Research map / Reinforcement learning / RL from feedback
Reinforcement-aware Knowledge Distillation for LLM Reasoning
Zhaoyang Zhang, Shuli Jiang, Yantao Shen, Yuting Zhang, Dhananjay Ram, Shuo Yang, Zhuowen Tu, Wei Xia, Stefano Soatto
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 |
|---|---|---|---|
| use RLAD (TRRD distillation) instead of plain GRPO | Qwen3-0.6B · K&K Logistics (8K) | average accuracy: 0.76 -> 0.94 | Improved |
| use RLAD (TRRD distillation) instead of plain GRPO | Qwen3-1.7B · K&K Logistics (8K) | average accuracy: 0.95 -> 0.99 | Improved |
| use RLAD (TRRD distillation) instead of plain GRPO | Qwen3-0.6B · K&K Logistics (2K) | average accuracy: 0.70 -> 0.90 | Improved |
| use RLAD instead of GRPO on hardest split | Qwen3-0.6B · K&K Logistics PPL8 (8K) | accuracy: 0.63 to 0.83 | Improved |
| use RLAD instead of GRPO | Qwen3-8B-Base · Long-context math avg (30K) | average score: 61.0 -> 66.5 | Improved |
| use RLAD instead of GRPO | Qwen3-1.7B-Base · Long-context math avg (30K) | average score: 36.5 -> 39.0 | Improved |
| use RLAD instead of GRPO | Qwen3-8B-Base · AIME25-Pass@32 | Pass@32: 48.5 to 66.4 | Improved |
| use RLAD instead of GRPO | Qwen3-8B-Base · MATH500 (30K) | Mean@32: 68.2 to 68.6 | No clear effect |
| use RLAD instead of GRPO | Qwen3-1.7B (post-trained) · Math avg (8K) | average score: 51.2 to 57.8 | Improved |
| add KDRL distillation from a weaker teacher (Qwen3-0.6B) (ablation) | Qwen3-1.7B · K&K Logistics (8K) | average accuracy: 0.950 -> 0.910 | Hurt |
| add RLAD distillation from a weaker teacher (Qwen3-0.6B) (ablation) | Qwen3-1.7B · K&K Logistics (8K) | average accuracy: 0.950 -> 0.946 | No clear effect |
| raise mixture coefficient alpha from 0.5 to 0.9 (ablation) | RLAD · MATH | Pass@1: 80.5 -> 77.4 | Hurt |
Closest papers by meaning
- RLKD: Distilling LLMs’ Reasoning via Reinforcement Learning (AAAI 2026 · similarity 0.92)
- Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models (ICML 2026 · similarity 0.92)
- Reinforcement Learning Teachers of Test Time Scaling (NeurIPS 2025 · similarity 0.92)
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.91)
- Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective (NeurIPS 2025 · similarity 0.91)
- KL for a KL: On-Policy Distillation with Control Variate Baseline (NeurIPS 2026 · similarity 0.91)
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.
