Research map / Large language models / Reasoning and chain of thought
Failure Profiling and Reachable Trajectory Selection for Reasoning Distillation
yijun Zhu, Jianxin Wang, Chengchao Shen
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.
Proposes failure-profiling and reachable-trajectory selection to improve reasoning distillation from a larger teacher to a smaller student for math reasoning.
Problem
Students distilled from a larger teacher's reasoning solutions overfit the teacher's output distribution and generalize poorly when their own inference-time behavior deviates from the training data; prior error-data methods only partially address this.
Approach
Profiles the student's failure modes and selects teacher trajectories that the student can actually reach (reachable trajectories), rather than simply adding error data or using all teacher outputs. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add behavioral failure profiling with mode-specific teacher prompts (ablation) | FPRT · Five-benchmark average, Llama-3.1-8B | accuracy (%): 72.9 -> 79.0 | Improved |
| add branching point reachability (BPR) selection instead of global log-prob scoring (ablation) | FPRT · Five-benchmark average, Llama-3.1-8B | accuracy (%): 77.3 -> 79.0 | Improved |
| add BPR on top of uniform prompting (no profiling) (ablation) | FPRT · Five-benchmark average, Llama-3.1-8B | accuracy (%): 72.4 -> 72.9 | No clear effect |
| BPR scoring versus global scoring (ablation) | FPRT · MAWPS, Llama-3.1-8B | accuracy (%): 93.9 vs 94.2 global | Hurt |
| select by BPR instead of random correct candidate (ablation) | FPRT · Five-benchmark average, Llama-3.1-8B | accuracy (%): 76.6 -> 79.0 | Improved |
| select by BPR instead of random correct candidate (ablation) | FPRT · Five-benchmark average, DeepSeek-Math-7B | accuracy (%): 79.0 -> 80.0 | Improved |
| use full failure-profiled pipeline instead of direct teacher-trace SFT (ablation) | FPRT · Five-benchmark average, Llama-3.1-8B | accuracy (%): 75.9 -> 79.0 | Improved |
| set BPR window to 8 sentences instead of shorter or longer (ablation) | FPRT · Five-benchmark average, Llama-3.1-8B | accuracy (%): L=2 77.1, L=4 76.7, L=8 79.0, L=16 78.6, L=32 78.2 | Improved |
| scale corpus from 16.5K to 88.90K examples | FPRT · Llama-3.1-8B average | accuracy (%): 71.8 -> 79.7 | Improved |
| scale corpus from 16.5K to 88.90K examples | FPRT · DeepSeek-Math-7B average | accuracy (%): 77.9 -> 80.6 | Improved |
| FPRT versus LEMMA on competition-level math | FPRT · MATH, Llama-3.1-8B | accuracy (%): 44.3 vs 48.8 (62.74K); 46.2 at 88.90K | Hurt |
| switch teacher from DeepSeek-V3.2 to DeepSeek-V4-Flash (ablation) | FPRT · DeepSeek-Math-7B average | accuracy (%): 80.6 -> 81.0 | No clear effect |
Closest papers by meaning
- The Signal is in the Steps: Local Scoring for Reasoning Data Selection (ICML 2026 · similarity 0.90)
- LARK: Learnability-Grounded Trajectory Selection for Efficient Reasoning Distillation (NeurIPS 2026 · similarity 0.90)
- RE-IMAGINE: Symbolic Benchmark Synthesis for Reasoning Evaluation (ICML 2025 · similarity 0.89)
- Physics of Language Models: Part 2.2, How to Learn From Mistakes on Grade-School Math Problems (ICLR 2025 · similarity 0.89)
- SuperCorrect: Advancing Small LLM Reasoning with Thought Template Distillation and Self-Correction (ICLR 2025 · similarity 0.89)
- MARGE: Improving Math Reasoning with Guided Exploration (ICML 2025 · similarity 0.89)
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.
