Research map / Large language models / Instruction and fine-tuning

Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning

NeurIPS 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that standard SFT on instruction-response data without reasoning traces causes reasoning models to lose valid explicit reasoning traces (reasoning-trace collapse), and that loss masking mitigates this without teacher traces, with a structural evaluation framework and the ThinkPack library.

Problem

Fine-tuning reasoning models on instruction-response data lacking explicit reasoning traces can silently erase their structurally valid reasoning, which answer-only metrics fail to reveal.

Approach

Defines reasoning-trace collapse and a structural evaluation framework (valid, empty, missing, truncated reasoning rates plus reasoning-conditioned pass rate), and shows that simple loss masking of empty/missing reasoning regions preserves valid reasoning as an alternative to teacher distillation.

Key results

• On Chemistry, Qwen3-8B and Nemotron-7B reach 0% valid reasoning under standard SFT while pass rate keeps improving to 50-57%.
• Llama-R1-8B drops to 0% Chemistry valid reasoning under no-think formatting but keeps 71% under empty-think.
• Masking strategies preserve substantially more valid reasoning than standard fine-tuning; distillation is strongest for Qwen3-8B and Llama-R1-8B but performs poorly for Olmo-3-7B (GSM8K VR 22%).

Surprising findings

• Model-default formatting is not reliably protective: Qwen3-8B is more stable with no-think, while Olmo-3-7B and Llama-R1-8B preserve reasoning better with empty-think.
• Teacher distillation with GPT-5-mini traces performs worst for Olmo-3-7B, failing to preserve GSM8K reasoning.

Limitations

• Only four small reasoning models and one main fine-tuning dataset; external validity limited.
• Evaluation on fixed 256-example subsets with greedy decoding and two seeds (ablations single seed).
• Metrics are structural only and do not establish correctness or faithfulness of traces.

Experimental setup

Datasets: SciKnowEval Chemistry L-3 GSM8K EvalPlus (HumanEval+, MBPP+) OrcaMath SelfOSS

Benchmarks: SciKnowEval Chemistry L-3 GSM8K EvalPlus

Metrics: valid reasoning rate empty reasoning rate missing reasoning rate truncated reasoning rate pass@1 reasoning-conditioned pass rate

Baselines: standard fine-tuning (empty-think) no-think formatting GPT-5-mini teacher distillation

Builds on: LoRA AdamW chain-of-thought DeepSeek-R1 distillation Hugging Face transformers

Closest papers by meaning

• Probing the Trajectories of Reasoning Traces in Large Language Models (NeurIPS 2026 · similarity 0.90)
• Reasoning or Retrieval? A Study of Answer Attribution on Large Reasoning Models (ICLR 2026 · similarity 0.90)
• Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.90)
• Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens (NeurIPS 2026 · similarity 0.90)
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.89)
• Hint Tuning: Less Data Makes Better Reasoners (NeurIPS 2026 · similarity 0.89)