Research map / Large language models / Instruction and fine-tuning
SkillFactory: Self-Distillation for Learning Cognitive Behaviors
Zayne Rea Sprague, Jack Lu, Manya Wadhwa, Sedrick Keh, Mengye Ren, Greg Durrett
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 SkillFactory SFT init before GRPO instead of RL only | Qwen2.5-1.5B-Instruct · Countdown 4-6arg | accuracy: 15.8 -> 25.1 | Improved |
| SkillFactory SFT init vs R1 distillation init before GRPO | Qwen2.5-1.5B-Instruct · Countdown 4-6arg | accuracy: 21.2 -> 25.1 | Improved |
| SkillFactory SFT alone has lower accuracy than R1 distillation SFT but better post-RL | Qwen2.5-1.5B-Instruct · Countdown 4-6arg | accuracy: SFT 2.8 vs 11.7; post-RL 25.1 vs 21.2 | Mixed |
| STaR self-distillation init before GRPO | Qwen2.5-1.5B-Instruct · Countdown 4-6arg | accuracy: 15.8 -> 9.7 | Hurt |
| SkillFactory init vs RL only on out-of-domain average | Qwen2.5-1.5B-Instruct · OOD tasks overall | overall accuracy: 31.9 -> 35.7 | Improved |
| SkillFactory init vs RL only on OOD average with larger models | Qwen2.5-7B-Instruct · OOD tasks overall | overall accuracy: 32.9 -> 54.1 (R1 Distill -> GRPO: 70.1) | Improved |
| add reflections to silver traces (ablation) | SkillFactory · OOD tasks overall (ablation eval) | overall accuracy: 23.3 -> 32.0 | Improved |
| order samples so traces end on a correct sample (ablation) | SkillFactory · OOD tasks overall (ablation eval) | overall accuracy: 24.1 -> 32.0 | Improved |
| use diverse solve prompts rather than a single prompt (ablation) | SkillFactory · OOD tasks overall (ablation eval) | overall accuracy: 29.0 -> 32.0 | Improved |
| add SFT on structured traces instead of only an instruction prompt (ablation) | RL on Qwen2.5-1.5B-Instruct · OOD tasks overall (ablation eval) | overall accuracy: 25.8 -> 32.0 | Improved |
| scale SkillFactory SFT data from 1k to 10k rows (ablation) | Qwen2.5-7B-Instruct · OpenThoughts math/science overall | overall accuracy: 42.1 -> 40.6 | No clear effect |
| SkillFactory vs QwQ distillation at 1k rows | Qwen2.5-7B-Instruct · OpenThoughts math/science overall | overall accuracy: 33.5 -> 42.1 | Improved |
Closest papers by meaning
- From f(x) and g(x) to f(g(x)): LLMs Learn New Skills in RL by Composing Old Ones (ICLR 2026 · similarity 0.90)
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models (ICML 2026 · similarity 0.90)
- RLKD: Distilling LLMs’ Reasoning via Reinforcement Learning (AAAI 2026 · similarity 0.89)
- How does RL Post-training Induce Skill Composition? A Case Study on Countdown (NeurIPS 2026 · similarity 0.89)
- Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.89)
- Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · 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.
