Better Accuracies, Worse Reasoning: A Step-Level Audit of Medical Chain-of-Thought Distillation

• URL: http://arxiv.org/abs/2605.28301v1
• Date: Wed, 27 May 2026 10:55:24 GMT
• Title: Better Accuracies, Worse Reasoning: A Step-Level Audit of Medical Chain-of-Thought Distillation
• Abstract summary: Chain-of-thought (CoT) distillation trains a smaller model to imitate a teacher's reasoning trace.<n>We ask whether gains in answer quality are accompanied by improvements in the trace.
• Score: 14.101636730819175
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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.

Related papers

• Route, Don't Fix: Regime-Dependent Decoding Correction and a Trajectory-Gated Router for Reliable Clinical LLM Answer Selection [1.4358829531570645]
  Large language models (LLMs) are often deemed unsafe for clinical question answering because of their tendency to hallucinate.<n>Inference-time correction uses the model's internal logit signals, but a fixed transformation need not suit every question.<n>We introduce ALTAS, which reads terminal entropy and late-layer linearity from one forward pass to choose per question between greedy decoding and late-layer trajectory correction.
  arXiv  Detail & Related papers  (2026-09-13T22:36:37Z)
• The Answer Is Not the Argument [0.00897008283486962]
  Chain-of-thought monitoring is proposed for AI oversight, yet evaluations often provide monitors with a trusted reference answer.<n>We ask whether answer access improves reasoning verification or mainly exposes incorrect conclusions.
  arXiv  Detail & Related papers  (2026-08-31T19:06:44Z)
• Right Diagnoses, Decorative Reasoning:A Perturbation Audit of Medical Chain-of-Thought [20.946393640258997]
  Clinicians read chain-of-thought (CoT) rationales as evidence of medical reasoning, but whether the visible chain plays that role is rarely tested.<n>Three independent tests converge: the Chain-Decoupling Rate (CDR; chain does not register the edit and the answer does not flip) is 72.9% panel-wide on clinically meaningful destructive edits, chain corruption leaves accuracy unchanged, and removing CoT prompting does not reduce accuracy.
  arXiv  Detail & Related papers  (2026-08-25T16:37:39Z)
• Cybersecurity Detection Classification with Reasoning-enabled Language Models [50.72675435043546]
  A major issue in Security Operations Centers (SOCs) is alert fatigue.<n>Prior work prompts or fine-tunes large language models (LLMs) to emit a triage label directly, but does not train them to reason about whether a detection is a genuine threat.<n>We train a chain-of-thought (CoT) reasoning-enabled triage classifier on real, human-labeled Windows endpoint detections.
  arXiv  Detail & Related papers  (2026-07-30T16:22:13Z)
• Outcome-Confounded Local Supervision in On-Policy Distillation [0.5937989874256571]
  On-policy distillation trains a student on its own trajectories while a teacher supplies dense token-level likelihoods at student-visited prefixes.<n>We show that both readings are confounded by the outcome of the completed trajectory.<n>We introduce an outcome-resolved diagnostic that crosses pointwise teacher-student divergence with final-answer correctness.
  arXiv  Detail & Related papers  (2026-07-26T16:03:33Z)
• When Confidence Takes the Wrong Path: Diagnosing Retrieval-State Lock-In in RAG [0.0]
  Many black-box uncertainty methods still read agreement among sampled answers as confidence.<n>The problem is recognised in deployed RAG, but it has lacked a name, a measurable signature, and a prevalence bound.<n>We name the failure retrieval-state lock-in and diagnose it by separating the three objects a single confidence score conflates.
  arXiv  Detail & Related papers  (2026-06-22T00:08:00Z)
• Beyond Scalar Scores: Exploring LLM-based Metrics for Clinical Significance Evaluation in Radiology Reports [49.5225801722164]
  Existing metrics obscure this requirement by reducing report quality to a medically ungrounded scalar.<n>We study this boundary using ReEvalMed benchmark as testbed and evaluate metric-level clinical significance.
  arXiv  Detail & Related papers  (2026-06-17T08:10:30Z)
• MAIGO: Mitigating Lost-in-Conversation with History-Cleaned On-Policy Self-Distillation [51.32912774464992]
  Large language models often solve tasks from a fully specified prompt but degrade when the same requirements unfold over multiple turns.<n>We propose MAIGO, an on-policy self-distillation method that reduces this contamination using history-cleaned references from the model's own policy.<n>For middle turns, MAIGO removes prior assistant replies while preserving the user-visible sharded prefix; for answer turns, it distills from paired full-view references conditioned on the completed user-side dialogue.
  arXiv  Detail & Related papers  (2026-05-26T15:38:46Z)
• DART: Mitigating Harm Drift in Difference-Aware LLMs via Distill-Audit-Repair Training [18.22030439716779]
  Large language models (LLMs) tuned for safety often avoid acknowledging demographic differences.<n>This identity-blindness yields incorrect responses, unnecessary refusals, or generic "equal-treatment" defaults.<n>We introduce DART (Distill--Audit-Repair Training), which distills label-conditioned reasoning from a teacher.
  arXiv  Detail & Related papers  (2026-04-18T05:28:53Z)
• Benchmarking Multi-turn Medical Diagnosis: Hold, Lure, and Self-Correction [72.89352076103889]
  Large language models (LLMs) achieve high accuracy in medical diagnosis when all clinical information is provided in a single turn.<n>We introduce MINT, a high-fidelity, multi-turn medical diagnosis benchmark comprising 1,035 cases.<n>We uncover three persistent behavioral patterns that significantly impact diagnostic decisions.
  arXiv  Detail & Related papers  (2026-04-06T00:23:10Z)
• Guideline-Grounded Evidence Accumulation for High-Stakes Agent Verification [60.18369393468405]
  Existing verifiers usually underperform owing to a lack of domain knowledge and limited calibration.<n>GLEAN compiles expert-curated protocols into trajectory-informed, well-calibrated correctness signals.<n>We empirically validate GLEAN with agentic clinical diagnosis across three diseases from the MIMIC-IV dataset.
  arXiv  Detail & Related papers  (2026-03-03T09:36:43Z)
• GATES: Self-Distillation under Privileged Context with Consensus Gating [89.62339954332248]
  We study self-distillation in settings where supervision is unreliable.<n>We focus on document-grounded question answering with asymmetric context.<n>We derive supervision online from tutor consensus by sampling multiple document-grounded reasoning traces.
  arXiv  Detail & Related papers  (2026-02-24T05:56:20Z)
• Measuring Stability Beyond Accuracy in Small Open-Source Medical Large Language Models for Pediatric Endocrinology [34.80893325510028]
  Small open-source medical large language models (LLMs) offer promising opportunities for low-resource deployment and broader accessibility.<n>We use coupled to human evaluation and clinical review to assess six small open-source medical LLMs.
  arXiv  Detail & Related papers  (2025-12-26T14:30:53Z)
• Evaluating Reasoning Faithfulness in Medical Vision-Language Models using Multimodal Perturbations [19.488236277427358]
  Vision-language models (VLMs) often produce chain-of-thought (CoT) explanations that sound plausible yet fail to reflect the underlying decision process.<n>We present a clinically grounded framework for chest X-ray visual question answering (VQA) that probes CoT faithfulness via controlled text and image modifications.
  arXiv  Detail & Related papers  (2025-10-13T09:28:22Z)
• Faithful Knowledge Distillation [75.59907631395849]
  We focus on two crucial questions with regard to a teacher-student pair: (i) do the teacher and student disagree at points close to correctly classified dataset examples, and (ii) is the distilled student as confident as the teacher around dataset examples? These are critical questions when considering the deployment of a smaller student network trained from a robust teacher within a safety-critical setting.
  arXiv  Detail & Related papers  (2023-06-07T13:41:55Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.