowledge conflicts.

Psychological defenses are strategies, often automatic, that people use to manage distress. Rigid or overuse of defenses is negatively linked to mental health and shapes what speakers disclose and how they accept or resist help. However, defenses are complex and difficult to reliably measure, particularly in clinical dialogues. We introduce PsyDefConv, a dialogue corpus with help seeker utterances labeled for defense level, and DMRS Co-Pilot, a four-stage pipeline that provides evidence-based pre-annotations. The corpus contains 200 dialogues and 4709 utterances, including 2336 help seeker turns, with labeling and Cohen's kappa 0.639. In a counterbalanced study, the co-pilot reduced average annotation time by 22.4%. In expert review, it averaged 4.62 for evidence, 4.44 for clinical plausibility, and 4.40 for insight on a seven-point scale. Benchmarks with strong language models in zero-shot and fine-tuning settings demonstrate clear headroom, with the best macro F1-score around 30% and a tendency to overpredict mature defenses. Corpus analyses confirm that mature defenses are most common and reveal emotion-specific deviations. We will release the corpus, annotations, code, and prompts to support research on defensive functioning in language.

We introduce Debate Speech Evaluation as a novel and challenging benchmark for assessing LLM judges. Evaluating debate speeches requires a deep understanding of the speech at multiple levels, including argument strength and relevance, the coherence and organization of the speech, the appropriateness of its style and tone, and so on. This task involves a unique set of cognitive abilities that have previously received limited attention in systematic LLM benchmarking. To explore such skills, we leverage a dataset of over 600 meticulously annotated debate speeches and present the first in-depth analysis of how state-of-the-art LLMs compare to human judges on this task. Our findings reveal a nuanced picture: while larger models can approximate individual human judgments in some respects, they differ substantially in their overall judgment behavior. We also investigate the ability of frontier LLMs to generate persuasive, opinionated speeches, showing that models may perform at a human level on this task.

Hate speech poses a significant threat to social harmony. Over the past two years, Indonesia has seen a ten-fold increase in the online hate speech ratio, underscoring the urgent need for effective detection mechanisms. However, progress is hindered by the limited availability of labeled data for Indonesian texts. The condition is even worse for marginalized minorities, such as Shia, LGBTQ, and other ethnic minorities because hate speech is underreported and less understood by detection tools. Furthermore, the lack of accommodation for subjectivity in current datasets compounds this issue. To address this, we introduce IndoToxic2024, a comprehensive Indonesian hate speech and toxicity classification dataset. Comprising 43,692 entries annotated by 19 diverse individuals, the dataset focuses on texts targeting vulnerable groups in Indonesia, specifically during the hottest political event in the country: the presidential election. We establish baselines for seven binary classification tasks, achieving a macro-F1 score of 0.78 with a BERT model (IndoBERTweet) fine-tuned for hate speech classification. Furthermore, we demonstrate how incorporating demographic information can enhance the zero-shot performance of the large language model, gpt-3.5-turbo. However, we also caution that an overemphasis on demographic information can negatively impact the fine-tuned model performance due to data fragmentation.

In this paper, we present a novel hostility detection dataset in Hindi language. We collect and manually annotate ~8200 online posts. The annotated dataset covers four hostility dimensions: fake news, hate speech, offensive, and defamation posts, along with a non-hostile label. The hostile posts are also considered for multi-label tags due to a significant overlap among the hostile classes. We release this dataset as part of the CONSTRAINT-2021 shared task on hostile post detection.