Research map / Label-efficient and robust learning / Knowledge distillation
From Solver Trajectories to Teaching Trajectories: Cognition-Aligned Reasoning Distillation
Yashuo Luo, Qianren Mao, ZiqiQin, Junnan Liu, Haitao Li, Qili Zhang, Yuan Gao, Zhijun Chen, Xu Wang
Read the paperCheck your idea against 137,217 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.
Argues that raw solver trajectories from large reasoning models are a poor distillation target and proposes cognition-aligned teaching trajectories for reasoning distillation.
Problem
Knowledge distillation from large reasoning models (LRMs) typically uses raw teacher trajectories as supervision, which mismatches the goal of distillation because LRMs are optimized as solvers via reward-driven post-training rather than as teachers.
Approach
Reframes distillation supervision from solver trajectories to teaching trajectories that are aligned with student cognition, rather than reusing the teacher's raw reward-optimized reasoning traces.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| reformat failed-step trajectories with CARD instead of raw SFT | Qwen2.5-7B student · Math average | accuracy: 51.80 -> 54.22 | Improved |
| reformat failed-step trajectories with CARD instead of raw SFT | Qwen2.5-7B student · General OOD average | accuracy: 56.34 -> 71.07 | Improved |
| reformat with CARD instead of raw SFT | Llama3.2-1B student · Math average | accuracy: 11.41 -> 13.77 | Improved |
| add Logical Coherence Reward (ablation) | CARD Reformatter · General OOD average (Qwen2.5-7B) | accuracy: 66.31 -> 71.07 | Improved |
| add answer-distribution KL reward (ablation) | CARD Reformatter · Math average (Qwen2.5-7B) | accuracy: 52.07 -> 54.22 | Improved |
| add rationale-distribution KL reward (ablation) | CARD Reformatter · Math average (Qwen2.5-7B) | accuracy: 52.23 -> 54.22 | Improved |
| add leakage penalty (ablation) | CARD Reformatter · Math average (Qwen2.5-7B) | accuracy: 52.66 -> 54.22 | Improved |
| reformat only FSF>0 trajectories instead of random selection (ablation) | CARD · General OOD average | accuracy: 52.66 -> 71.07 (Qwen2.5-7B) | Improved |
| reformat failed-step trajectories instead of dropping them (ablation) | Vanilla SFT · Math average (Qwen2.5-7B) | accuracy: 51.80 -> 50.93 | Hurt |
| replace failed-step trajectories with clean ones (Replace-FSF) (ablation) | Vanilla SFT · Math average (Qwen2.5-7B) | accuracy: 51.80 -> 52.74 | Improved |
| replace failed-step rationales with TokenSkip compression | Vanilla SFT · Math average (Llama3.2-3B) | accuracy: 31.45 -> 29.36 | Hurt |
| flip 20% of FSF judge decisions (ablation) | CARD · General OOD average (Qwen2.5-7B) | accuracy: 71.07 -> 66.42 | Hurt |
Closest papers by meaning
- The Signal is in the Steps: Local Scoring for Reasoning Data Selection (ICML 2026 · similarity 0.89)
- LARK: Learnability-Grounded Trajectory Selection for Efficient Reasoning Distillation (NeurIPS 2026 · similarity 0.89)
- RLKD: Distilling LLMs’ Reasoning via Reinforcement Learning (AAAI 2026 · similarity 0.89)
- PADD: Path-Aligned Decompression Distillation for Non-Router Teacher to Guide MoE Student Learning (ICML 2026 · similarity 0.88)
- Your Teacher Can’t Help You Here: Combating Supervision Fidelity Decay in On-Policy Distillation (NeurIPS 2026 · similarity 0.87)
- Reinforcement Learning Teachers of Test Time Scaling (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
