---
library_name: pytorch
base_model: microsoft/deberta-v3-large
tags:
- retrieval-augmented-generation
- reranking
- robust-retrieval
- evidence-critic
- corm-rag
- arxiv:2605.01302
---

# CoRM-RAG Evidence Critic

This repository hosts the released Evidence Critic checkpoint for **CoRM-RAG**:

**Beyond Semantic Relevance: Counterfactual Risk Minimization for Robust Retrieval-Augmented Generation**  
Peiyang Liu, Qiang Yan, Ziqiang Cui, Di Liang, Xi Wang, Wei Ye  
arXiv: <https://arxiv.org/abs/2605.01302>

Code: <https://github.com/PeiYangLiu/CoRM-RAG>

## Model Description

CoRM-RAG aligns retrieval with decision safety rather than semantic similarity alone. The Evidence Critic is a lightweight reranking model trained to score whether a document remains useful under cognitively biased query perturbations, such as false premises, confirmation bias, and distracting assumptions.

The released checkpoint uses a `microsoft/deberta-v3-large` backbone and outputs a robustness score for a `(query, document)` pair. It is intended to be used inside the CoRM-RAG pipeline for evidence reranking and risk-aware retrieval.

## Files

```text
critic-v12-mixed/checkpoint-latest/state.pt
```

This file is a PyTorch checkpoint consumed by the CoRM-RAG codebase.

## Usage

Install the code from GitHub and download the checkpoint:

```bash
git clone https://github.com/PeiYangLiu/CoRM-RAG.git
cd CoRM-RAG

huggingface-cli download PeiyangLiu/CoRM-RAG \
  critic-v12-mixed/checkpoint-latest/state.pt \
  --local-dir checkpoints/hf
```

Run evaluation by pointing `CRITIC_PATH` to the downloaded checkpoint:

```bash
CRITIC_PATH=checkpoints/hf/critic-v12-mixed/checkpoint-latest/state.pt bash src/run_eval.sh
```

For training-data construction, critic training, and end-to-end evaluation details, see the GitHub repository.

## Intended Use

This checkpoint is intended for research on robust retrieval-augmented generation, evidence reranking, and risk-aware retrieval under biased or perturbed user queries. It is not a standalone generative model.

## Limitations

The critic score reflects robustness patterns learned from the CoRM-RAG training pipeline and should be interpreted within that retrieval setting. Performance may vary across domains, corpora, retrievers, and perturbation distributions.

## Citation

```bibtex
@misc{liu2026cormrag,
  title={Beyond Semantic Relevance: Counterfactual Risk Minimization for Robust Retrieval-Augmented Generation},
  author={Peiyang Liu and Qiang Yan and Ziqiang Cui and Di Liang and Xi Wang and Wei Ye},
  year={2026},
  eprint={2605.01302},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2605.01302}
}
```