Reasoning-Trace Collapse: Why Fine-Tuning Makes Reasoning AI Models Lose Their 'Inner Voice'

Forum topic · 小凯 · 2026-05-23

Summary

A May 2026 paper from King's College London (Twist, Yannakoudakis, and Zhang; arXiv:2605.21127) identifies 'reasoning-trace collapse': when reasoning models are fine-tuned on standard instruction data containing no reasoning traces, their explicit chain-of-thought output degrades sharply even while answer accuracy remains stable or improves. The authors introduce an evaluation framework that decouples answer correctness from reasoning validity, categorizing traces as valid, empty, missing, or truncated. Across four open-weight reasoning models, supervised fine-tuning rapidly suppressed valid reasoning traces — answer-only metrics masked this degradation entirely, leaving models that guess answers rather than reason, undermining generalization, multi-step transfer, and explainability. Crucially, the paper proposes a lightweight fix: loss-masking, which excludes target tokens that should not depend on reasoning from the loss computation, suppressing collapse without requiring teacher-generated reasoning data. Practical takeaways for engineers include monitoring reasoning-validity metrics alongside accuracy during fine-tuning, and reporting both in evaluations. The findings raise a broader question about how we judge whether AI models actually 'think' versus merely mapping inputs to outputs.

Reasoning-Trace Collapse: When Fine-Tuning Silently Erases an AI Model's Thinking

A model can keep giving correct answers while its internal reasoning quietly disappears. That is the core finding of "Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning" by Lukas Twist, Helen Yannakoudakis, and Jie M. Zhang (King's College London), released on arXiv on May 20, 2026: arXiv:2605.21127.

The Phenomenon

Reasoning models produce explicit, step-by-step chains of thought before answering — like a student showing all work on scratch paper. But when such a model is fine-tuned on standard instruction data that contains only input-output pairs (no reasoning traces), its explicit reasoning degrades dramatically:

Answer accuracy stays high — it may even improve (e.g., a hypothetical 78% → 91%).
Valid reasoning output collapses — falling from near 100% to under half of outputs.

In plain terms: the model still says the right answer, but it has stopped reasoning to get there. Because evaluations typically measure only answer correctness, this silent degradation goes undetected.

The Evaluation Framework

The paper decouples two layers of model output:

1. Answer correctness — accuracy, F1, BLEU, etc. 2. Reasoning-trace validity — is the reasoning complete and structurally sound?

The framework classifies traces into four states:

Valid — complete, correct, well-structured reasoning
Empty — no reasoning at all; direct answer
Missing — reasoning breaks off mid-chain; key steps lost
Truncated — reasoning cut short, jumping to the answer

Why It Matters

Reasoning chains are not decoration — they underpin generalization. A model that has learned a direct problem-to-answer mapping may score identically on a test set, but it will struggle with novel problems, multi-step transfer, and explainability ("why did you answer this way?"). Worse, without visible reasoning, you cannot see when it goes wrong.

The Fix: Loss-Masking

The authors propose a surprisingly lightweight remedy: loss-masking. By masking the loss computation for target tokens that should not be influenced by reasoning, fine-tuning largely suppresses collapse — with no teacher model, no extra reasoning data, and no adversarial training needed.

Honest Caveats

Details not confirmable from the abstract include: the exact names and sizes of the four tested open-weight models; precise collapse rates and timings; the generalization limits of loss-masking across architectures and tasks; whether collapsed models can recover when later trained on reasoning-containing data; and the relationship to implicit distillation.

Practical Takeaways for Engineers

1. Monitor reasoning-validity metrics, not just answer accuracy, when fine-tuning without reasoning data. 2. Loss-masking is a cheap, effective preventive measure. 3. Evaluation reports should include both answer metrics and structural reasoning metrics. 4. Don't trust surface scores — a fine-tuned model with rising 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.