Reasoning models think longer, but longer chains bury more factual mistakes-and one wrong fact early on poisons everything that follows. What if you could catch and fix those errors while the model is still reasoning? Introducing CheckRLM, from @TsinghuaNLP (OpenBMB member) with Beijing Normal University, the Institute of Information Engineering (CAS), University of Chinese Academy of Sciences, and Northeastern University. It's a RAG framework that checks and corrects factual errors inside the reasoning chain-before they cascade into a wrong answer. 1⃣️ In-process, not post-hoc. Instead of waiting for the full chain to finish, CheckRLM extracts explicit factual claims from each reasoning paragraph as it's generated, then localizes potential knowledge inconsistencies right where they appear-catching errors early rather than auditing a corrupted trajectory after the fact. Checks run at the paragraph level to balance context completeness against timely verification. 2⃣️ Localized, minimal-cost correction via retrieval. It builds a retrieval query set from the extracted claims plus the original question, pulls top-k documents (deduplicated), and feeds them with the current reasoning unit into a correction model that makes token-level fixes only-preserving the reasoning structure and leaving correct or irrelevant content untouched. 3⃣️ Jointly optimized with DPO. A single unified model handles both claim recognition and coherence correction, trained on a small high-quality preference set (5,311 samples from 2WikiMQA, annotated by GPT-4o-mini)-yielding +5.7 f1 on in-domain 2WikiMQA plus gains on out-of-domain HotpotQA and SimpleQA over inference-only. 4⃣️ Better and cheaper. On 2WikiMQA with QwQ-32B, CheckRLM hits 73.4 f1 while using fewer tokens (1364 vs 2008) and less time (3.3s vs 3.9s) than Search-o1-beating the strongest baseline by +6.3 f1 on the hardest MuSiQue set and topping every one of five QA benchmarks. arXiv: https://arxiv.org/abs/2607.02262 Code & data: https://github.com/AI9Stars/CheckRLM #AI #THUNLP #OpenBMB #LLM #RAG #ReasoningModels #NLP
清华NLP等机构推出CheckRLM：推理中实时纠错RAG框架
清华NLP等机构推出CheckRLM，一个在推理过程中实时检查并纠正事实错误的RAG框架，而非事后审计。在2WikiMQA上，CheckRLM用QwQ-32B达到73.4 F1，比Search-o1用更少token（1364 vs 2008）和更短时间（3.3s vs 3.9s），并在最难MuSiQue集上超越最强基线+6.3 F1。
Reasoning models think longer, but longer chains bury more factual mistakes-and one wrong fact early on poisons everything that follows. What if you could catch and fix those errors while the model is still reasoning? Introducing CheckRLM, from @TsinghuaNLP (OpenBMB member) with Beijing Normal University, the Institute of Information Engineering (CAS), University of Chinese Academy of Sciences, and Northeastern University. It's a RAG framework that checks and corrects factual errors inside the reasoning chain-before they cascade into a wrong answer. 1⃣️ In-process, not post-hoc. Instead of waiting for the full chain to finish, CheckRLM extracts explicit factual claims from each reasoning paragraph as it's generated, then localizes potential knowledge inconsistencies right where they appear-catching errors early rather than auditing a corrupted trajectory after the fact. Checks run at the paragraph level to balance context completeness against timely verification. 2⃣️ Localized, minimal-cost correction via retrieval. It builds a retrieval query set from the extracted claims plus the original question, pulls top-k documents (deduplicated), and feeds them with the current reasoning unit into a correction model that makes token-level fixes only-preserving the reasoning structure and leaving correct or irrelevant content untouched. 3⃣️ Jointly optimized with DPO. A single unified model handles both claim recognition and coherence correction, trained on a small high-quality preference set (5,311 samples from 2WikiMQA, annotated by GPT-4o-mini)-yielding +5.7 f1 on in-domain 2WikiMQA plus gains on out-of-domain HotpotQA and SimpleQA over inference-only. 4⃣️ Better and cheaper. On 2WikiMQA with QwQ-32B, CheckRLM hits 73.4 f1 while using fewer tokens (1364 vs 2008) and less time (3.3s vs 3.9s) than Search-o1-beating the strongest baseline by +6.3 f1 on the hardest MuSiQue set and topping every one of five QA benchmarks. arXiv: https://arxiv.org/abs/2607.02262 Code & data: https://github.com/AI9Stars/CheckRLM #AI #THUNLP #OpenBMB #LLM #RAG #ReasoningModels #NLP
来源：OpenBMB· x.com