| --- |
| license: apache-2.0 |
| language: |
| - en |
| metrics: |
| - macro-f1 |
| library_name: transformers |
| base_model: |
| - Qwen/Qwen2.5-7B-Instruct |
| tags: |
| - fact-checking |
| - misinformation |
| - reinforcement-learning |
| - grpo |
| - retrieval-augmented-generation |
| - iterative-retrieval |
| - multi-hop-reasoning |
| - llm |
| --- |
| # Model Card |
|
|
| <!-- Provide a quick summary of what the model is/does. --> |
| R³Check is a large language model specifically trained for complex multi-hop fact-checking through iterative evidence retrieval and structured reasoning. |
|
|
| It is built upon [Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) and trained using a rule-guided reinforcement learning framework based on Group Relative Policy Optimization (GRPO). Unlike supervised fine-tuning approaches that require annotated reasoning trajectories, R³Check relies only on final claim-veracity labels and rule-based format rewards. |
|
|
| The model follows a structured retrieval–reasoning protocol: |
|
|
| 1. **Structured reasoning:** The model analyzes the claim and decomposes it into verifiable components within `<think>` tags. |
| 2. **Iterative retrieval:** When additional evidence is required, the model generates search queries within `<search>` tags and receives retrieved passages through `<information>` tags. |
| 3. **Evidence-grounded verification:** The model reasons over the accumulated evidence and may issue additional searches when information is incomplete. |
| 4. **Final prediction:** The model produces a binary veracity judgment, `True` or `False`, within `<answer>` tags. |
|
|
| R³Check is trained using a two-stage difficulty-incremental curriculum: |
|
|
| - **Stage 1 — Reasoning Reinforcement:** The model learns structured, evidence-grounded reasoning in a closed fact-checking setting using gold evidence. |
| - **Stage 2 — Reasoning–Retrieval Coordination:** The model jointly learns iterative retrieval and reasoning by interacting with an external retriever in real time. |
|
|
| An importance-based bucket sampling strategy is additionally used to emphasize informative training examples and improve the effectiveness of GRPO optimization. |
|
|
|
|
| ## Model Sources |
|
|
| <!-- Provide the basic links for the model. --> |
| - **Paper:** https://dl.acm.org/doi/10.1145/3805712.3809601 |
| - **Project:** https://pengqi.website/R3Check/ |
|
|
|
|
| ## Citation |
|
|
| <!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. --> |
| ``` |
| @inproceedings{qi2026r3check, |
| author = {Qi, Peng and Zhao, Yuyang and Hsu, Wynne and Lee, Mong Li}, |
| title = {R3Check: Reinforcement Learning for Iterative Retrieval and Structured Reasoning in Complex Fact Checking}, |
| booktitle = {Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval}, |
| year = {2026}, |
| pages = {1473--1484}, |
| publisher = {Association for Computing Machinery}, |
| doi = {10.1145/3805712.3809601} |
| } |
| ``` |
|
|
