Research map / Reinforcement learning / RL from feedback
TMS: Trajectory-Mixed Supervision for On-Policy Self Distillation
Rana Shahroz, Zijie Liu, Zhen Tan, Charles Fleming, Tianlong Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| mix supervision across the training trajectory instead of rewriting labels once (add trajectory mixing vs Rephrase-SFT) (ablation) | TMS · MATH-trained, Qwen-2.5-3B-Instruct, ARC-C | accuracy (%): 60.6 -> 79.2 | Improved |
| add trajectory mixing vs base-snapshot T-matched controls (ablation) | TMS · MATH-trained, Qwen-2.5-3B-Instruct, ARC-C | accuracy (%): 64.7 -> 79.2 | Improved |
| add trajectory mixing vs final-snapshot T-matched controls (ablation) | TMS · MATH-trained, Qwen-2.5-3B-Instruct, ARC-C | accuracy (%): 61.5 -> 79.2 | Improved |
| add trajectory mixing vs base-snapshot controls (ablation) | TMS · MATH-trained, Qwen-2.5-3B-Instruct, cross-task | Cross-task Avg delta (lower better): 14.7 -> 2.3 | Improved |
| add oracle label mixing (alpha 0.25 vs 0.0) (ablation) | TMS · MATH-trained, Qwen-2.5-3B-Instruct, MATH | accuracy (%): 80.6 -> 81.1 | Improved |
| mix oracle labels at alpha 0.25 rather than oracle-only SFT (ablation) | TMS · MATH-trained, Qwen-2.5-3B-Instruct, ARC-C | accuracy (%): 42.7 -> 79.2 | Improved |
| sample checkpoints uniformly rather than early-only (ablation) | TMS · MATH-trained, Qwen-2.5-3B-Instruct, ARC-C | accuracy (%): 68.1 -> 79.2 | Improved |
| sample checkpoints uniformly rather than late-only (ablation) | TMS · MATH-trained, Qwen-2.5-3B-Instruct, ARC-C | accuracy (%): 70.3 -> 79.2 | Improved |
| sample checkpoints uniformly rather than mid-only (ablation) | TMS · MATH-trained, Qwen-2.5-3B-Instruct, ARC-C | accuracy (%): 78.4 -> 79.2 | No clear effect |
| use TMS instead of standard SFT | Qwen-2.5-1.5B-Instruct · ARC-C retention | accuracy (%): 42.3 -> 65.4 | Improved |
| use TMS instead of standard SFT | Qwen-2.5-1.5B-Instruct · MATH Pass@100 | Pass@100 (%): 66.0 -> 74.0 | Improved |
| use TMS instead of standard SFT | LLaMA-3.1-8B-Instruct · ARC-C retention | accuracy (%): 53.2 -> 83.1 | Improved |
Closest papers by meaning
- Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents (NeurIPS 2026 · similarity 0.89)
- Trajectory Bellman Residual Minimization: A Simple Value-Based Method for LLM Reasoning (NeurIPS 2025 · similarity 0.89)
- Learning to Foresee: Unveiling the Unlocking Efficiency of On-Policy Distillation (NeurIPS 2026 · similarity 0.89)
- Procedural Memory Distillation: Online Reflection for Self-Improving Language Models (NeurIPS 2026 · similarity 0.88)
- TTRL: Test-Time Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Learning Agentic Policy from Action Guidance (NeurIPS 2026 · similarity 0.88)
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.
