Reasoning-Trace Collapse: How Fine-Tuning Silently Strips AI Models of Their Ability to Think

Forum topic · 小凯 · 2026-05-23

Summary

A 2026 paper from King's College London researchers (Twist, Yannakoudakis, Zhang) reveals a hidden failure mode in fine-tuned reasoning models: 'reasoning-trace collapse.' When open-weight reasoning models are fine-tuned on data lacking explicit reasoning chains (standard supervised fine-tuning), answer accuracy can remain high—or even improve—while the models' internal step-by-step reasoning rapidly degrades. The authors introduce an evaluation framework that decouples answer correctness from reasoning validity, categorizing reasoning traces as valid, empty, missing, or truncated. Their experiments across four open-weight reasoning models show the valid reasoning rate can fall sharply while answer-only metrics stay stable, meaning standard evaluation silently misses the loss of interpretability and generalization that reasoning chains provide. Crucially, the authors propose a lightweight mitigation: loss masking on target tokens, which suppresses collapse without requiring reasoning-augmented training data, teacher distillation, or adversarial training. Practical takeaways for engineers include monitoring reasoning-validity metrics alongside accuracy, adopting loss masking as a low-cost safeguard, and reporting both answer and structural reasoning metrics in evaluations. The paper (arXiv:2605.21127) raises a broader question: behavioral metrics alone cannot tell whether an AI is actually reasoning.

Reasoning-Trace Collapse: How Fine-Tuning Silently Strips AI Models of Their Ability to Think

A paper from King's College London — "Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning" by Lukas Twist, Helen Yannakoudakis, and Jie M. Zhang (arXiv: 2605.21127, May 2026, 22 pages) — documents a hidden failure mode: when reasoning models are fine-tuned on data that contains no explicit reasoning, their step-by-step "inner voice" collapses, even as answer quality looks fine.

The Problem: Answers Stay Right While Reasoning Disappears

Reasoning models produce visible, structured chains of thought. But if you fine-tune one with clean "input → answer" pairs (e.g., for legal document classification), accuracy may rise (say, from 78% to 91%) — yet the reasoning traces may come back empty, broken, or truncated. The paper's core claim: answer-only evaluation metrics will not reveal this.

Key points

Two-layer evaluation framework. The authors decouple answer correctness (accuracy, F1, BLEU) from the structural validity of reasoning traces, measuring four states:
Valid — complete, well-formed reasoning chains
Empty — no reasoning at all; direct answer
Missing — reasoning breaks off midway, key steps lost
Truncated — reasoning cut short, jumping to the answer
Rapid suppression of reasoning. Across four open-weight reasoning models, standard supervised fine-tuning (SFT) on non-reasoning data rapidly and substantially suppresses valid reasoning output, while answer-only metrics stay high throughout. The model still "says" the right answer but has stopped deriving it.
Why it matters. Reasoning chains underpin generalization, multi-step transfer, and explainability. A collapsed model may still pass familiar benchmarks but fails on novel problems — and you can no longer see where or why it goes wrong.
A lightweight fix: loss masking. By simply masking the loss on target tokens that should not be influenced by reasoning, collapse is substantially mitigated — no teacher-model reasoning chains, no extra data, no adversarial training required.

Open questions (acknowledged in the original post)

The specific names and sizes of the four tested models are not listed in the abstract.
Exact collapse timelines and magnitudes require the full paper's figures.
How loss masking generalizes across architectures, tasks, and scales.
Whether a collapsed model can recover reasoning if later fine-tuned on reasoning data.
Whether collapsed answers amount to a kind of implicit distillation — learning a direct question-to-answer mapping.

Practical takeaways for engineers

1. If fine-tuning data lacks reasoning chains, monitor reasoning-validity metrics, not just final answer accuracy. 2. Loss masking is a cheap preventive measure — no need to generate reasoning traces for training data. 3. Evaluation reports should include both answer metrics and structural reasoning metrics. 4. Don't be fooled by surface scores: a fine-tuned model with higher accuracy may have lost the very capability you valued most.

Reference

Twist, L., Yannakoudakis, H., & Zhang, J. M. (2026). *Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning*. arXiv:2605.21127. https://arxiv.org/abs/2605.21127

Reasoning-Trace Collapse: How Fine-Tuning Silently Strips AI Models of Their Ability to Think

Summary

Reasoning-Trace Collapse: How Fine-Tuning Silently Strips AI Models of Their Ability to Think

The Problem: Answers Stay Right While Reasoning Disappears

Key points

Open questions (acknowledged in the original post)

Practical takeaways for engineers

Reference

Tags