Research map / Large language models / Agents, code and math
Student-Centered Distillation Narrows the Agentic Gap Between Small and Large LLMs
Yuanjie Lyu, Chengyu Wang, Jun Huang, Tong Xu
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 |
|---|---|---|---|
| replace full-trajectory behavior cloning with SCoRe-SFT on student-generated, first-error-corrected data | Qwen2.5-7B-Instruct · 8 reasoning benchmarks average | average score: 43.0 -> 46.1 | Improved |
| add short-horizon RL with key-step rewards after SCoRe-SFT (ablation) | Qwen2.5-7B-Instruct · 8 reasoning benchmarks average | average score: 46.1 -> 52.3 | Improved |
| use SCoRe-RL instead of BC + GRPO | Qwen2.5-7B-Instruct · 8 reasoning benchmarks average | average score: 48.2 -> 52.3 | Improved |
| use SCoRe-RL instead of BC + GRPO | Qwen3-8B non-thinking · deep search average | average score: 22.2 -> 30.5 | Improved |
| add short-horizon rollouts from the verified prefix in RL (ablation) | SCoRe-RL · 8 reasoning benchmarks average | average score: 48.4 -> 52.3 | Improved |
| add key-step rewards in RL (ablation) | SCoRe-RL · 8 reasoning benchmarks average | average score: 49.7 -> 52.3 | Improved |
| train on MPS-generated data instead of the cold-start BC data only (ablation) | Qwen2.5-7B-Instruct · 8 reasoning benchmarks average | average score: 41.7 -> 46.1 | Improved |
| add RL after SFT on MPS data (ablation) | Hard-to-Teach held-out subset · math and search hard subset | accuracy: math 17.3% -> 24.3%; search 20.4% -> 26.3% | Improved |
| add GRPO full-horizon RL to behavior cloning on deep search | Qwen3-8B non-thinking · deep search average | average score: 22.8 -> 22.2 | No clear effect |
| set key-step reward scale to 0.5 versus 0.3 or 0.7 (ablation) | SCoRe-RL · math benchmarks (2k-sample search) | accuracy | Improved |
Closest papers by meaning
- Progress-Aware Distillation for Mitigating Stagnation in Small Language Model Agents (NeurIPS 2026 · similarity 0.91)
- Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents (NeurIPS 2026 · similarity 0.90)
- Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.89)
- Reinforcement Learning Teachers of Test Time Scaling (NeurIPS 2025 · similarity 0.88)
- Learning from Language Feedback via Variational Policy Distillation (NeurIPS 2026 · similarity 0.88)
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (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.
