Daily Papers

We introduce VL-JEPA, a vision-language model built on a Joint Embedding Predictive Architecture (JEPA). Instead of autoregressively generating tokens as in classical VLMs, VL-JEPA predicts continuous embeddings of the target texts. By learning in an abstract representation space, the model focuses on task-relevant semantics while abstracting away surface-level linguistic variability. In a strictly controlled comparison against standard token-space VLM training with the same vision encoder and training data, VL-JEPA achieves stronger performance while having 50% fewer trainable parameters. At inference time, a lightweight text decoder is invoked only when needed to translate VL-JEPA predicted embeddings into text. We show that VL-JEPA natively supports selective decoding that reduces the number of decoding operations by 2.85x while maintaining similar performance compared to non-adaptive uniform decoding. Beyond generation, the VL-JEPA's embedding space naturally supports open-vocabulary classification, text-to-video retrieval, and discriminative VQA without any architecture modification. On eight video classification and eight video retrieval datasets, the average performance VL-JEPA surpasses that of CLIP, SigLIP2, and Perception Encoder. At the same time, the model achieves comparable performance as classical VLMs (InstructBLIP, QwenVL) on four VQA datasets: GQA, TallyQA, POPE and POPEv2, despite only having 1.6B parameters.

Recent progress in text-to-music generation has enabled models to synthesize high-quality musical segments, full compositions, and even respond to fine-grained control signals, e.g. chord progressions. State-of-the-art (SOTA) systems differ significantly across many dimensions, such as training datasets, modeling paradigms, and architectural choices. This diversity complicates efforts to evaluate models fairly and pinpoint which design choices most influence performance. While factors like data and architecture are important, in this study we focus exclusively on the modeling paradigm. We conduct a systematic empirical analysis to isolate its effects, offering insights into associated trade-offs and emergent behaviors that can guide future text-to-music generation systems. Specifically, we compare the two arguably most common modeling paradigms: Auto-Regressive decoding and Conditional Flow-Matching. We conduct a controlled comparison by training all models from scratch using identical datasets, training configurations, and similar backbone architectures. Performance is evaluated across multiple axes, including generation quality, robustness to inference configurations, scalability, adherence to both textual and temporally aligned conditioning, and editing capabilities in the form of audio inpainting. This comparative study sheds light on distinct strengths and limitations of each paradigm, providing actionable insights that can inform future architectural and training decisions in the evolving landscape of text-to-music generation. Audio sampled examples are available at: https://huggingface.co/spaces/ortal1602/ARvsFM

Sequential LLM agents fail on long-horizon planning with hard constraints like budgets and diversity requirements. As planning progresses and context grows, these agents drift from global constraints. We propose HiMAP-Travel, a hierarchical multi-agent framework that splits planning into strategic coordination and parallel day-level execution. A Coordinator allocates resources across days, while Day Executors plan independently in parallel. Three key mechanisms enable this: a transactional monitor enforcing budget and uniqueness constraints across parallel agents, a bargaining protocol allowing agents to reject infeasible sub-goals and trigger re-planning, and a single policy trained with GRPO that powers all agents through role conditioning. On TravelPlanner, HiMAP-Travel with Qwen3-8B achieves 52.78% validation and 52.65% test Final Pass Rate (FPR). In a controlled comparison with identical model, training, and tools, it outperforms the sequential DeepTravel baseline by +8.67~pp. It also surpasses ATLAS by +17.65~pp and MTP by +10.0~pp. On FlexTravelBench multi-turn scenarios, it achieves 44.34% (2-turn) and 37.42% (3-turn) FPR while reducing latency 2.5x through parallelization.

Real-world decision-making, from tax compliance assessment to medical diagnosis, requires aggregating multiple noisy and potentially contradictory evidence sources. Existing approaches either lack explicit uncertainty quantification (neural aggregation methods) or rely on manually engineered discrete predicates (probabilistic logic frameworks), limiting scalability to unstructured data. We introduce Latent Posterior Factors (LPF), a framework that transforms Variational Autoencoder (VAE) latent posteriors into soft likelihood factors for Sum-Product Network (SPN) inference, enabling tractable probabilistic reasoning over unstructured evidence while preserving calibrated uncertainty estimates. We instantiate LPF as LPF-SPN (structured factor-based inference) and LPF-Learned (end-to-end learned aggregation), enabling a principled comparison between explicit probabilistic reasoning and learned aggregation under a shared uncertainty representation. Across eight domains (seven synthetic and the FEVER benchmark), LPF-SPN achieves high accuracy (up to 97.8%), low calibration error (ECE 1.4%), and strong probabilistic fit, substantially outperforming evidential deep learning, LLMs and graph-based baselines over 15 random seeds. Contributions: (1) A framework bridging latent uncertainty representations with structured probabilistic reasoning. (2) Dual architectures enabling controlled comparison of reasoning paradigms. (3) Reproducible training methodology with seed selection. (4) Evaluation against EDL, BERT, R-GCN, and large language model baselines. (5) Cross-domain validation. (6) Formal guarantees in a companion paper.

The effort devoted to hand-crafting neural network image classifiers has motivated the use of architecture search to discover them automatically. Although evolutionary algorithms have been repeatedly applied to neural network topologies, the image classifiers thus discovered have remained inferior to human-crafted ones. Here, we evolve an image classifier---AmoebaNet-A---that surpasses hand-designs for the first time. To do this, we modify the tournament selection evolutionary algorithm by introducing an age property to favor the younger genotypes. Matching size, AmoebaNet-A has comparable accuracy to current state-of-the-art ImageNet models discovered with more complex architecture-search methods. Scaled to larger size, AmoebaNet-A sets a new state-of-the-art 83.9% / 96.6% top-5 ImageNet accuracy. In a controlled comparison against a well known reinforcement learning algorithm, we give evidence that evolution can obtain results faster with the same hardware, especially at the earlier stages of the search. This is relevant when fewer compute resources are available. Evolution is, thus, a simple method to effectively discover high-quality architectures.

We introduce MCUBench, a benchmark featuring over 100 YOLO-based object detection models evaluated on the VOC dataset across seven different MCUs. This benchmark provides detailed data on average precision, latency, RAM, and Flash usage for various input resolutions and YOLO-based one-stage detectors. By conducting a controlled comparison with a fixed training pipeline, we collect comprehensive performance metrics. Our Pareto-optimal analysis shows that integrating modern detection heads and training techniques allows various YOLO architectures, including legacy models like YOLOv3, to achieve a highly efficient tradeoff between mean Average Precision (mAP) and latency. MCUBench serves as a valuable tool for benchmarking the MCU performance of contemporary object detectors and aids in model selection based on specific constraints.

Instruction fine-tuning of large language models (LLMs) often involves selecting a subset of instruction training data from a large candidate pool, using a small query set from the target task. Despite growing interest, the literature on targeted instruction selection remains fragmented and opaque: methods vary widely in selection budgets, often omit zero-shot baselines, and frequently entangle the contributions of key components. As a result, practitioners lack actionable guidance on selecting instructions for their target tasks. In this work, we aim to bring clarity to this landscape by disentangling and systematically analyzing the two core ingredients: data representation and selection algorithms. Our framework enables controlled comparisons across models, tasks, and budgets. We find that only gradient-based data representations choose subsets whose similarity to the query consistently predicts performance across datasets and models. While no single method dominates, gradient-based representations paired with a greedy round-robin selection algorithm tend to perform best on average at low budgets, but these benefits diminish at larger budgets. Finally, we unify several existing selection algorithms as forms of approximate distance minimization between the selected subset and the query set, and support this view with new generalization bounds. More broadly, our findings provide critical insights and a foundation for more principled data selection in LLM fine-tuning. The code is available at https://github.com/dcml-lab/targeted-instruction-selection.

World Action Models (WAMs) have emerged as a promising alternative to Vision-Language-Action (VLA) models for embodied control because they explicitly model how visual observations may evolve under action. Most existing WAMs follow an imagine-then-execute paradigm, incurring substantial test-time latency from iterative video denoising, yet it remains unclear whether explicit future imagination is actually necessary for strong action performance. In this paper, we ask whether WAMs need explicit future imagination at test time, or whether their benefit comes primarily from video modeling during training. We disentangle the role of video modeling during training from explicit future generation during inference by proposing Fast-WAM, a WAM architecture that retains video co-training during training but skips future prediction at test time. We further instantiate several Fast-WAM variants to enable a controlled comparison of these two factors. Across these variants, we find that Fast-WAM remains competitive with imagine-then-execute variants, while removing video co-training causes a much larger performance drop. Empirically, Fast-WAM achieves competitive results with state-of-the-art methods both on simulation benchmarks (LIBERO and RoboTwin) and real-world tasks, without embodied pretraining. It runs in real time with 190ms latency, over 4times faster than existing imagine-then-execute WAMs. These results suggest that the main value of video prediction in WAMs may lie in improving world representations during training rather than generating future observations at test time. Project page: https://yuantianyuan01.github.io/FastWAM/

Contrastive pretraining on image-text pairs from the web is one of the most popular large-scale pretraining strategies for vision backbones, especially in the context of large multimodal models. At the same time, image captioning on this type of data is commonly considered an inferior pretraining strategy. In this paper, we perform a fair comparison of these two pretraining strategies, carefully matching training data, compute, and model capacity. Using a standard encoder-decoder transformer, we find that captioning alone is surprisingly effective: on classification tasks, captioning produces vision encoders competitive with contrastively pretrained encoders, while surpassing them on vision & language tasks. We further analyze the effect of the model architecture and scale, as well as the pretraining data on the representation quality, and find that captioning exhibits the same or better scaling behavior along these axes. Overall our results show that plain image captioning is a more powerful pretraining strategy than was previously believed.

We analyze automatic speech recognition (ASR) modeling choices under domain mismatch, comparing classic modular and novel sequence-to-sequence (seq2seq) architectures. Across the different ASR architectures, we examine a spectrum of modeling choices, including label units, context length, and topology. To isolate language domain effects from acoustic variation, we synthesize target domain audio using a text-to-speech system trained on LibriSpeech. We incorporate target domain n-gram and neural language models for domain adaptation without retraining the acoustic model. To our knowledge, this is the first controlled comparison of optimized ASR systems across state-of-the-art architectures under domain shift, offering insights into their generalization. The results show that, under domain shift, rather than the decoder architecture choice or the distinction between classic modular and novel seq2seq models, it is specific modeling choices that influence performance.

In this work, we provide a systematic and comprehensive empirical comparison of pretrained multilingual language models versus their monolingual counterparts with regard to their monolingual task performance. We study a set of nine typologically diverse languages with readily available pretrained monolingual models on a set of five diverse monolingual downstream tasks. We first aim to establish, via fair and controlled comparisons, if a gap between the multilingual and the corresponding monolingual representation of that language exists, and subsequently investigate the reason for any performance difference. To disentangle conflating factors, we train new monolingual models on the same data, with monolingually and multilingually trained tokenizers. We find that while the pretraining data size is an important factor, a designated monolingual tokenizer plays an equally important role in the downstream performance. Our results show that languages that are adequately represented in the multilingual model's vocabulary exhibit negligible performance decreases over their monolingual counterparts. We further find that replacing the original multilingual tokenizer with the specialized monolingual tokenizer improves the downstream performance of the multilingual model for almost every task and language.

LLM agents now perform strongly in software engineering, deep research, GUI automation, and various other applications, while recent agent scaffolds and models are increasingly integrating these capabilities into unified systems. Yet, most evaluations still test these capabilities in isolation, which leaves a gap for more diverse use cases that require agents to combine different capabilities. We introduce CocoaBench, a benchmark for unified digital agents built from human-designed, long-horizon tasks that require flexible composition of vision, search, and coding. Tasks are specified only by an instruction and an automatic evaluation function over the final output, enabling reliable and scalable evaluation across diverse agent infrastructures. We also present CocoaAgent, a lightweight shared scaffold for controlled comparison across model backbones. Experiments show that current agents remain far from reliable on CocoaBench, with the best evaluated system achieving only 45.1% success rate. Our analysis further points to substantial room for improvement in reasoning and planning, tool use and execution, and visual grounding.

The success of large language models has shifted the evaluation paradigms in natural language processing (NLP). The community's interest has drifted towards comparing NLP models across many tasks, domains, and datasets, often at an extreme scale. This imposes new engineering challenges: efforts in constructing datasets and models have been fragmented, and their formats and interfaces are incompatible. As a result, it often takes extensive (re)implementation efforts to make fair and controlled comparisons at scale. Catwalk aims to address these issues. Catwalk provides a unified interface to a broad range of existing NLP datasets and models, ranging from both canonical supervised training and fine-tuning, to more modern paradigms like in-context learning. Its carefully-designed abstractions allow for easy extensions to many others. Catwalk substantially lowers the barriers to conducting controlled experiments at scale. For example, we finetuned and evaluated over 64 models on over 86 datasets with a single command, without writing any code. Maintained by the AllenNLP team at the Allen Institute for Artificial Intelligence (AI2), Catwalk is an ongoing open-source effort: https://github.com/allenai/catwalk.

The integration of audio perception capabilities into Large Language Models (LLMs) has enabled significant advances in Audio-LLMs. Although application-focused developments, particularly in curating training data for specific capabilities e.g., audio reasoning, have progressed rapidly, the underlying mechanisms that govern efficient transfer of rich semantic representations from audio encoders to LLMs remain under-explored. We conceptualize effective audio-LLM interaction as the LLM's ability to proficiently probe the audio encoder representations to satisfy textual queries. This paper presents a systematic investigation on how architectural design choices can affect that. Beginning with a standard Pengi/LLaVA-style audio-LLM architecture, we propose and evaluate several modifications guided by hypotheses derived from mechanistic interpretability studies and LLM operational principles. Our experiments demonstrate that: (1) delaying audio integration until the LLM's initial layers establish textual context that enhances its ability to probe the audio representations for relevant information; (2) the LLM can proficiently probe audio representations exclusively through LLM layer's attention submodule, without requiring propagation to its Feed-Forward Network (FFN) submodule; (3) an efficiently integrated ensemble of diverse audio encoders provides richer, complementary representations, thereby broadening the LLM's capacity to probe a wider spectrum of audio information. All hypotheses are evaluated using an identical three-stage training curriculum on a dataset of 5.6 million audio-text pairs, ensuring controlled comparisons. Our final architecture, which incorporates all proposed modifications, achieves relative improvements from 10\% to 60\% over the baseline, validating our approach to optimizing cross-modal information transfer in audio-LLMs. Project page: https://ta012.github.io/PAL/

Representation Autoencoders (RAEs) have shown distinct advantages in diffusion modeling on ImageNet by training in high-dimensional semantic latent spaces. In this work, we investigate whether this framework can scale to large-scale, freeform text-to-image (T2I) generation. We first scale RAE decoders on the frozen representation encoder (SigLIP-2) beyond ImageNet by training on web, synthetic, and text-rendering data, finding that while scale improves general fidelity, targeted data composition is essential for specific domains like text. We then rigorously stress-test the RAE design choices originally proposed for ImageNet. Our analysis reveals that scaling simplifies the framework: while dimension-dependent noise scheduling remains critical, architectural complexities such as wide diffusion heads and noise-augmented decoding offer negligible benefits at scale Building on this simplified framework, we conduct a controlled comparison of RAE against the state-of-the-art FLUX VAE across diffusion transformer scales from 0.5B to 9.8B parameters. RAEs consistently outperform VAEs during pretraining across all model scales. Further, during finetuning on high-quality datasets, VAE-based models catastrophically overfit after 64 epochs, while RAE models remain stable through 256 epochs and achieve consistently better performance. Across all experiments, RAE-based diffusion models demonstrate faster convergence and better generation quality, establishing RAEs as a simpler and stronger foundation than VAEs for large-scale T2I generation. Additionally, because both visual understanding and generation can operate in a shared representation space, the multimodal model can directly reason over generated latents, opening new possibilities for unified models.

Despite the prominence of decoder-only language models, encoders remain crucial for resource-constrained applications. We introduce ModernGBERT (134M, 1B), a fully transparent family of German encoder models trained from scratch, incorporating architectural innovations from ModernBERT. To evaluate the practical trade-offs of training encoders from scratch, we also present LL\"aMmlein2Vec (120M, 1B, 7B), a family of encoders derived from German decoder-only models via LLM2Vec. We benchmark all models on natural language understanding, text embedding, and long-context reasoning tasks, enabling a controlled comparison between dedicated encoders and converted decoders. Our results show that ModernGBERT 1B outperforms prior state-of-the-art German encoders as well as encoders adapted via LLM2Vec, with regard to performance and parameter-efficiency. All models, training data, checkpoints and code are publicly available, advancing the German NLP ecosystem with transparent, high-performance encoder models.

AI agents hold the potential to revolutionize scientific productivity by automating literature reviews, replicating experiments, analyzing data, and even proposing new directions of inquiry; indeed, there are now many such agents, ranging from general-purpose "deep research" systems to specialized science-specific agents, such as AI Scientist and AIGS. Rigorous evaluation of these agents is critical for progress. Yet existing benchmarks fall short on several fronts: they (1) fail to provide holistic, product-informed measures of real-world use cases such as science research; (2) lack reproducible agent tools necessary for a controlled comparison of core agentic capabilities; (3) do not account for confounding variables such as model cost and tool access; (4) do not provide standardized interfaces for quick agent prototyping and evaluation; and (5) lack comprehensive baseline agents necessary to identify true advances. In response, we define principles and tooling for more rigorously benchmarking agents. Using these, we present AstaBench, a suite that provides the first holistic measure of agentic ability to perform scientific research, comprising 2400+ problems spanning the entire scientific discovery process and multiple scientific domains, and including many problems inspired by actual user requests to deployed Asta agents. Our suite comes with the first scientific research environment with production-grade search tools that enable controlled, reproducible evaluation, better accounting for confounders. Alongside, we provide a comprehensive suite of nine science-optimized classes of Asta agents and numerous baselines. Our extensive evaluation of 57 agents across 22 agent classes reveals several interesting findings, most importantly that despite meaningful progress on certain individual aspects, AI remains far from solving the challenge of science research assistance.

The scarcity of high-quality and multi-task singing datasets significantly hinders the development of diverse controllable and personalized singing tasks, as existing singing datasets suffer from low quality, limited diversity of languages and singers, absence of multi-technique information and realistic music scores, and poor task suitability. To tackle these problems, we present GTSinger, a large Global, multi-Technique, free-to-use, high-quality singing corpus with realistic music scores, designed for all singing tasks, along with its benchmarks. Particularly, (1) we collect 80.59 hours of high-quality singing voices, forming the largest recorded singing dataset; (2) 20 professional singers across nine widely spoken languages offer diverse timbres and styles; (3) we provide controlled comparison and phoneme-level annotations of six commonly used singing techniques, helping technique modeling and control; (4) GTSinger offers realistic music scores, assisting real-world musical composition; (5) singing voices are accompanied by manual phoneme-to-audio alignments, global style labels, and 16.16 hours of paired speech for various singing tasks. Moreover, to facilitate the use of GTSinger, we conduct four benchmark experiments: technique-controllable singing voice synthesis, technique recognition, style transfer, and speech-to-singing conversion. The corpus and demos can be found at http://gtsinger.github.io. We provide the dataset and the code for processing data and conducting benchmarks at https://huggingface.co/datasets/GTSinger/GTSinger and https://github.com/GTSinger/GTSinger.

We present YOLOBench, a benchmark comprised of 550+ YOLO-based object detection models on 4 different datasets and 4 different embedded hardware platforms (x86 CPU, ARM CPU, Nvidia GPU, NPU). We collect accuracy and latency numbers for a variety of YOLO-based one-stage detectors at different model scales by performing a fair, controlled comparison of these detectors with a fixed training environment (code and training hyperparameters). Pareto-optimality analysis of the collected data reveals that, if modern detection heads and training techniques are incorporated into the learning process, multiple architectures of the YOLO series achieve a good accuracy-latency trade-off, including older models like YOLOv3 and YOLOv4. We also evaluate training-free accuracy estimators used in neural architecture search on YOLOBench and demonstrate that, while most state-of-the-art zero-cost accuracy estimators are outperformed by a simple baseline like MAC count, some of them can be effectively used to predict Pareto-optimal detection models. We showcase that by using a zero-cost proxy to identify a YOLO architecture competitive against a state-of-the-art YOLOv8 model on a Raspberry Pi 4 CPU. The code and data are available at https://github.com/Deeplite/deeplite-torch-zoo

Political scientists are rapidly adopting large language models (LLMs) for text annotation, yet the sensitivity of annotation results to implementation choices remains poorly understood. Most evaluations test a single model or configuration; how model choice, model size, learning approach, and prompt style interact, and whether popular "best practices" survive controlled comparison, are largely unexplored. We present a controlled evaluation of these pipeline choices, testing six open-weight models across four political science annotation tasks under identical quantisation, hardware, and prompt-template conditions. Our central finding is methodological: interaction effects dominate main effects, so seemingly reasonable pipeline choices can become consequential researcher degrees of freedom. No single model, prompt style, or learning approach is uniformly superior, and the best-performing model varies across tasks. Two corollaries follow. First, model size is an unreliable guide both to cost and to performance: cross-family efficiency differences are so large that some larger models are less resource-intensive than much smaller alternatives, while within model families mid-range variants often match or exceed larger counterparts. Second, widely recommended prompt engineering techniques yield inconsistent and sometimes negative effects on annotation performance. We use these benchmark results to develop a validation-first framework - with a principled ordering of pipeline decisions, guidance on prompt freezing and held-out evaluation, reporting standards, and open-source tools - to help researchers navigate this decision space transparently.

We use (multi)modal deep neural networks (DNNs) to probe for sites of multimodal integration in the human brain by predicting stereoencephalography (SEEG) recordings taken while human subjects watched movies. We operationalize sites of multimodal integration as regions where a multimodal vision-language model predicts recordings better than unimodal language, unimodal vision, or linearly-integrated language-vision models. Our target DNN models span different architectures (e.g., convolutional networks and transformers) and multimodal training techniques (e.g., cross-attention and contrastive learning). As a key enabling step, we first demonstrate that trained vision and language models systematically outperform their randomly initialized counterparts in their ability to predict SEEG signals. We then compare unimodal and multimodal models against one another. Because our target DNN models often have different architectures, number of parameters, and training sets (possibly obscuring those differences attributable to integration), we carry out a controlled comparison of two models (SLIP and SimCLR), which keep all of these attributes the same aside from input modality. Using this approach, we identify a sizable number of neural sites (on average 141 out of 1090 total sites or 12.94%) and brain regions where multimodal integration seems to occur. Additionally, we find that among the variants of multimodal training techniques we assess, CLIP-style training is the best suited for downstream prediction of the neural activity in these sites.

This paper does not describe a new method; instead, it provides a thorough exploration of an important yet understudied design space related to recent advances in text-to-image synthesis -- specifically, the deep fusion of large language models (LLMs) and diffusion transformers (DiTs) for multi-modal generation. Previous studies mainly focused on overall system performance rather than detailed comparisons with alternative methods, and key design details and training recipes were often left undisclosed. These gaps create uncertainty about the real potential of this approach. To fill these gaps, we conduct an empirical study on text-to-image generation, performing controlled comparisons with established baselines, analyzing important design choices, and providing a clear, reproducible recipe for training at scale. We hope this work offers meaningful data points and practical guidelines for future research in multi-modal generation.

Multi-Agent Systems (MAS) powered by Large Language Models have unlocked advanced collaborative reasoning, yet they remain shackled by the inefficiency of discrete text communication, which imposes significant runtime overhead and information quantization loss. While latent state transfer offers a high-bandwidth alternative, existing approaches either assume homogeneous sender-receiver architectures or rely on pair-specific learned translators, limiting scalability and modularity across diverse model families with disjoint manifolds. In this work, we propose the Vision Wormhole, a novel framework that repurposes the visual interface of Vision-Language Models (VLMs) to enable model-agnostic, text-free communication. By introducing a Universal Visual Codec, we map heterogeneous reasoning traces into a shared continuous latent space and inject them directly into the receiver's visual pathway, effectively treating the vision encoder as a universal port for inter-agent telepathy. Our framework adopts a hub-and-spoke topology to reduce pairwise alignment complexity from O(N^2) to O(N) and leverages a label-free, teacher-student distillation objective to align the high-speed visual channel with the robust reasoning patterns of the text pathway. Extensive experiments across heterogeneous model families (e.g., Qwen-VL, Gemma) demonstrate that the Vision Wormhole reduces end-to-end wall-clock time in controlled comparisons while maintaining reasoning fidelity comparable to standard text-based MAS. Code is available at https://github.com/xz-liu/heterogeneous-latent-mas

Tokenization is a pivotal design choice for neural language modeling in morphologically rich languages (MRLs) such as Turkish, where productive agglutination challenges both vocabulary efficiency and morphological fidelity. Prior studies have explored tokenizer families and vocabulary sizes but typically (i) vary vocabulary without systematically controlling the tokenizer's training corpus, (ii) provide limited intrinsic diagnostics, and (iii) evaluate a narrow slice of downstream tasks. We present the first comprehensive, principled study of Turkish subword tokenization; a "subwords manifest", that jointly varies vocabulary size and tokenizer training corpus size (data and vocabulary coupling), compares multiple tokenizer families under matched parameter budgets (WordPiece, morphology level, and character baselines), and evaluates across semantic (NLI, STS, sentiment analysis, NER), syntactic (POS, dependency parsing), and morphology-sensitive probes. To explain why tokenizers succeed or fail, we introduce a morphology-aware diagnostic toolkit that goes beyond coarse aggregates to boundary-level micro/macro F1, decoupled lemma atomicity vs. surface boundary hits, over/under-segmentation indices, character/word edit distances (CER/WER), continuation rates, and affix-type coverage and token-level atomicity. Our contributions are fourfold: (i) a systematic investigation of the vocabulary-corpus-success triad; (ii) a unified, morphology-aware evaluation framework linking intrinsic diagnostics to extrinsic outcomes; (iii) controlled comparisons identifying when character-level and morphology-level tokenization pay off; and (iv) an open-source release of evaluation code, tokenizer pipelines, and models. As the first work of its kind, this "subwords manifest" delivers actionable guidance for building effective tokenizers in MRLs and establishes a reproducible foundation for future research.

Large language models for code generation increasingly rely on synthetic data, where both problem solutions and verification tests are generated by models. While this enables scalable data creation, it introduces a previously unexplored bottleneck: the verification ceiling, in which the quality and diversity of training data are fundamentally constrained by the capabilities of synthetic verifiers. In this work, we systematically study how verification design and strategies influence model performance. We investigate (i) what we verify by analyzing the impact of test complexity and quantity: richer test suites improve code generation capabilities (on average +3 pass@1), while quantity alone yields diminishing returns, (ii) how we verify by exploring relaxed pass thresholds: rigid 100% pass criteria can be overly restrictive. By allowing for relaxed thresholds or incorporating LLM-based soft verification, we can recover valuable training data, leading to a 2-4 point improvement in pass@1 performance. However, this benefit is contingent upon the strength and diversity of the test cases used, and (iii) why verification remains necessary through controlled comparisons of formally correct versus incorrect solutions and human evaluation: retaining diverse correct solutions per problem yields consistent generalization gains. Our results show that Verification as currently practiced is too rigid, filtering out valuable diversity. But it cannot be discarded, only recalibrated. By combining calibrated verification with diverse, challenging problem-solution pairs, we outline a path to break the verification ceiling and unlock stronger code generation models.

Can multi-agent communication pressure extract discrete, compositional representations of invisible physical properties from frozen video features? We show that agents communicating through a Gumbel-Softmax bottleneck with iterated learning develop positionally disentangled protocols for latent properties (elasticity, friction, mass ratio) without property labels or supervision on message structure. With 4 agents, 100% of 80 seeds converge to near-perfect compositionality (PosDis=0.999, holdout 98.3%). Controls confirm multi-agent structure -- not bandwidth or temporal coverage -- drives this effect. Causal intervention shows surgical property disruption (~15% drop on targeted property, <3% on others). A controlled backbone comparison reveals that the perceptual prior determines what is communicable: DINOv2 dominates on spatially-visible ramp physics (98.3% vs 95.1%), while V-JEPA 2 dominates on dynamics-only collision physics (87.4% vs 77.7%, d=2.74). Scale-matched (d=3.37) and frame-matched (d=6.53) controls attribute this gap entirely to video-native pretraining. The frozen protocol supports action-conditioned planning (91.5%) with counterfactual velocity reasoning (r=0.780). Validation on Physics 101 real camera footage confirms 85.6% mass-comparison accuracy on unseen objects, temporal dynamics contributing +11.2% beyond static appearance, agent-scaling compositionality replicating at 90% for 4 agents, and causal intervention extending to real video (d=1.87, p=0.022).

Designing a unified neural network to efficiently and inherently process sequential data with arbitrary lengths is a central and challenging problem in sequence modeling. The design choices in Transformer, including quadratic complexity and weak length extrapolation, have limited their ability to scale to long sequences. In this work, we propose Gecko, a neural architecture that inherits the design of Mega and Megalodon (exponential moving average with gated attention), and further introduces multiple technical components to improve its capability to capture long range dependencies, including timestep decay normalization, sliding chunk attention mechanism, and adaptive working memory. In a controlled pretraining comparison with Llama2 and Megalodon in the scale of 7 billion parameters and 2 trillion training tokens, Gecko achieves better efficiency and long-context scalability. Gecko reaches a training loss of 1.68, significantly outperforming Llama2-7B (1.75) and Megalodon-7B (1.70), and landing close to Llama2-13B (1.67). Notably, without relying on any context-extension techniques, Gecko exhibits inherent long-context processing and retrieval capabilities, stably handling sequences of up to 4 million tokens and retrieving information from contexts up to 4times longer than its attention window. Code: https://github.com/XuezheMax/gecko-llm

In this work, we present a systematic and comprehensive empirical evaluation of state-of-the-art reranking methods, encompassing large language model (LLM)-based, lightweight contextual, and zero-shot approaches, with respect to their performance in information retrieval tasks. We evaluate in total 22 methods, including 40 variants (depending on used LLM) across several established benchmarks, including TREC DL19, DL20, and BEIR, as well as a novel dataset designed to test queries unseen by pretrained models. Our primary goal is to determine, through controlled and fair comparisons, whether a performance disparity exists between LLM-based rerankers and their lightweight counterparts, particularly on novel queries, and to elucidate the underlying causes of any observed differences. To disentangle confounding factors, we analyze the effects of training data overlap, model architecture, and computational efficiency on reranking performance. Our findings indicate that while LLM-based rerankers demonstrate superior performance on familiar queries, their generalization ability to novel queries varies, with lightweight models offering comparable efficiency. We further identify that the novelty of queries significantly impacts reranking effectiveness, highlighting limitations in existing approaches. https://github.com/DataScienceUIBK/llm-reranking-generalization-study

Unified Multimodal Models (UMMs) offer powerful cross-modality capabilities but introduce new safety risks not observed in single-task models. Despite their emergence, existing safety benchmarks remain fragmented across tasks and modalities, limiting the comprehensive evaluation of complex system-level vulnerabilities. To address this gap, we introduce UniSAFE, the first comprehensive benchmark for system-level safety evaluation of UMMs across 7 I/O modality combinations, spanning conventional tasks and novel multimodal-context image generation settings. UniSAFE is built with a shared-target design that projects common risk scenarios across task-specific I/O configurations, enabling controlled cross-task comparisons of safety failures. Comprising 6,802 curated instances, we use UniSAFE to evaluate 15 state-of-the-art UMMs, both proprietary and open-source. Our results reveal critical vulnerabilities across current UMMs, including elevated safety violations in multi-image composition and multi-turn settings, with image-output tasks consistently more vulnerable than text-output tasks. These findings highlight the need for stronger system-level safety alignment for UMMs. Our code and data are publicly available at https://github.com/segyulee/UniSAFE

We introduce CompareBench, a benchmark for evaluating visual comparison reasoning in vision-language models (VLMs), a fundamental yet understudied skill. CompareBench consists of 1000 QA pairs across four tasks: quantity (600), temporal (100), geometric (200), and spatial (100). It is derived from two auxiliary datasets that we constructed: TallyBench (2000 counting images with QA) and HistCaps (515 historical images with bilingual captions). We evaluate both closed-source APIs (OpenAI, Gemini, Claude) and open-source models (Qwen2.5-VL and Qwen3-VL series). Results show clear scaling trends but also reveal critical limitations: even the strongest models consistently fail at temporal ordering and spatial relations, and they often make mistakes in basic counting and geometric comparisons that are trivial for humans. These findings demonstrate that visual comparison remains a systematic blind spot for current VLMs. By providing controlled, diverse, and diagnostic evaluation, CompareBench establishes a foundation for advancing more reliable multimodal reasoning.

Pretrained language models (PLMs) display impressive performances and have captured the attention of the NLP community. Establishing best practices in pretraining has, therefore, become a major focus of NLP research, especially since insights gained from monolingual English models may not necessarily apply to more complex multilingual models. One significant caveat of the current state of the art is that different works are rarely comparable: they often discuss different parameter counts, training data, and evaluation methodology. This paper proposes a comparison of multilingual pretraining objectives in a controlled methodological environment. We ensure that training data and model architectures are comparable, and discuss the downstream performances across 6 languages that we observe in probing and fine-tuning scenarios. We make two key observations: (1) the architecture dictates which pretraining objective is optimal; (2) multilingual translation is a very effective pretraining objective under the right conditions. We make our code, data, and model weights available at \url{https://github.com/Helsinki-NLP/lm-vs-mt}.

Evaluating the reasoning ability of language models (LMs) is complicated by their extensive parametric world knowledge, where benchmark performance often reflects factual recall rather than genuine reasoning. Existing datasets and approaches (e.g., temporal filtering, paraphrasing, adversarial substitution) cannot cleanly separate the two. We present SynthWorlds, a framework that disentangles task reasoning complexity from factual knowledge. In SynthWorlds, we construct parallel corpora representing two worlds with identical interconnected structure: a real-mapped world, where models may exploit parametric knowledge, and a synthetic-mapped world, where such knowledge is meaningless. On top of these corpora, we design two mirrored tasks as case studies: multi-hop question answering and page navigation, which maintain equal reasoning difficulty across worlds. Experiments in parametric-only (e.g., closed-book QA) and knowledge-augmented (e.g., retrieval-augmented) LM settings reveal a persistent knowledge advantage gap, defined as the performance boost models gain from memorized parametric world knowledge. Knowledge acquisition and integration mechanisms reduce but do not eliminate this gap, highlighting opportunities for system improvements. Fully automatic and scalable, SynthWorlds provides a controlled environment for evaluating LMs in ways that were previously challenging, enabling precise and testable comparisons of reasoning and memorization.

Tokenization strategies shape how models process electronic health records, yet fair comparisons of their effectiveness remain limited. We present a systematic evaluation of tokenization approaches for clinical time series modeling using transformer-based architectures, revealing task-dependent and sometimes counterintuitive findings about temporal and value feature importance. Through controlled ablations across four clinical prediction tasks on MIMIC-IV, we demonstrate that explicit time encodings provide no consistent statistically significant benefit for the evaluated downstream tasks. Value features show task-dependent importance, affecting mortality prediction but not readmission, suggesting code sequences alone can carry sufficient predictive signal. We further show that frozen pretrained code encoders dramatically outperform their trainable counterparts while requiring dramatically fewer parameters. Larger clinical encoders provide consistent improvements across tasks, benefiting from frozen embeddings that eliminate computational overhead. Our controlled evaluation enables fairer tokenization comparisons and demonstrates that simpler, parameter-efficient approaches can, in many cases, achieve strong performance, though the optimal tokenization strategy remains task-dependent.

The quadratic complexity and weak length extrapolation of Transformers limits their ability to scale to long sequences, and while sub-quadratic solutions like linear attention and state space models exist, they empirically underperform Transformers in pretraining efficiency and downstream task accuracy. We introduce Megalodon, a neural architecture for efficient sequence modeling with unlimited context length. Megalodon inherits the architecture of Mega (exponential moving average with gated attention), and further introduces multiple technical components to improve its capability and stability, including complex exponential moving average (CEMA), timestep normalization layer, normalized attention mechanism and pre-norm with two-hop residual configuration. In a controlled head-to-head comparison with Llama2, Megalodon achieves better efficiency than Transformer in the scale of 7 billion parameters and 2 trillion training tokens. Megalodon reaches a training loss of 1.70, landing mid-way between Llama2-7B (1.75) and 13B (1.67). Code: https://github.com/XuezheMax/megalodon

Matrix-based optimizers have attracted growing interest for improving LLM training efficiency, with significant progress centered on orthogonalization/whitening based methods. While yielding substantial performance gains, a fundamental question arises: can we develop new paradigms beyond orthogonalization, pushing the efficiency frontier further? We present Adaptively Rotated Optimization (ARO, a new matrix optimization framework that treats gradient rotation as a first class design principle. ARO accelerates LLM training by performing normed steepest descent in a rotated coordinate system, where the rotation is determined by a novel norm-informed policy. This perspective yields update rules that go beyond existing orthogonalization and whitening optimizers, improving sample efficiency in practice. To make comparisons reliable, we propose a rigorously controlled benchmarking protocol that reduces confounding and bias. Under this protocol, ARO consistently outperforms AdamW (by 1.3 sim1.35times) and orthogonalization methods (by 1.1sim1.15times) in LLM pretraining at up to 8B activated parameters, and up to 8times overtrain budget, without evidence of diminishing returns. Finally, we discuss how ARO can be reformulated as a symmetry-aware optimizer grounded in rotational symmetries of residual streams, motivating advanced designs that enable computationally efficient exploitation of cross-layer/cross module couplings.

The rapid advancement of generative AI has revolutionized image creation, enabling high-quality synthesis from text prompts while raising critical challenges for media authenticity. We present Ai-GenBench, a novel benchmark designed to address the urgent need for robust detection of AI-generated images in real-world scenarios. Unlike existing solutions that evaluate models on static datasets, Ai-GenBench introduces a temporal evaluation framework where detection methods are incrementally trained on synthetic images, historically ordered by their generative models, to test their ability to generalize to new generative models, such as the transition from GANs to diffusion models. Our benchmark focuses on high-quality, diverse visual content and overcomes key limitations of current approaches, including arbitrary dataset splits, unfair comparisons, and excessive computational demands. Ai-GenBench provides a comprehensive dataset, a standardized evaluation protocol, and accessible tools for both researchers and non-experts (e.g., journalists, fact-checkers), ensuring reproducibility while maintaining practical training requirements. By establishing clear evaluation rules and controlled augmentation strategies, Ai-GenBench enables meaningful comparison of detection methods and scalable solutions. Code and data are publicly available to ensure reproducibility and to support the development of robust forensic detectors to keep pace with the rise of new synthetic generators.

In-context learning (ICL) -- the capacity of a model to infer and apply abstract patterns from examples provided within its input -- has been extensively studied in large language models trained for next-token prediction on human text. In fact, prior work often attributes this emergent behavior to distinctive statistical properties in human language. This raises a fundamental question: can ICL arise organically in other sequence domains purely through large-scale predictive training? To explore this, we turn to genomic sequences, an alternative symbolic domain rich in statistical structure. Specifically, we study the Evo2 genomic model, trained predominantly on next-nucleotide (A/T/C/G) prediction, at a scale comparable to mid-sized LLMs. We develop a controlled experimental framework comprising symbolic reasoning tasks instantiated in both linguistic and genomic forms, enabling direct comparison of ICL across genomic and linguistic models. Our results show that genomic models, like their linguistic counterparts, exhibit log-linear gains in pattern induction as the number of in-context demonstrations increases. To the best of our knowledge, this is the first evidence of organically emergent ICL in genomic sequences, supporting the hypothesis that ICL arises as a consequence of large-scale predictive modeling over rich data. These findings extend emergent meta-learning beyond language, pointing toward a unified, modality-agnostic view of in-context learning.

Previous work has highlighted that existing post-hoc explanation methods exhibit disparities in explanation fidelity (across 'race' and 'gender' as sensitive attributes), and while a large body of work focuses on mitigating these issues at the explanation metric level, the role of the data generating process and black box model in relation to explanation disparities remains largely unexplored. Accordingly, through both simulations as well as experiments on a real-world dataset, we specifically assess challenges to explanation disparities that originate from properties of the data: limited sample size, covariate shift, concept shift, omitted variable bias, and challenges based on model properties: inclusion of the sensitive attribute and appropriate functional form. Through controlled simulation analyses, our study demonstrates that increased covariate shift, concept shift, and omission of covariates increase explanation disparities, with the effect pronounced higher for neural network models that are better able to capture the underlying functional form in comparison to linear models. We also observe consistent findings regarding the effect of concept shift and omitted variable bias on explanation disparities in the Adult income dataset. Overall, results indicate that disparities in model explanations can also depend on data and model properties. Based on this systematic investigation, we provide recommendations for the design of explanation methods that mitigate undesirable disparities.

We present an approach to enhancing the realism of synthetic images. The images are enhanced by a convolutional network that leverages intermediate representations produced by conventional rendering pipelines. The network is trained via a novel adversarial objective, which provides strong supervision at multiple perceptual levels. We analyze scene layout distributions in commonly used datasets and find that they differ in important ways. We hypothesize that this is one of the causes of strong artifacts that can be observed in the results of many prior methods. To address this we propose a new strategy for sampling image patches during training. We also introduce multiple architectural improvements in the deep network modules used for photorealism enhancement. We confirm the benefits of our contributions in controlled experiments and report substantial gains in stability and realism in comparison to recent image-to-image translation methods and a variety of other baselines.

Offline reinforcement learning (RL) aims at learning a good policy from a batch of collected data, without extra interactions with the environment during training. However, current offline RL benchmarks commonly have a large reality gap, because they involve large datasets collected by highly exploratory policies, and the trained policy is directly evaluated in the environment. In real-world situations, running a highly exploratory policy is prohibited to ensure system safety, the data is commonly very limited, and a trained policy should be well validated before deployment. In this paper, we present a near real-world offline RL benchmark, named NeoRL, which contains datasets from various domains with controlled sizes, and extra test datasets for policy validation. We evaluate existing offline RL algorithms on NeoRL and argue that the performance of a policy should also be compared with the deterministic version of the behavior policy, instead of the dataset reward. The empirical results demonstrate that the tested offline RL algorithms become less competitive to the deterministic policy on many datasets, and the offline policy evaluation hardly helps. The NeoRL suit can be found at http://polixir.ai/research/neorl. We hope this work will shed some light on future research and draw more attention when deploying RL in real-world systems.

Large Language Models used in ChatGPT have traditionally been trained to learn a refusal boundary: depending on the user's intent, the model is taught to either fully comply or outright refuse. While this is a strong mitigation for explicitly malicious prompts, focusing safety training on refusals can lead to brittleness for prompts with obscured user intent. Binary refusal boundaries are especially ill-suited for dual-use cases (such as biology or cybersecurity), where a user request can be answered safely at a high level, but in some cases can lead to malicious uplift if sufficiently detailed or actionable. As an alternative, we propose safe-completions: a safety-training approach that centers on the safety of the assistant's output, rather than a binary classification of the user's intent. Safe-completions seek to maximize helpfulness within the safety policy's constraints. We incorporated this approach into GPT-5 and find that across both production comparisons and internally controlled experiments, safe-completion training improves safety (especially on dual-use prompts), reduces the severity of residual safety failures, and substantially increases model helpfulness.

Parameter-efficient fine-tuning (PEFT) has become a standard approach for adapting large pre-trained models. Amongst PEFT methods, low-rank adaptation (LoRA) has achieved notable success. However, recent studies have highlighted its limitations compared against full-rank alternatives, particularly when applied to multimodal and large language models. In this work, we present a quantitative comparison amongst full-rank and low-rank PEFT methods using a synthetic matrix approximation benchmark with controlled spectral properties. Our results confirm that LoRA struggles to approximate matrices with relatively flat spectrums or high frequency components -- signs of high effective ranks. To this end, we introduce KRAdapter, a novel PEFT algorithm that leverages the Khatri-Rao product to produce weight updates, which, by construction, tends to produce matrix product with a high effective rank. We demonstrate performance gains with KRAdapter on vision-language models up to 1B parameters and on large language models up to 8B parameters, particularly on unseen common-sense reasoning tasks. In addition, KRAdapter maintains the memory and compute efficiency of LoRA, making it a practical and robust alternative to fine-tune billion-scale parameter models.

Individual neurons participate in the representation of multiple high-level concepts. To what extent can different interpretability methods successfully disentangle these roles? To help address this question, we introduce RAVEL (Resolving Attribute-Value Entanglements in Language Models), a dataset that enables tightly controlled, quantitative comparisons between a variety of existing interpretability methods. We use the resulting conceptual framework to define the new method of Multi-task Distributed Alignment Search (MDAS), which allows us to find distributed representations satisfying multiple causal criteria. With Llama2-7B as the target language model, MDAS achieves state-of-the-art results on RAVEL, demonstrating the importance of going beyond neuron-level analyses to identify features distributed across activations. We release our benchmark at https://github.com/explanare/ravel.

Chess provides an ideal testbed for evaluating the reasoning, modeling, and abstraction capabilities of large language models (LLMs), as it has well-defined structure and objective ground truth while admitting a wide spectrum of skill levels. However, existing evaluations of LLM ability in chess are ad hoc and narrow in scope, making it difficult to accurately measure LLM chess understanding and how it varies with scale, post-training methodologies, or architecture choices. We present ChessQA, a comprehensive benchmark that assesses LLM chess understanding across five task categories (Structural, Motifs, Short Tactics, Position Judgment, and Semantic), which approximately correspond to the ascending abstractions that players master as they accumulate chess knowledge, from understanding basic rules and learning tactical motifs to correctly calculating tactics, evaluating positions, and semantically describing high-level concepts. In this way, ChessQA captures a more comprehensive picture of chess ability and understanding, going significantly beyond the simple move quality evaluations done previously, and offers a controlled, consistent setting for diagnosis and comparison. Furthermore, ChessQA is inherently dynamic, with prompts, answer keys, and construction scripts that can evolve as models improve. Evaluating a range of contemporary LLMs, we find persistent weaknesses across all five categories and provide results and error analyses by category. We will release the code, periodically refreshed datasets, and a public leaderboard to support further research.

Large Language Models (LLMs) are increasingly deployed in real-world fact-checking systems, yet existing evaluations focus predominantly on claim verification and overlook the broader fact-checking workflow, including claim extraction and evidence retrieval. This narrow focus prevents current benchmarks from revealing systematic reasoning failures, factual blind spots, and robustness limitations of modern LLMs. To bridge this gap, we present FactArena, a fully automated arena-style evaluation framework that conducts comprehensive, stage-wise benchmarking of LLMs across the complete fact-checking pipeline. FactArena integrates three key components: (i) an LLM-driven fact-checking process that standardizes claim decomposition, evidence retrieval via tool-augmented interactions, and justification-based verdict prediction; (ii) an arena-styled judgment mechanism guided by consolidated reference guidelines to ensure unbiased and consistent pairwise comparisons across heterogeneous judge agents; and (iii) an arena-driven claim-evolution module that adaptively generates more challenging and semantically controlled claims to probe LLMs' factual robustness beyond fixed seed data. Across 16 state-of-the-art LLMs spanning seven model families, FactArena produces stable and interpretable rankings. Our analyses further reveal significant discrepancies between static claim-verification accuracy and end-to-end fact-checking competence, highlighting the necessity of holistic evaluation. The proposed framework offers a scalable and trustworthy paradigm for diagnosing LLMs' factual reasoning, guiding future model development, and advancing the reliable deployment of LLMs in safety-critical fact-checking applications.

Null Hypothesis Significance Testing is the de facto tool for assessing effectiveness differences between Information Retrieval systems. Researchers use statistical tests to check whether those differences will generalise to online settings or are just due to the samples observed in the laboratory. Much work has been devoted to studying which test is the most reliable when comparing a pair of systems, but most of the IR real-world experiments involve more than two. In the multiple comparisons scenario, testing several systems simultaneously may inflate the errors committed by the tests. In this paper, we use a new approach to assess the reliability of multiple comparison procedures using simulated and real TREC data. Experiments show that Wilcoxon plus the Benjamini-Hochberg correction yields Type I error rates according to the significance level for typical sample sizes while being the best test in terms of statistical power.

Finding relevant literature underpins the practice of evidence-based medicine. From 2014 to 2016, TREC conducted a clinical decision support track, wherein participants were tasked with finding articles relevant to clinical questions posed by physicians. In total, 87 teams have participated over the past three years, generating 395 runs. During this period, each team has trialled a variety of methods. While there was significant overlap in the methods employed by different teams, the results were varied. Due to the diversity of the platforms used, the results arising from the different techniques are not directly comparable, reducing the ability to build on previous work. By using a stable platform, we have been able to compare different document and query processing techniques, allowing us to experiment with different search parameters. We have used our system to reproduce leading teams runs, and compare the results obtained. By benchmarking our indexing and search techniques, we can statistically test a variety of hypotheses, paving the way for further research.

We present a generic framework for creating differentially private versions of any hypothesis test in a black-box way. We analyze the resulting tests analytically and experimentally. Most crucially, we show good practical performance for small data sets, showing that at epsilon = 1 we only need 5-6 times as much data as in the fully public setting. We compare our work to the one existing framework of this type, as well as to several individually-designed private hypothesis tests. Our framework is higher power than other generic solutions and at least competitive with (and often better than) individually-designed tests.

Large language models are being widely used across industries to generate content that contributes directly to key performance metrics, such as conversion rates. Pretrained models, however, often fall short when it comes to aligning with human preferences or optimizing for business objectives. As a result, fine-tuning with good-quality labeled data is essential to guide models to generate content that achieves better results. Controlled experiments, like A/B tests, can provide such data, but they are often expensive and come with significant engineering and logistical challenges. Meanwhile, companies have access to a vast amount of historical (observational) data that remains underutilized. In this work, we study the challenges and opportunities of fine-tuning LLMs using observational data. We show that while observational outcomes can provide valuable supervision, directly fine-tuning models on such data can lead them to learn spurious correlations. We present empirical evidence of this issue using various real-world datasets and propose DeconfoundLM, a method that explicitly removes the effect of known confounders from reward signals. Using simulation experiments, we demonstrate that DeconfoundLM improves the recovery of causal relationships and mitigates failure modes found in fine-tuning methods that ignore or naively incorporate confounding variables. Our findings highlight that while observational data presents risks, with the right causal corrections, it can be a powerful source of signal for LLM alignment. Please refer to the project page for code and related resources.

Creating test collections for offline retrieval evaluation requires human effort to judge documents' relevance. This expensive activity motivated much work in developing methods for constructing benchmarks with fewer assessment costs. In this respect, adjudication methods actively decide both which documents and the order in which experts review them, in order to better exploit the assessment budget or to lower it. Researchers evaluate the quality of those methods by measuring the correlation between the known gold ranking of systems under the full collection and the observed ranking of systems under the lower-cost one. This traditional analysis ignores whether and how the low-cost judgements impact on the statistically significant differences among systems with respect to the full collection. We fill this void by proposing a novel methodology to evaluate how the low-cost adjudication methods preserve the pairwise significant differences between systems as the full collection. In other terms, while traditional approaches look for stability in answering the question "is system A better than system B?", our proposed approach looks for stability in answering the question "is system A significantly better than system B?", which is the ultimate questions researchers need to answer to guarantee the generalisability of their results. Among other results, we found that the best methods in terms of ranking of systems correlation do not always match those preserving statistical significance.

Large language models (LLMs) are currently applied to scientific paper evaluation by assigning an absolute score to each paper independently. However, since score scales vary across conferences, time periods, and evaluation criteria, models trained on absolute scores are prone to fitting narrow, context-specific rules rather than developing robust scholarly judgment. To overcome this limitation, we propose shifting paper evaluation from isolated scoring to collaborative ranking. In particular, we design Comparison-Native framework for Paper Evaluation (CNPE), integrating comparison into both data construction and model learning. We first propose a graph-based similarity ranking algorithm to facilitate the sampling of more informative and discriminative paper pairs from a collection. We then enhance relative quality judgment through supervised fine-tuning and reinforcement learning with comparison-based rewards. At inference, the model performs pairwise comparisons over sampled paper pairs and aggregates these preference signals into a global relative quality ranking. Experimental results demonstrate that our framework achieves an average relative improvement of 21.8\% over the strong baseline DeepReview-14B, while exhibiting robust generalization to five previously unseen datasets. https://github.com/ECNU-Text-Computing/ComparisonReview{Code}.

We present Code Comparison Tuning (CCT), a simple and effective tuning method for code large language models (Code LLMs) to better handle subtle code errors. Specifically, we integrate the concept of comparison into instruction tuning, both at the token and sequence levels, enabling the model to discern even the slightest deviations in code. To compare the original code with an erroneous version containing manually added code errors, we use token-level preference loss for detailed token-level comparisons. Additionally, we combine code segments to create a new instruction tuning sample for sequence-level comparisons, enhancing the model's bug-fixing capability. Experimental results on the HumanEvalFix benchmark show that CCT surpasses instruction tuning in pass@1 scores by up to 4 points across diverse code LLMs, and extensive analysis demonstrates the effectiveness of our method.

The reproducibility of many experimental results in Deep Reinforcement Learning (RL) is under question. To solve this reproducibility crisis, we propose a theoretically sound methodology to compare multiple Deep RL algorithms. The performance of one execution of a Deep RL algorithm is random so that independent executions are needed to assess it precisely. When comparing several RL algorithms, a major question is how many executions must be made and how can we assure that the results of such a comparison is theoretically sound. Researchers in Deep RL often use less than 5 independent executions to compare algorithms: we claim that this is not enough in general. Moreover, when comparing several algorithms at once, the error of each comparison accumulates and must be taken into account with a multiple tests procedure to preserve low error guarantees. To address this problem in a statistically sound way, we introduce AdaStop, a new statistical test based on multiple group sequential tests. When comparing algorithms, AdaStop adapts the number of executions to stop as early as possible while ensuring that we have enough information to distinguish algorithms that perform better than the others in a statistical significant way. We prove both theoretically and empirically that AdaStop has a low probability of making an error (Family-Wise Error). Finally, we illustrate the effectiveness of AdaStop in multiple use-cases, including toy examples and difficult cases such as Mujoco environments.

To enhance the controllability of text-to-image diffusion models, current ControlNet-like models have explored various control signals to dictate image attributes. However, existing methods either handle conditions inefficiently or use a fixed number of conditions, which does not fully address the complexity of multiple conditions and their potential conflicts. This underscores the need for innovative approaches to manage multiple conditions effectively for more reliable and detailed image synthesis. To address this issue, we propose a novel framework, DynamicControl, which supports dynamic combinations of diverse control signals, allowing adaptive selection of different numbers and types of conditions. Our approach begins with a double-cycle controller that generates an initial real score sorting for all input conditions by leveraging pre-trained conditional generation models and discriminative models. This controller evaluates the similarity between extracted conditions and input conditions, as well as the pixel-level similarity with the source image. Then, we integrate a Multimodal Large Language Model (MLLM) to build an efficient condition evaluator. This evaluator optimizes the ordering of conditions based on the double-cycle controller's score ranking. Our method jointly optimizes MLLMs and diffusion models, utilizing MLLMs' reasoning capabilities to facilitate multi-condition text-to-image (T2I) tasks. The final sorted conditions are fed into a parallel multi-control adapter, which learns feature maps from dynamic visual conditions and integrates them to modulate ControlNet, thereby enhancing control over generated images. Through both quantitative and qualitative comparisons, DynamicControl demonstrates its superiority over existing methods in terms of controllability, generation quality and composability under various conditional controls.

Perception of toxicity evolves over time and often differs between geographies and cultural backgrounds. Similarly, black-box commercially available APIs for detecting toxicity, such as the Perspective API, are not static, but frequently retrained to address any unattended weaknesses and biases. We evaluate the implications of these changes on the reproducibility of findings that compare the relative merits of models and methods that aim to curb toxicity. Our findings suggest that research that relied on inherited automatic toxicity scores to compare models and techniques may have resulted in inaccurate findings. Rescoring all models from HELM, a widely respected living benchmark, for toxicity with the recent version of the API led to a different ranking of widely used foundation models. We suggest caution in applying apples-to-apples comparisons between studies and lay recommendations for a more structured approach to evaluating toxicity over time. Code and data are available at https://github.com/for-ai/black-box-api-challenges.

Peer review is a cornerstone of quality control in scientific publishing. With the increasing workload, the unintended use of `quick' heuristics, referred to as lazy thinking, has emerged as a recurring issue compromising review quality. Automated methods to detect such heuristics can help improve the peer-reviewing process. However, there is limited NLP research on this issue, and no real-world dataset exists to support the development of detection tools. This work introduces LazyReview, a dataset of peer-review sentences annotated with fine-grained lazy thinking categories. Our analysis reveals that Large Language Models (LLMs) struggle to detect these instances in a zero-shot setting. However, instruction-based fine-tuning on our dataset significantly boosts performance by 10-20 performance points, highlighting the importance of high-quality training data. Furthermore, a controlled experiment demonstrates that reviews revised with lazy thinking feedback are more comprehensive and actionable than those written without such feedback. We will release our dataset and the enhanced guidelines that can be used to train junior reviewers in the community. (Code available here: https://github.com/UKPLab/arxiv2025-lazy-review)

In natural language processing, a recently popular line of work explores how to best report the experimental results of neural networks. One exemplar publication, titled "Show Your Work: Improved Reporting of Experimental Results," advocates for reporting the expected validation effectiveness of the best-tuned model, with respect to the computational budget. In the present work, we critically examine this paper. As far as statistical generalizability is concerned, we find unspoken pitfalls and caveats with this approach. We analytically show that their estimator is biased and uses error-prone assumptions. We find that the estimator favors negative errors and yields poor bootstrapped confidence intervals. We derive an unbiased alternative and bolster our claims with empirical evidence from statistical simulation. Our codebase is at http://github.com/castorini/meanmax.

We introduce a new type of test, called a Turing Experiment (TE), for evaluating how well a language model, such as GPT-3, can simulate different aspects of human behavior. Unlike the Turing Test, which involves simulating a single arbitrary individual, a TE requires simulating a representative sample of participants in human subject research. We give TEs that attempt to replicate well-established findings in prior studies. We design a methodology for simulating TEs and illustrate its use to compare how well different language models are able to reproduce classic economic, psycholinguistic, and social psychology experiments: Ultimatum Game, Garden Path Sentences, Milgram Shock Experiment, and Wisdom of Crowds. In the first three TEs, the existing findings were replicated using recent models, while the last TE reveals a "hyper-accuracy distortion" present in some language models.

Hypothesis ranking is a crucial component of automated scientific discovery, particularly in natural sciences where wet-lab experiments are costly and throughput-limited. Existing approaches focus on pre-experiment ranking, relying solely on large language model's internal reasoning without incorporating empirical outcomes from experiments. We introduce the task of experiment-guided ranking, which aims to prioritize candidate hypotheses based on the results of previously tested ones. However, developing such strategies is challenging due to the impracticality of repeatedly conducting real experiments in natural science domains. To address this, we propose a simulator grounded in three domain-informed assumptions, modeling hypothesis performance as a function of similarity to a known ground truth hypothesis, perturbed by noise. We curate a dataset of 124 chemistry hypotheses with experimentally reported outcomes to validate the simulator. Building on this simulator, we develop a pseudo experiment-guided ranking method that clusters hypotheses by shared functional characteristics and prioritizes candidates based on insights derived from simulated experimental feedback. Experiments show that our method outperforms pre-experiment baselines and strong ablations.

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.

Comparative settings (e.g. pairwise choice, listwise ranking) have been adopted by a wide range of subjective studies for image quality assessment (IQA), as it inherently standardizes the evaluation criteria across different observers and offer more clear-cut responses. In this work, we extend the edge of emerging large multi-modality models (LMMs) to further advance visual quality comparison into open-ended settings, that 1) can respond to open-range questions on quality comparison; 2) can provide detailed reasonings beyond direct answers. To this end, we propose the Co-Instruct. To train this first-of-its-kind open-source open-ended visual quality comparer, we collect the Co-Instruct-562K dataset, from two sources: (a) LMM-merged single image quality description, (b) GPT-4V "teacher" responses on unlabeled data. Furthermore, to better evaluate this setting, we propose the MICBench, the first benchmark on multi-image comparison for LMMs. We demonstrate that Co-Instruct not only achieves 30% higher superior accuracy than state-of-the-art open-source LMMs, but also outperforms GPT-4V (its teacher), on both existing related benchmarks and the proposed MICBench. Our model is published at https://huggingface.co/q-future/co-instruct.

Recently, Agentic AI has become an increasingly popular research field. However, we argue that current agent research practices lack standardization and scientific rigor, making it hard to conduct fair comparisons among methods. As a result, it is still unclear how different design choices in agent frameworks affect effectiveness, and measuring their progress remains challenging. In this work, we conduct a systematic empirical study on GAIA benchmark and BrowseComp to examine the impact of popular design choices in key agent components in a fair and rigorous manner. We find that the lack of a standard evaluation protocol makes previous works, even open-sourced ones, non-reproducible, with significant variance between random runs. Therefore, we introduce a more robust evaluation protocol to stabilize comparisons. Our study reveals which components and designs are crucial for effective agents, while others are redundant, despite seeming logical. Based on our findings, we build and open-source OAgents, a new foundation agent framework that achieves state-of-the-art performance among open-source projects. OAgents offers a modular design for various agent components, promoting future research in Agentic AI.

Controllable text generation systems often leverage control codes to direct various properties of the output like style and length. Inspired by recent work on causal inference for NLP, this paper reveals a previously overlooked flaw in these control code-based conditional text generation algorithms. Spurious correlations in the training data can lead models to incorrectly rely on parts of the input other than the control code for attribute selection, significantly undermining downstream generation quality and controllability. We demonstrate the severity of this issue with a series of case studies and then propose two simple techniques to reduce these correlations in training sets. The first technique is based on resampling the data according to an example's propensity towards each linguistic attribute (IPS). The second produces multiple counterfactual versions of each example and then uses an additional feedback mechanism to remove noisy examples (feedback aware self-training, FAST). We evaluate on 3 tasks -- news headline, meta review, and search ads generation -- and demonstrate that FAST can significantly improve the controllability and language quality of generated outputs when compared to state-of-the-art controllable text generation approaches.

Retrieval-enhanced methods have become a primary approach in fact verification (FV); it requires reasoning over multiple retrieved pieces of evidence to verify the integrity of a claim. To retrieve evidence, existing work often employs off-the-shelf retrieval models whose design is based on the probability ranking principle. We argue that, rather than relevance, for FV we need to focus on the utility that a claim verifier derives from the retrieved evidence. We introduce the feedback-based evidence retriever(FER) that optimizes the evidence retrieval process by incorporating feedback from the claim verifier. As a feedback signal we use the divergence in utility between how effectively the verifier utilizes the retrieved evidence and the ground-truth evidence to produce the final claim label. Empirical studies demonstrate the superiority of FER over prevailing baselines.

We consider the task of evaluating policies of algorithmic resource allocation through randomized controlled trials (RCTs). Such policies are tasked with optimizing the utilization of limited intervention resources, with the goal of maximizing the benefits derived. Evaluation of such allocation policies through RCTs proves difficult, notwithstanding the scale of the trial, because the individuals' outcomes are inextricably interlinked through resource constraints controlling the policy decisions. Our key contribution is to present a new estimator leveraging our proposed novel concept, that involves retrospective reshuffling of participants across experimental arms at the end of an RCT. We identify conditions under which such reassignments are permissible and can be leveraged to construct counterfactual trials, whose outcomes can be accurately ascertained, for free. We prove theoretically that such an estimator is more accurate than common estimators based on sample means -- we show that it returns an unbiased estimate and simultaneously reduces variance. We demonstrate the value of our approach through empirical experiments on synthetic, semi-synthetic as well as real case study data and show improved estimation accuracy across the board.

This is the 1st part of the dissertation for my master degree and compares the power consumption using the default floating point (32bit) and Nvidia mixed precision (16bit and 32bit) while training a classification ML model. A custom PC with specific hardware was built to perform the experiments, and different ML hyper-parameters, such as batch size, neurons, and epochs, were chosen to build Deep Neural Networks (DNN). Additionally, various software was used during the experiments to collect the power consumption data in Watts from the Graphics Processing Unit (GPU), Central Processing Unit (CPU), Random Access Memory (RAM) and manually from a wattmeter connected to the wall. A benchmarking test with default hyper parameter values for the DNN was used as a reference, while the experiments used a combination of different settings. The results were recorded in Excel, and descriptive statistics were chosen to calculate the mean between the groups and compare them using graphs and tables. The outcome was positive when using mixed precision combined with specific hyper-parameters. Compared to the benchmarking, the optimisation for the classification reduced the power consumption between 7 and 11 Watts. Similarly, the carbon footprint is reduced because the calculation uses the same power consumption data. Still, a consideration is required when configuring hyper-parameters because it can negatively affect hardware performance. However, this research required inferential statistics, specifically ANOVA and T-test, to compare the relationship between the means. Furthermore, tests indicated no statistical significance of the relationship between the benchmarking and experiments. However, a more extensive implementation with a cluster of GPUs can increase the sample size significantly, as it is an essential factor and can change the outcome of the statistical analysis.

Medical coding is the task of assigning medical codes to clinical free-text documentation. Healthcare professionals manually assign such codes to track patient diagnoses and treatments. Automated medical coding can considerably alleviate this administrative burden. In this paper, we reproduce, compare, and analyze state-of-the-art automated medical coding machine learning models. We show that several models underperform due to weak configurations, poorly sampled train-test splits, and insufficient evaluation. In previous work, the macro F1 score has been calculated sub-optimally, and our correction doubles it. We contribute a revised model comparison using stratified sampling and identical experimental setups, including hyperparameters and decision boundary tuning. We analyze prediction errors to validate and falsify assumptions of previous works. The analysis confirms that all models struggle with rare codes, while long documents only have a negligible impact. Finally, we present the first comprehensive results on the newly released MIMIC-IV dataset using the reproduced models. We release our code, model parameters, and new MIMIC-III and MIMIC-IV training and evaluation pipelines to accommodate fair future comparisons.

In this research, we compare four different evaluation methods in coevolution on the Majority Function problem. The size of the problem is selected such that evaluation against all possible test cases is feasible. Two measures are used for the comparisons, i.e., the objective fitness derived from evaluating solutions against all test cases, and the objective fitness correlation (OFC), which is defined as the correlation coefficient between subjective and objective fitness. The results of our experiments suggest that a combination of average score and weighted informativeness may provide a more accurate evaluation in coevolution. In order to confirm this difference, a series of t-tests on the preference between each pair of the evaluation methods is performed. The resulting significance is affirmative, and the tests for two quality measures show similar preference on four evaluation methods. %This study is the first time OFC is actually computed on a real problem. Experiments on Majority Function problems with larger sizes and Parity problems are in progress, and their results will be added in the final version.

Deep-Research agents, which integrate large language models (LLMs) with search tools, have shown success in improving the effectiveness of handling complex queries that require iterative search planning and reasoning over search results. Evaluations on current benchmarks like BrowseComp relies on black-box live web search APIs, have notable limitations in (1) fairness: dynamic and opaque web APIs hinder fair comparisons and reproducibility of deep research methods; (2) transparency: lack of control over the document corpus makes it difficult to isolate retriever contributions. In other words, the current evaluations may compare a complete deep research system at a given time, but they do not foster well-controlled experiments to provide insights into the capability of underlying deep research LLMs. To address these challenges, we introduce BrowseComp-Plus, a benchmark derived from BrowseComp, employing a fixed, carefully curated corpus. Each query in BrowseComp-Plus includes human-verified supporting documents and mined challenging negatives, enabling controlled experimentation. The benchmark is shown to be effective in distinguishing the performance of deep research systems. For instance, the open-source model Search-R1, when paired with the BM25 retriever, achieves 3.86% accuracy, whereas the GPT-5 achieves 55.9%. Integrating the GPT-5 with the Qwen3-Embedding-8B retriever further enhances its accuracy to 70.1% with fewer search calls. This benchmark allows comprehensive evaluation and disentangled analysis of deep research agents and retrieval methods, fostering insights into retrieval effectiveness, citation accuracy, and context engineering in Deep-Research system.

The correct use of model evaluation, model selection, and algorithm selection techniques is vital in academic machine learning research as well as in many industrial settings. This article reviews different techniques that can be used for each of these three subtasks and discusses the main advantages and disadvantages of each technique with references to theoretical and empirical studies. Further, recommendations are given to encourage best yet feasible practices in research and applications of machine learning. Common methods such as the holdout method for model evaluation and selection are covered, which are not recommended when working with small datasets. Different flavors of the bootstrap technique are introduced for estimating the uncertainty of performance estimates, as an alternative to confidence intervals via normal approximation if bootstrapping is computationally feasible. Common cross-validation techniques such as leave-one-out cross-validation and k-fold cross-validation are reviewed, the bias-variance trade-off for choosing k is discussed, and practical tips for the optimal choice of k are given based on empirical evidence. Different statistical tests for algorithm comparisons are presented, and strategies for dealing with multiple comparisons such as omnibus tests and multiple-comparison corrections are discussed. Finally, alternative methods for algorithm selection, such as the combined F-test 5x2 cross-validation and nested cross-validation, are recommended for comparing machine learning algorithms when datasets are small.

Clinical trial matching is the task of identifying trials for which patients may be potentially eligible. Typically, this task is labor-intensive and requires detailed verification of patient electronic health records (EHRs) against the stringent inclusion and exclusion criteria of clinical trials. This process is manual, time-intensive, and challenging to scale up, resulting in many patients missing out on potential therapeutic options. Recent advancements in Large Language Models (LLMs) have made automating patient-trial matching possible, as shown in multiple concurrent research studies. However, the current approaches are confined to constrained, often synthetic datasets that do not adequately mirror the complexities encountered in real-world medical data. In this study, we present the first, end-to-end large-scale empirical evaluation of clinical trial matching using real-world EHRs. Our study showcases the capability of LLMs to accurately match patients with appropriate clinical trials. We perform experiments with proprietary LLMs, including GPT-4 and GPT-3.5, as well as our custom fine-tuned model called OncoLLM and show that OncoLLM, despite its significantly smaller size, not only outperforms GPT-3.5 but also matches the performance of qualified medical doctors. All experiments were carried out on real-world EHRs that include clinical notes and available clinical trials from a single cancer center in the United States.

This study investigates GPT-4's assessment of its performance in healthcare applications. A simple prompting technique was used to prompt the LLM with questions taken from the United States Medical Licensing Examination (USMLE) questionnaire and it was tasked to evaluate its confidence score before posing the question and after asking the question. The questionnaire was categorized into two groups-questions with feedback (WF) and questions with no feedback(NF) post-question. The model was asked to provide absolute and relative confidence scores before and after each question. The experimental findings were analyzed using statistical tools to study the variability of confidence in WF and NF groups. Additionally, a sequential analysis was conducted to observe the performance variation for the WF and NF groups. Results indicate that feedback influences relative confidence but doesn't consistently increase or decrease it. Understanding the performance of LLM is paramount in exploring its utility in sensitive areas like healthcare. This study contributes to the ongoing discourse on the reliability of AI, particularly of LLMs like GPT-4, within healthcare, offering insights into how feedback mechanisms might be optimized to enhance AI-assisted medical education and decision support.

Dataset summarisation is a fruitful approach to dataset inspection. However, when applied to a single dataset the discovery of visual concepts is restricted to those most prominent. We argue that a comparative approach can expand upon this paradigm to enable richer forms of dataset inspection that go beyond the most prominent concepts. To enable dataset comparison we present a module that learns concept-level prototypes across datasets. We leverage self-supervised learning to discover these prototypes without supervision, and we demonstrate the benefits of our approach in two case-studies. Our findings show that dataset comparison extends dataset inspection and we hope to encourage more works in this direction. Code and usage instructions available at https://github.com/Nanne/ProtoSim

Objective estimators of multimedia quality are often judged by comparing estimates with subjective "truth data," most often via Pearson correlation coefficient (PCC) or mean-squared error (MSE). But subjective test results contain noise, so striving for a PCC of 1.0 or an MSE of 0.0 is neither realistic nor repeatable. Numerous efforts have been made to acknowledge and appropriately accommodate subjective test noise in objective-subjective comparisons, typically resulting in new analysis frameworks and figures-of-merit. We take a different approach. By making only basic assumptions, we derive bounds on PCC and MSE that can be expected for a subjective test. Consistent with intuition, these bounds are functions of subjective vote variance. When a subjective test includes vote variance information, the calculation of the bounds is easy, and in this case we say the resulting bounds are "fully data-driven." We provide two options for calculating bounds in cases where vote variance information is not available. One option is to use vote variance information from other subjective tests that do provide such information, and the second option is to use a model for subjective votes. Thus we introduce a binomial-based model for subjective votes (BinoVotes) that naturally leads to a mean opinion score (MOS) model, named BinoMOS, with multiple unique desirable properties. BinoMOS reproduces the discrete nature of MOS values and its dependence on the number of votes per file. This modeling provides vote variance information required by the PCC and MSE bounds and we compare this modeling with data from 18 subjective tests. The modeling yields PCC and MSE bounds that agree very well with those found from the data directly. These results allow one to set expectations for the PCC and MSE that might be achieved for any subjective test, even those where vote variance information is not available.

Randomized controlled trials (RCTs) are the benchmark for causal inference, yet field implementation can deviate. We here present a remote audit - a design-based, preregistrable diagnostic that uses only pre-treatment satellite imagery to test whether assignment is independent of local conditions. The conditional randomization test of the remote audit evaluates whether treatment assignment is more predictable from pre-treatment satellite features than expected under the experiment's registered mechanism, providing a finite-sample valid, design-based diagnostic that requires no parametric assumptions. The procedure is finite-sample valid, honors blocks and clusters, and controls multiplicity across image models and resolutions via a max-statistic. We illustrate with two RCTs: Uganda's Youth Opportunities Program, where the audit corroborates randomization and flags selection and missing-data risks; and a school-based trial in Bangladesh, where assignment is highly predictable from pre-treatment features relative to the stated design, consistent with independent concerns about irregularities. Remote audits complement balance tests, lower early-stage costs, and enable rapid design checks when baseline surveys are expensive or infeasible.

Reward models are critical for aligning models to follow instructions, and are typically trained following one of two popular paradigms: Bradley-Terry style or Regression style. However, there is a lack of evidence that either approach is better than the other, when adequately matched for data. This is primarily because these approaches require data collected in different (but incompatible) formats, meaning that adequately matched data is not available in existing public datasets. To tackle this problem, we release preference annotations (designed for Bradley-Terry training) to complement existing ratings (designed for Regression style training) in the HelpSteer2 dataset. To improve data interpretability, preference annotations are accompanied with human-written justifications. Using this data, we conduct the first head-to-head comparison of Bradley-Terry and Regression models when adequately matched for data. Based on insights derived from such a comparison, we propose a novel approach to combine Bradley-Terry and Regression reward modeling. A Llama-3.1-70B-Instruct model tuned with this approach scores 94.1 on RewardBench, emerging top of more than 140 reward models as of 1 Oct 2024. We also demonstrate the effectiveness of this reward model at aligning models to follow instructions in RLHF. We open-source this dataset (CC-BY-4.0 license) at https://huggingface.co/datasets/nvidia/HelpSteer2 and openly release the trained Reward Model at https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Reward

Understanding causality should be a core requirement of any attempt to build real impact through AI. Due to the inherent unobservability of counterfactuals, large randomised trials (RCTs) are the standard for causal inference. But large experiments are generically expensive, and randomisation carries its own costs, e.g. when suboptimal decisions are trialed. Recent work has proposed more sample-efficient alternatives to RCTs, but these are not adaptable to the downstream application for which the causal effect is sought. In this work, we develop a task-specific approach to experimental design and derive sampling strategies customised to particular downstream applications. Across a range of important tasks, real-world datasets, and sample sizes, our method outperforms other benchmarks, e.g. requiring an order-of-magnitude less data to match RCT performance on targeted marketing tasks.

Behavior Cloning (BC) on curated (or filtered) data is the predominant paradigm for supervised fine-tuning (SFT) of large language models; as well as for imitation learning of control policies. Here, we draw on a connection between this successful strategy and the theory and practice of finding optimal policies via Reinforcement Learning (RL). Building on existing literature, we clarify that SFT can be understood as maximizing a lower bound on the RL objective in a sparse reward setting. Giving support to its often observed good performance. From this viewpoint, we realize that a small modification to SFT leads to an importance weighted variant that behaves closer to training with RL as it: i) optimizes a tighter bound to the RL objective and, ii) can improve performance compared to SFT on curated data. We refer to this variant as importance weighted supervised fine-tuning (iw-SFT). We show that it is easy to implement and can be further generalized to training with quality scored data. The resulting SFT variants are competitive with more advanced RL algorithms for large language models and for training policies in continuous control tasks. For example achieving 66.7% on the AIME 2024 dataset.

To effectively optimize and personalize treatments, it is necessary to investigate the heterogeneity of treatment effects. With the wide range of users being treated over many online controlled experiments, the typical approach of manually investigating each dimension of heterogeneity becomes overly cumbersome and prone to subjective human biases. We need an efficient way to search through thousands of experiments with hundreds of target covariates and hundreds of breakdown dimensions. In this paper, we propose a systematic paradigm for detecting, surfacing and characterizing heterogeneous treatment effects. First, we detect if treatment effect variation is present in an experiment, prior to specifying any breakdowns. Second, we surface the most relevant dimensions for heterogeneity. Finally, we characterize the heterogeneity beyond just the conditional average treatment effects (CATE) by studying the conditional distributions of the estimated individual treatment effects. We show the effectiveness of our methods using simulated data and empirical studies.

In this study, we present a novel paradigm for industrial robotic embodied agents, encapsulating an 'agent as cerebrum, controller as cerebellum' architecture. Our approach harnesses the power of Large Multimodal Models (LMMs) within an agent framework known as AeroAgent, tailored for drone technology in industrial settings. To facilitate seamless integration with robotic systems, we introduce ROSchain, a bespoke linkage framework connecting LMM-based agents to the Robot Operating System (ROS). We report findings from extensive empirical research, including simulated experiments on the Airgen and real-world case study, particularly in individual search and rescue operations. The results demonstrate AeroAgent's superior performance in comparison to existing Deep Reinforcement Learning (DRL)-based agents, highlighting the advantages of the embodied LMM in complex, real-world scenarios.

The past two decades have witnessed the rapid development of personalized recommendation techniques. Despite significant progress made in both research and practice of recommender systems, to date, there is a lack of a widely-recognized benchmarking standard in this field. Many existing studies perform model evaluations and comparisons in an ad-hoc manner, for example, by employing their own private data splits or using different experimental settings. Such conventions not only increase the difficulty in reproducing existing studies, but also lead to inconsistent experimental results among them. This largely limits the credibility and practical value of research results in this field. To tackle these issues, we present an initiative project (namely BARS) aiming for open benchmarking for recommender systems. In comparison to some earlier attempts towards this goal, we take a further step by setting up a standardized benchmarking pipeline for reproducible research, which integrates all the details about datasets, source code, hyper-parameter settings, running logs, and evaluation results. The benchmark is designed with comprehensiveness and sustainability in mind. It covers both matching and ranking tasks, and also enables researchers to easily follow and contribute to the research in this field. This project will not only reduce the redundant efforts of researchers to re-implement or re-run existing baselines, but also drive more solid and reproducible research on recommender systems. We would like to call upon everyone to use the BARS benchmark for future evaluation, and contribute to the project through the portal at: https://openbenchmark.github.io/BARS.

Black-box risk scoring models permeate our lives, yet are typically proprietary or opaque. We propose Distill-and-Compare, a model distillation and comparison approach to audit such models. To gain insight into black-box models, we treat them as teachers, training transparent student models to mimic the risk scores assigned by black-box models. We compare the student model trained with distillation to a second un-distilled transparent model trained on ground-truth outcomes, and use differences between the two models to gain insight into the black-box model. Our approach can be applied in a realistic setting, without probing the black-box model API. We demonstrate the approach on four public data sets: COMPAS, Stop-and-Frisk, Chicago Police, and Lending Club. We also propose a statistical test to determine if a data set is missing key features used to train the black-box model. Our test finds that the ProPublica data is likely missing key feature(s) used in COMPAS.

Pretrained neural networks have attracted significant interest in chemistry and small molecule drug design. Embeddings from these models are widely used for molecular property prediction, virtual screening, and small data learning in molecular chemistry. This study presents the most extensive comparison of such models to date, evaluating 25 models across 25 datasets. Under a fair comparison framework, we assess models spanning various modalities, architectures, and pretraining strategies. Using a dedicated hierarchical Bayesian statistical testing model, we arrive at a surprising result: nearly all neural models show negligible or no improvement over the baseline ECFP molecular fingerprint. Only the CLAMP model, which is also based on molecular fingerprints, performs statistically significantly better than the alternatives. These findings raise concerns about the evaluation rigor in existing studies. We discuss potential causes, propose solutions, and offer practical recommendations.

Machine learning has demonstrated remarkable performance over finite datasets, yet whether the scores over the fixed benchmarks can sufficiently indicate the model's performance in the real world is still in discussion. In reality, an ideal robust model will probably behave similarly to the oracle (e.g., the human users), thus a good evaluation protocol is probably to evaluate the models' behaviors in comparison to the oracle. In this paper, we introduce a new robustness measurement that directly measures the image classification model's performance compared with a surrogate oracle (i.e., a foundation model). Besides, we design a simple method that can accomplish the evaluation beyond the scope of the benchmarks. Our method extends the image datasets with new samples that are sufficiently perturbed to be distinct from the ones in the original sets, but are still bounded within the same image-label structure the original test image represents, constrained by a foundation model pretrained with a large amount of samples. As a result, our new method will offer us a new way to evaluate the models' robustness performance, free of limitations of fixed benchmarks or constrained perturbations, although scoped by the power of the oracle. In addition to the evaluation results, we also leverage our generated data to understand the behaviors of the model and our new evaluation strategies.

Despite its importance to experimental design, statistical power (the probability that, given a real effect, an experiment will reject the null hypothesis) has largely been ignored by the NLP community. Underpowered experiments make it more difficult to discern the difference between statistical noise and meaningful model improvements, and increase the chances of exaggerated findings. By meta-analyzing a set of existing NLP papers and datasets, we characterize typical power for a variety of settings and conclude that underpowered experiments are common in the NLP literature. In particular, for several tasks in the popular GLUE benchmark, small test sets mean that most attempted comparisons to state of the art models will not be adequately powered. Similarly, based on reasonable assumptions, we find that the most typical experimental design for human rating studies will be underpowered to detect small model differences, of the sort that are frequently studied. For machine translation, we find that typical test sets of 2000 sentences have approximately 75% power to detect differences of 1 BLEU point. To improve the situation going forward, we give an overview of best practices for power analysis in NLP and release a series of notebooks to assist with future power analyses.

In Reinforcement Learning (RL), an agent acts in an unknown environment to maximize the expected cumulative discounted sum of an external reward signal, i.e., the expected return. In practice, in many tasks of interest, such as policy optimization, the agent usually spends its interaction budget by collecting episodes of fixed length within a simulator (i.e., Monte Carlo simulation). However, given the discounted nature of the RL objective, this data collection strategy might not be the best option. Indeed, the rewards taken in early simulation steps weigh exponentially more than future rewards. Taking a cue from this intuition, in this paper, we design an a-priori budget allocation strategy that leads to the collection of trajectories of different lengths, i.e., truncated. The proposed approach provably minimizes the width of the confidence intervals around the empirical estimates of the expected return of a policy. After discussing the theoretical properties of our method, we make use of our trajectory truncation mechanism to extend Policy Optimization via Importance Sampling (POIS, Metelli et al., 2018) algorithm. Finally, we conduct a numerical comparison between our algorithm and POIS: the results are consistent with our theory and show that an appropriate truncation of the trajectories can succeed in improving performance.

Rigorous and reproducible evaluation is critical for assessing the state of the art and for guiding scientific advances in Artificial Intelligence. Evaluation is challenging in practice due to several reasons, including benchmark saturation, lack of transparency in methods used for measurement, development challenges in extracting measurements for generative tasks, and, more generally, the extensive number of capabilities required for a well-rounded comparison across models. We make three contributions to alleviate the above challenges. First, we present Eureka, an open-source framework for standardizing evaluations of large foundation models beyond single-score reporting and rankings. Second, we introduce Eureka-Bench as an extensible collection of benchmarks testing capabilities that (i) are still challenging for state-of-the-art models and (ii) represent fundamental but overlooked language and multimodal capabilities. The inherent space for improvement in non-saturated benchmarks enables us to discover meaningful differences between models at a capability level. Third, using Eureka, we conduct an analysis of 12 state-of-the-art models, providing in-depth insights into failure understanding and model comparison, which can be leveraged to plan targeted improvements. In contrast to recent trends in reports and leaderboards showing absolute rankings and claims for one model or another to be the best, our analysis shows that there is no such best model. Different models have different strengths, but there are models that appear more often than others as best performers for some capabilities. Despite the recent improvements, current models still struggle with several fundamental capabilities including detailed image understanding, benefiting from multimodal input when available rather than fully relying on language, factuality and grounding for information retrieval, and over refusals.

We have compared three common visualisations for hierarchical quantitative data, treemaps, icicle plots and sunburst charts as well as a semicircular variant of sunburst charts we call the sundown chart. In a pilot study, we found that the sunburst chart was least preferred. In a controlled study with 12 participants, we compared treemaps, icicle plots and sundown charts. Treemap was the least preferred and had a slower performance on a basic navigation task and slower performance and accuracy in hierarchy understanding tasks. The icicle plot and sundown chart had similar performance with slight user preference for the icicle plot.

In modern recommendation systems, the standard pipeline involves training machine learning models on historical data to predict user behaviors and improve recommendations continuously. However, these data training loops can introduce interference in A/B tests, where data generated by control and treatment algorithms, potentially with different distributions, are combined. To address these challenges, we introduce a novel approach called weighted training. This approach entails training a model to predict the probability of each data point appearing in either the treatment or control data and subsequently applying weighted losses during model training. We demonstrate that this approach achieves the least variance among all estimators that do not cause shifts in the training distributions. Through simulation studies, we demonstrate the lower bias and variance of our approach compared to other methods.

In recent years, large language models have greatly improved in their ability to perform complex multi-step reasoning. However, even state-of-the-art models still regularly produce logical mistakes. To train more reliable models, we can turn either to outcome supervision, which provides feedback for a final result, or process supervision, which provides feedback for each intermediate reasoning step. Given the importance of training reliable models, and given the high cost of human feedback, it is important to carefully compare the both methods. Recent work has already begun this comparison, but many questions still remain. We conduct our own investigation, finding that process supervision significantly outperforms outcome supervision for training models to solve problems from the challenging MATH dataset. Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Additionally, we show that active learning significantly improves the efficacy of process supervision. To support related research, we also release PRM800K, the complete dataset of 800,000 step-level human feedback labels used to train our best reward model.

Online medical consultation (OMC) restricts doctors to gathering patient information solely through inquiries, making the already complex sequential decision-making process of diagnosis even more challenging. Recently, the rapid advancement of large language models has demonstrated a significant potential to transform OMC. However, most studies have primarily focused on improving diagnostic accuracy under conditions of relatively sufficient information, while paying limited attention to the "inquiry" phase of the consultation process. This lack of focus has left the relationship between "inquiry" and "diagnosis" insufficiently explored. In this paper, we first extract real patient interaction strategies from authentic doctor-patient conversations and use these strategies to guide the training of a patient simulator that closely mirrors real-world behavior. By inputting medical records into our patient simulator to simulate patient responses, we conduct extensive experiments to explore the relationship between "inquiry" and "diagnosis" in the consultation process. Experimental results demonstrate that inquiry and diagnosis adhere to the Liebig's law: poor inquiry quality limits the effectiveness of diagnosis, regardless of diagnostic capability, and vice versa. Furthermore, the experiments reveal significant differences in the inquiry performance of various models. To investigate this phenomenon, we categorize the inquiry process into four types: (1) chief complaint inquiry; (2) specification of known symptoms; (3) inquiry about accompanying symptoms; and (4) gathering family or medical history. We analyze the distribution of inquiries across the four types for different models to explore the reasons behind their significant performance differences. We plan to open-source the weights and related code of our patient simulator at https://github.com/LIO-H-ZEN/PatientSimulator.

How do we most effectively treat a disease or condition? Ideally, we could consult a database of evidence gleaned from clinical trials to answer such questions. Unfortunately, no such database exists; clinical trial results are instead disseminated primarily via lengthy natural language articles. Perusing all such articles would be prohibitively time-consuming for healthcare practitioners; they instead tend to depend on manually compiled systematic reviews of medical literature to inform care. NLP may speed this process up, and eventually facilitate immediate consult of published evidence. The Evidence Inference dataset was recently released to facilitate research toward this end. This task entails inferring the comparative performance of two treatments, with respect to a given outcome, from a particular article (describing a clinical trial) and identifying supporting evidence. For instance: Does this article report that chemotherapy performed better than surgery for five-year survival rates of operable cancers? In this paper, we collect additional annotations to expand the Evidence Inference dataset by 25\%, provide stronger baseline models, systematically inspect the errors that these make, and probe dataset quality. We also release an abstract only (as opposed to full-texts) version of the task for rapid model prototyping. The updated corpus, documentation, and code for new baselines and evaluations are available at http://evidence-inference.ebm-nlp.com/.

Medical systematic reviews typically require assessing all the documents retrieved by a search. The reason is two-fold: the task aims for ``total recall''; and documents retrieved using Boolean search are an unordered set, and thus it is unclear how an assessor could examine only a subset. Screening prioritisation is the process of ranking the (unordered) set of retrieved documents, allowing assessors to begin the downstream processes of the systematic review creation earlier, leading to earlier completion of the review, or even avoiding screening documents ranked least relevant. Screening prioritisation requires highly effective ranking methods. Pre-trained language models are state-of-the-art on many IR tasks but have yet to be applied to systematic review screening prioritisation. In this paper, we apply several pre-trained language models to the systematic review document ranking task, both directly and fine-tuned. An empirical analysis compares how effective neural methods compare to traditional methods for this task. We also investigate different types of document representations for neural methods and their impact on ranking performance. Our results show that BERT-based rankers outperform the current state-of-the-art screening prioritisation methods. However, BERT rankers and existing methods can actually be complementary, and thus, further improvements may be achieved if used in conjunction.

Recently, a myriad of conditional image generation and editing models have been developed to serve different downstream tasks, including text-to-image generation, text-guided image editing, subject-driven image generation, control-guided image generation, etc. However, we observe huge inconsistencies in experimental conditions: datasets, inference, and evaluation metrics - render fair comparisons difficult. This paper proposes ImagenHub, which is a one-stop library to standardize the inference and evaluation of all the conditional image generation models. Firstly, we define seven prominent tasks and curate high-quality evaluation datasets for them. Secondly, we built a unified inference pipeline to ensure fair comparison. Thirdly, we design two human evaluation scores, i.e. Semantic Consistency and Perceptual Quality, along with comprehensive guidelines to evaluate generated images. We train expert raters to evaluate the model outputs based on the proposed metrics. Our human evaluation achieves a high inter-worker agreement of Krippendorff's alpha on 76% models with a value higher than 0.4. We comprehensively evaluated a total of around 30 models and observed three key takeaways: (1) the existing models' performance is generally unsatisfying except for Text-guided Image Generation and Subject-driven Image Generation, with 74% models achieving an overall score lower than 0.5. (2) we examined the claims from published papers and found 83% of them hold with a few exceptions. (3) None of the existing automatic metrics has a Spearman's correlation higher than 0.2 except subject-driven image generation. Moving forward, we will continue our efforts to evaluate newly published models and update our leaderboard to keep track of the progress in conditional image generation.

We present a straightforward statistical test to detect certain violations of the assumption that the data are Independent and Identically Distributed (IID). The specific form of violation considered is common across real-world applications: whether the examples are ordered in the dataset such that almost adjacent examples tend to have more similar feature values (e.g. due to distributional drift, or attractive interactions between datapoints). Based on a k-Nearest Neighbors estimate, our approach can be used to audit any multivariate numeric data as well as other data types (image, text, audio, etc.) that can be numerically represented, perhaps with model embeddings. Compared with existing methods to detect drift or auto-correlation, our approach is both applicable to more types of data and also able to detect a wider variety of IID violations in practice. Code: https://github.com/cleanlab/cleanlab

Noise plagues many numerical datasets, where the recorded values in the data may fail to match the true underlying values due to reasons including: erroneous sensors, data entry/processing mistakes, or imperfect human estimates. We consider general regression settings with covariates and a potentially corrupted response whose observed values may contain errors. By accounting for various uncertainties, we introduced veracity scores that distinguish between genuine errors and natural data fluctuations, conditioned on the available covariate information in the dataset. We propose a simple yet efficient filtering procedure for eliminating potential errors, and establish theoretical guarantees for our method. We also contribute a new error detection benchmark involving 5 regression datasets with real-world numerical errors (for which the true values are also known). In this benchmark and additional simulation studies, our method identifies incorrect values with better precision/recall than other approaches.

We consider identification of average treatment effects on the treated (ATT) within the difference-in-differences (DiD) framework in the presence of endogenous sample selection. First, we establish that the usual DiD estimand fails to recover meaningful treatment effects, even if selection and treatment assignment are independent. Next, we partially identify the ATT for individuals who are always observed post-treatment regardless of their treatment status, and derive bounds on this parameter under different sets of assumptions about the relationship between sample selection and treatment assignment. Extensions to the repeated cross-section and two-by-two comparisons in the staggered adoption case are explored. Furthermore, we provide identification results for the ATT of three additional empirically relevant latent groups by incorporating outcome mean dominance assumptions which have intuitive appeal in applications. Finally, two empirical illustrations demonstrate the approach's usefulness by revisiting (i) the effect of a job training program on earnings(Calonico & Smith, 2017) and (ii) the effect of a working-from-home policy on employee performance (Bloom, Liang, Roberts, & Ying, 2015).

Many tasks revolve around editing a document, whether code or text. We formulate the revision similarity problem to unify a wide range of machine learning evaluation problems whose goal is to assess a revision to an existing document. We observe that revisions usually change only a small portion of an existing document, so the existing document and its immediate revisions share a majority of their content. We formulate five adequacy criteria for revision similarity measures, designed to align them with human judgement. We show that popular pairwise measures, like BLEU, fail to meet these criteria, because their scores are dominated by the shared content. They report high similarity between two revisions when humans would assess them as quite different. This is a fundamental flaw we address. We propose a novel static measure, Excision Score (ES), which computes longest common subsequence (LCS) to remove content shared by an existing document with the ground truth and predicted revisions, before comparing only the remaining divergent regions. This is analogous to a surgeon creating a sterile field to focus on the work area. We use approximation to speed the standard cubic LCS computation to quadratic. In code-editing evaluation, where static measures are often used as a cheap proxy for passing tests, we demonstrate that ES surpasses existing measures. When aligned with test execution on HumanEvalFix, ES improves over its nearest competitor, SARI, by 12% Pearson correlation and by >21% over standard measures like BLEU. The key criterion is invariance to shared context; when we perturb HumanEvalFix with increased shared context, ES' improvement over SARI increases to 20% and >30% over standard measures. ES also handles other corner cases that other measures do not, such as correctly aligning moved code blocks, and appropriately rewarding matching insertions or deletions.

Medical reasoning models (MRMs) achieve superior performance on medical benchmarks compared to medical LLMs; however, high accuracy alone is insufficient for practical deployment. One of such requirements for real-world application is robustness to varying output constraints. Specifically, posing the same medical question while requesting different answer formats should not affect the underlying correctness of the response. We investigate this phenomenon in this paper, focusing on MRMs. To quantify this behavior, we propose the metric answer-format robustness: the ability to reliably generate correct outputs across varying specified formats. We examine three representative formats: multiple-choice, open-ended question-answering, and ranked lists. Across 15 proprietary and open-weight models, we observe substantial variation in format robustness (35-100%). Furthermore, we conduct controlled fine-tuning experiments on a shared backbone with matched training data to isolate the effects of the fine-tuning paradigm. We find that supervised fine-tuning yields more stable behavior across formats, whereas reinforcement fine-tuning often exhibits higher cross-format brittleness, with the degree of instability strongly dependent on reward design. Overall, answer-format robustness in MRMs is trainable yet brittle and requires careful evaluation for practical medical use.

We study regressions with multiple treatments and a set of controls that is flexible enough to purge omitted variable bias. We show that these regressions generally fail to estimate convex averages of heterogeneous treatment effects -- instead, estimates of each treatment's effect are contaminated by non-convex averages of the effects of other treatments. We discuss three estimation approaches that avoid such contamination bias, including the targeting of easiest-to-estimate weighted average effects. A re-analysis of nine empirical applications finds economically and statistically meaningful contamination bias in observational studies; contamination bias in experimental studies is more limited due to smaller variability in propensity scores.

In regression and causal inference, controlled subgroup selection aims to identify, with inferential guarantees, a subgroup (defined as a subset of the covariate space) on which the average response or treatment effect is above a given threshold. E.g., in a clinical trial, it may be of interest to find a subgroup with a positive average treatment effect. However, existing methods either lack inferential guarantees, heavily restrict the search for the subgroup, or sacrifice efficiency by naive data splitting. We propose a novel framework called chiseling that allows the analyst to interactively refine and test a candidate subgroup by iteratively shrinking it. The sole restriction is that the shrinkage direction only depends on the points outside the current subgroup, but otherwise the analyst may leverage any prior information or machine learning algorithm. Despite this flexibility, chiseling controls the probability that the discovered subgroup is null (e.g., has a non-positive average treatment effect) under minimal assumptions: for example, in randomized experiments, this inferential validity guarantee holds under only bounded moment conditions. When applied to a variety of simulated datasets and a real survey experiment, chiseling identifies substantially better subgroups than existing methods with inferential guarantees.

Partial correlations quantify linear association between two variables adjusting for the influence of the remaining variables. They form the backbone for graphical models and are readily obtained from the inverse of the covariance matrix. For compositional data, the covariance structure is specified from log ratios of variables, so unless we try to "open" the data via a normalization, this implies changes in the definition and interpretation of partial correlations. In the present work, we elucidate how results derived by Aitchison (1986) lead to a natural definition of partial correlation that has a number of advantages over current measures of association. For this, we show that the residuals of log-ratios between a variable with a reference, when adjusting for all remaining variables including the reference, are reference-independent. Since the reference itself can be controlled for, correlations between residuals are defined for the variables directly without the necessity to recur to ratios except when specifying which variables are partialled out. Thus, perhaps surprisingly, partial correlations do not have the problems commonly found with measures of pairwise association on compositional data. They are well-defined between two variables, are properly scaled, and allow for negative association. By design, they are subcompositionally incoherent, but they share this property with conventional partial correlations (where results change when adjusting for the influence of fewer variables). We discuss the equivalence with normalization-based approaches whenever the normalizing variables are controlled for. We also discuss the partial variances and correlations we obtain from a previously studied data set of Roman glass cups.

Today, comprehensive evaluation of large-scale machine learning models is possible thanks to the open datasets produced using crowdsourcing, such as SQuAD, MS COCO, ImageNet, SuperGLUE, etc. These datasets capture objective responses, assuming the single correct answer, which does not allow to capture the subjective human perception. In turn, pairwise comparison tasks, in which one has to choose between only two options, allow taking peoples' preferences into account for very challenging artificial intelligence tasks, such as information retrieval and recommender system evaluation. Unfortunately, the available datasets are either small or proprietary, slowing down progress in gathering better feedback from human users. In this paper, we present IMDB-WIKI-SbS, a new large-scale dataset for evaluating pairwise comparisons. It contains 9,150 images appearing in 250,249 pairs annotated on a crowdsourcing platform. Our dataset has balanced distributions of age and gender using the well-known IMDB-WIKI dataset as ground truth. We describe how our dataset is built and then compare several baseline methods, indicating its suitability for model evaluation.

Recent work has shown that asking language models to generate reasoning steps improves performance on many reasoning tasks. When moving beyond prompting, this raises the question of how we should supervise such models: outcome-based approaches which supervise the final result, or process-based approaches which supervise the reasoning process itself? Differences between these approaches might naturally be expected not just in final-answer errors but also in reasoning errors, which can be difficult to detect and are problematic in many real-world domains such as education. We run the first comprehensive comparison between process- and outcome-based approaches trained on a natural language task, GSM8K. We find that pure outcome-based supervision produces similar final-answer error rates with less label supervision. However, for correct reasoning steps we find it necessary to use process-based supervision or supervision from learned reward models that emulate process-based feedback. In total, we improve the previous best results from 16.8% to 12.7% final-answer error and 14.0% to 3.4% reasoning error among final-answer-correct solutions.

Evaluations are critical for understanding the capabilities of large language models (LLMs). Fundamentally, evaluations are experiments; but the literature on evaluations has largely ignored the literature from other sciences on experiment analysis and planning. This article shows researchers with some training in statistics how to think about and analyze data from language model evaluations. Conceptualizing evaluation questions as having been drawn from an unseen super-population, we present formulas for analyzing evaluation data, measuring differences between two models, and planning an evaluation experiment. We make a number of specific recommendations for running language model evaluations and reporting experiment results in a way that minimizes statistical noise and maximizes informativeness.

In the last decade, several organizations have produced documents intended to standardize, in the normative sense, and promote guidance to our recent and rapid AI development. However, the full spectrum of ideas presented in these documents has not yet been analyzed, except for a few meta-analyses and critical reviews of the field. In this work, we seek to expand on the work done by past researchers and create a tool for better data visualization of the contents and nature of these documents, to understand whether there is consensus or similarity between the principles espoused by various institutions, which may inspire debates on future regulations. We also provide some preliminary thoughts and questions that could guide the continuity of the research through a critical analysis of the results acquired by our methodology into a sample size of 200 documents.

Seven years ago, researchers proposed a postprocessing method to equalize the error rates of a model across different demographic groups. The work launched hundreds of papers purporting to improve over the postprocessing baseline. We empirically evaluate these claims through thousands of model evaluations on several tabular datasets. We find that the fairness-accuracy Pareto frontier achieved by postprocessing contains all other methods we were feasibly able to evaluate. In doing so, we address two common methodological errors that have confounded previous observations. One relates to the comparison of methods with different unconstrained base models. The other concerns methods achieving different levels of constraint relaxation. At the heart of our study is a simple idea we call unprocessing that roughly corresponds to the inverse of postprocessing. Unprocessing allows for a direct comparison of methods using different underlying models and levels of relaxation.

Language model evaluation is a daunting task: prompts are brittle, corpus-level perplexities are vague, and the choice of benchmarks are endless. Finding examples that show meaningful, generalizable differences between two LMs is crucial to understanding where one model succeeds and another fails. Can this process be done automatically? In this work, we propose methodology for automated comparison of language models that uses performance-aware contextual embeddings to find fine-grained features of text where one LM outperforms another. Our method, which we name BehaviorBox, extracts coherent features that demonstrate differences with respect to the ease of generation between two LMs. Specifically, BehaviorBox finds features that describe groups of words in fine-grained contexts, such as "conditional 'were' in the phrase 'if you were'" and "exclamation marks after emotional statements", where one model outperforms another within a particular datatset. We apply BehaviorBox to compare models that vary in size, model family, and post-training, and enumerate insights into specific contexts that illustrate meaningful differences in performance which cannot be found by measures such as corpus-level perplexity alone.

Collections of probability distributions arise in a variety of applications ranging from user activity pattern analysis to brain connectomics. In practice these distributions can be defined over diverse domain types including finite intervals, circles, cylinders, spheres, other manifolds, and graphs. This paper introduces an approach for detecting differences between two collections of distributions over such general domains. To this end, we propose the intrinsic slicing construction that yields a novel class of Wasserstein distances on manifolds and graphs. These distances are Hilbert embeddable, allowing us to reduce the distribution collection comparison problem to a more familiar mean testing problem in a Hilbert space. We provide two testing procedures one based on resampling and another on combining p-values from coordinate-wise tests. Our experiments in various synthetic and real data settings show that the resulting tests are powerful and the p-values are well-calibrated.

This study presents a comparative analysis of 55 Wikipedia language editions employing a citation index alongside a synthetic quality measure. Specifically, we identified the most significant Wikipedia articles within distinct topical areas, selecting the top 10, top 25, and top 100 most cited articles in each topic and language version. This index was built on the basis of wikilinks between Wikipedia articles in each language version and in order to do that we processed 6.6 billion page-to-page link records. Next, we used a quality score for each Wikipedia article - a synthetic measure scaled from 0 to 100. This approach enabled quality comparison of Wikipedia articles even between language versions with different quality grading schemes. Our results highlight disparities among Wikipedia language editions, revealing strengths and gaps in content coverage and quality across topics.

As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant through self-improvement, without any human labels identifying harmful outputs. The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'. The process involves both a supervised learning and a reinforcement learning phase. In the supervised phase we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses. In the RL phase, we sample from the finetuned model, use a model to evaluate which of the two samples is better, and then train a preference model from this dataset of AI preferences. We then train with RL using the preference model as the reward signal, i.e. we use 'RL from AI Feedback' (RLAIF). As a result we are able to train a harmless but non-evasive AI assistant that engages with harmful queries by explaining its objections to them. Both the SL and RL methods can leverage chain-of-thought style reasoning to improve the human-judged performance and transparency of AI decision making. These methods make it possible to control AI behavior more precisely and with far fewer human labels.

In this report, we survey Bayesian Optimization methods focussed on the Multi-Armed Bandit Problem. We take the help of the paper "Portfolio Allocation for Bayesian Optimization". We report a small literature survey on the acquisition functions and the types of portfolio strategies used in papers discussing Bayesian Optimization. We also replicate the experiments and report our findings and compare them to the results in the paper. Code link: https://colab.research.google.com/drive/1GZ14klEDoe3dcBeZKo5l8qqrKf_GmBDn?usp=sharing#scrollTo=XgIBau3O45_V.

The rapid advancement of Large Language Models (LLMs) necessitates robust evaluation methodologies. Current benchmarking approaches often rely on comparing model outputs against predefined prompts and reference outputs. Relying on predefined reference outputs hinders flexible adaptation of benchmarks to the rapidly evolving capabilities of LLMs. This limitation necessitates periodic efforts to prepare new benchmarks. To keep pace with rapidly evolving LLM capabilities, we propose a more flexible benchmarking approach. Our method, \textbf{Varco Arena}, provides reference-free benchmarking of LLMs in tournament style. \textbf{Varco Arena} directly compares LLM outputs across a diverse set of prompts, determining model rankings through a single-elimination tournament structure. This direct pairwise comparison offers two key advantages: (1) Direct comparison, unmediated by reference text, more effectively orders competing LLMs, resulting in more reliable rankings, and (2) reference-free approach to benchmarking adds flexibility in updating benchmark prompts by eliminating the need for quality references. Our empirical results, supported by simulation experiments, demonstrate that the \textbf{Varco Arena} tournament approach aligns better with the current Elo model for benchmarking LLMs. The alignment is measured in terms of Spearman correlation, showing improvement over current practice of benchmarking that use reference outputs as comparison anchors.

There is a multitude of novel generative models for open-domain conversational systems; however, there is no systematic evaluation of different systems. Systematic comparisons require consistency in experimental design, evaluation sets, conversational systems and their outputs, and statistical analysis. We lay out a protocol for the evaluation of conversational models using head-to-head pairwise comparison. We analyze ten recent models that claim state-of-the-art performance using a paired head-to-head performance (win-loss-tie) on five evaluation datasets. Our findings show that DialoGPT and Blender are superior systems using Bradley-Terry model and TrueSkill ranking methods. These findings demonstrate the feasibility of our protocol to evaluate conversational agents and evaluation sets. Finally, we make all code and evaluations publicly available for researchers to compare their model to other state-of-the-art dialog models.

In many Reinforcement Learning (RL) papers, learning curves are useful indicators to measure the effectiveness of RL algorithms. However, the complete raw data of the learning curves are rarely available. As a result, it is usually necessary to reproduce the experiments from scratch, which can be time-consuming and error-prone. We present Open RL Benchmark, a set of fully tracked RL experiments, including not only the usual data such as episodic return, but also all algorithm-specific and system metrics. Open RL Benchmark is community-driven: anyone can download, use, and contribute to the data. At the time of writing, more than 25,000 runs have been tracked, for a cumulative duration of more than 8 years. Open RL Benchmark covers a wide range of RL libraries and reference implementations. Special care is taken to ensure that each experiment is precisely reproducible by providing not only the full parameters, but also the versions of the dependencies used to generate it. In addition, Open RL Benchmark comes with a command-line interface (CLI) for easy fetching and generating figures to present the results. In this document, we include two case studies to demonstrate the usefulness of Open RL Benchmark in practice. To the best of our knowledge, Open RL Benchmark is the first RL benchmark of its kind, and the authors hope that it will improve and facilitate the work of researchers in the field.

Cross-lingual evaluation of large language models (LLMs) typically conflates two sources of variance: genuine model performance differences and measurement instability. We investigate evaluation reliability by holding generation conditions constant while varying target language. Using synthetic customer-support dialogues generated with identical parameters across Estonian, Finnish, and Hungarian, we test whether automatic metrics and LLM-as-a-judge scoring produce stable model rankings across these morphologically rich, related Finno-Ugric languages. With a small set of Estonian native speaker annotations as a reference point, we find systematic ranking instabilities: surface-level metrics (lexical diversity, surface and semantic similarity) maintain cross-language stability, but pragmatic judgments (coherence, instruction-following) exhibit rank inversions and near-zero correlations. Because generation is controlled, these inconsistencies reflect how judge scoring behaves differently across languages rather than true model differences. This controlled design provides a diagnostic probe: evaluation methods that fail to maintain stability under identical generation conditions signal transfer failure before deployment. Our findings suggest that zero-shot judge transfer is unreliable for discourse-level assessment in morphologically rich languages, motivating language-specific calibration against targeted human baselines. We release our controlled generation protocol, synthetic data, and evaluation framework to enable replication across language families at https://github.com/isaac-chung/cross-lingual-stability-judges.

Evaluation efforts such as TREC, CLEF, NTCIR and FIRE, alongside public leaderboard such as MS MARCO, are intended to encourage research and track our progress, addressing big questions in our field. However, the goal is not simply to identify which run is "best", achieving the top score. The goal is to move the field forward by developing new robust techniques, that work in many different settings, and are adopted in research and practice. This paper uses the MS MARCO and TREC Deep Learning Track as our case study, comparing it to the case of TREC ad hoc ranking in the 1990s. We show how the design of the evaluation effort can encourage or discourage certain outcomes, and raising questions about internal and external validity of results. We provide some analysis of certain pitfalls, and a statement of best practices for avoiding such pitfalls. We summarize the progress of the effort so far, and describe our desired end state of "robust usefulness", along with steps that might be required to get us there.

In science and medicine, model interpretations may be reported as discoveries of natural phenomena or used to guide patient treatments. In such high-stakes tasks, false discoveries may lead investigators astray. These applications would therefore benefit from control over the finite-sample error rate of interpretations. We reframe black box model interpretability as a multiple hypothesis testing problem. The task is to discover "important" features by testing whether the model prediction is significantly different from what would be expected if the features were replaced with uninformative counterfactuals. We propose two testing methods: one that provably controls the false discovery rate but which is not yet feasible for large-scale applications, and an approximate testing method which can be applied to real-world data sets. In simulation, both tests have high power relative to existing interpretability methods. When applied to state-of-the-art vision and language models, the framework selects features that intuitively explain model predictions. The resulting explanations have the additional advantage that they are themselves easy to interpret.

Chain-of-thought (CoT) monitoring is a promising tool for detecting misbehaviors and understanding the motivations of modern reasoning models. However, if models can control what they verbalize in their CoT, it could undermine CoT monitorability. To measure this undesirable capability -- CoT controllability -- we introduce the CoT-Control evaluation suite, which includes tasks that require models to solve problems while adhering to CoT instructions, e.g., reasoning about a genetics question without using the word 'chromosome'. We show that reasoning models possess significantly lower CoT controllability than output controllability; for instance, Claude Sonnet 4.5 can control its CoT only 2.7% of the time but 61.9% when controlling its final output. We also find that CoT controllability is higher for larger models and decreases with more RL training, test-time compute, and increased problem difficulty. CoT controllability failures extend even to situations in which models are given incentives (as opposed to direct requests) to evade CoT monitors, although models exhibit slightly higher controllability when they are told they are being monitored. Similarly, eliciting controllability by adversarially optimizing prompts does not meaningfully increase controllability. Our results leave us cautiously optimistic that CoT controllability is currently unlikely to be a failure mode of CoT monitorability. However, the mechanism behind low controllability is not well understood. Given its importance for maintaining CoT monitorability, we recommend that frontier labs track CoT controllability in future models.

The MineRL BASALT competition has served to catalyze advances in learning from human feedback through four hard-to-specify tasks in Minecraft, such as create and photograph a waterfall. Given the completion of two years of BASALT competitions, we offer to the community a formalized benchmark through the BASALT Evaluation and Demonstrations Dataset (BEDD), which serves as a resource for algorithm development and performance assessment. BEDD consists of a collection of 26 million image-action pairs from nearly 14,000 videos of human players completing the BASALT tasks in Minecraft. It also includes over 3,000 dense pairwise human evaluations of human and algorithmic agents. These comparisons serve as a fixed, preliminary leaderboard for evaluating newly-developed algorithms. To enable this comparison, we present a streamlined codebase for benchmarking new algorithms against the leaderboard. In addition to presenting these datasets, we conduct a detailed analysis of the data from both datasets to guide algorithm development and evaluation. The released code and data are available at https://github.com/minerllabs/basalt-benchmark .