Changyue Wang

📢Our work on detecting hallucinations in Large Reasoning Models (LRMs) has been accepted at #AAAI2026 as an Oral Presentation! 📷 We introduce RACE, the first hallucination detection framework specifically designed for opaque-box Large Reasoning Models (LRMs). (1/3)

🎸RACE consistently outperforms existing answer-only baselines across numerous LRMs and datasets. We're excited to share the full details at @RealAAAI Conference! Paper: arxiv.org/abs/2506.04832 Github: github.com/bebr2/RACE Huggingface: huggingface.co/bebr2/RACE-CoT… (3/3)

Existing detectors only check the final answer, missing fatal flaws hidden in the reasoning trace. We built the solution. 👋 RACE (Reasoning and Answer Consistency Evaluation) is the first hallucination detector to jointly evaluate LRM's full reasoning-answer behavior. (2/3)

📢Our work on detecting hallucinations in Large Reasoning Models (LRMs) has been accepted at #AAAI2026 as an Oral Presentation! 📷 We introduce RACE, the first hallucination detection framework specifically designed for opaque-box Large Reasoning Models (LRMs). (1/3)

👉Code: github.com/bebr2/EditCoT EditCoT works by generating an initial CoT and then iteratively refining it using a trained editor based on new knowledge. This flexible approach allows LLMs to adapt reasoning across various tasks and languages without task-specific adjustments.

🎉 Our paper on "Knowledge Editing through Chain-of-Thought" has been accepted at #EMNLP2025 for main conference! We introduce EditCoT, a novel framework that efficiently updates LLMs by editing their Chain-of-thought.@thuir_lab @emnlpmeeting 👉Preprint: arxiv.org/pdf/2412.17727