Saved in:
Bibliographic Details
Main Authors: Jiang, Zhaoyang, Peng, Xuanqi, Teng, Fei, Fu, Zhizhong, Kim, Yunsoo, Mi, Jiacong, Li, Zicheng, Wu, Honghan
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.28301
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910266359283712
author Jiang, Zhaoyang
Peng, Xuanqi
Teng, Fei
Fu, Zhizhong
Kim, Yunsoo
Mi, Jiacong
Li, Zicheng
Wu, Honghan
author_facet Jiang, Zhaoyang
Peng, Xuanqi
Teng, Fei
Fu, Zhizhong
Kim, Yunsoo
Mi, Jiacong
Li, Zicheng
Wu, Honghan
contents Chain-of-thought (CoT) distillation trains a smaller model to imitate a teacher's reasoning trace, but it is typically evaluated by final-answer metrics including accuracy. We ask whether gains in answer quality are accompanied by improvements in the trace. In medical QA, where short answer options can leave a richer clinical justification under-specified, a Qwen3-8B student distilled from a DeepSeek-V3-family teacher improves on MedQA-USMLE answer metrics (SC@64 74.7% to 84.4%; expected calibration error (ECE) 0.096 to 0.034). Yet under a Kimi-K2.6 style-blind LLM-judge audit, its error rate over non-abstained steps rises from 30.6% to 50.3%. In this primary medical setting, answer quality and trace factuality move in opposite directions. This before--after pattern persists across evaluators, teacher strengths, student scales and families, medical benchmarks, and style, segmentation, and answer-correctness controls. A 150-step blinded audit by a clinical expert reproduces the same ordering. Boundary checks narrow the scope of the claim: the risk appears when a compact answer under-constrains the rationale and a capable student can imitate expert-like form without reliably grounding each local claim. Standard answer metrics and aggregate hedging rates do not reveal the shift. When such traces are released or reused, answer-level metrics alone are insufficient.
format Preprint
id arxiv_https___arxiv_org_abs_2605_28301
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Better Accuracies, Worse Reasoning: A Step-Level Audit of Medical Chain-of-Thought Distillation
Jiang, Zhaoyang
Peng, Xuanqi
Teng, Fei
Fu, Zhizhong
Kim, Yunsoo
Mi, Jiacong
Li, Zicheng
Wu, Honghan
Artificial Intelligence
Chain-of-thought (CoT) distillation trains a smaller model to imitate a teacher's reasoning trace, but it is typically evaluated by final-answer metrics including accuracy. We ask whether gains in answer quality are accompanied by improvements in the trace. In medical QA, where short answer options can leave a richer clinical justification under-specified, a Qwen3-8B student distilled from a DeepSeek-V3-family teacher improves on MedQA-USMLE answer metrics (SC@64 74.7% to 84.4%; expected calibration error (ECE) 0.096 to 0.034). Yet under a Kimi-K2.6 style-blind LLM-judge audit, its error rate over non-abstained steps rises from 30.6% to 50.3%. In this primary medical setting, answer quality and trace factuality move in opposite directions. This before--after pattern persists across evaluators, teacher strengths, student scales and families, medical benchmarks, and style, segmentation, and answer-correctness controls. A 150-step blinded audit by a clinical expert reproduces the same ordering. Boundary checks narrow the scope of the claim: the risk appears when a compact answer under-constrains the rationale and a capable student can imitate expert-like form without reliably grounding each local claim. Standard answer metrics and aggregate hedging rates do not reveal the shift. When such traces are released or reused, answer-level metrics alone are insufficient.
title Better Accuracies, Worse Reasoning: A Step-Level Audit of Medical Chain-of-Thought Distillation
topic Artificial Intelligence
url https://arxiv.org/abs/2605.28301