Research map / Large language models / Instruction and fine-tuning
STAT: Skill-Targeted Adaptive Training
Yinghui He, Abhishek Panigrahi, Yong Lin, Sanjeev Arora
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 embedding-similarity selection with skill-profile selection (STAT-Sel vs Embed-Sel) (ablation) | Llama-3.2-3B-Instruct SFT · MATH test | accuracy: 46.0 -> 51.5 | Improved |
| Add skill-targeted STAT-Sel data instead of plain MATH-Train SFT | Llama-3.2-3B-Instruct · MATH test | accuracy: 44.8 -> 51.5 | Improved |
| Naive SFT on MATH-Train | Qwen2.5-3B · MATH test | accuracy: 55.8 -> 50.0 | Hurt |
| Restrict SFT to Level 4-5 questions (MATH-Hard) (ablation) | Llama-3.2-3B-Instruct · MATH test | accuracy: 44.8 -> 45.6 | No clear effect |
| Add synthesized skill-targeted questions (STAT-Syn vs STAT-Sel) (ablation) | Llama-3.2-3B-Instruct SFT · MATH^D (difficult subset) | accuracy: 26.6 -> 31.7 | Improved |
| Add synthetic data, STAT-Syn vs STAT-Sel on MATH (ablation) | Llama-3.2-3B-Instruct SFT · MATH test | accuracy: 51.5 -> 50.2 | No clear effect |
| Apply STAT-Sel before GRPO | Llama-3.2-3B-Instruct · Average over 9 math benchmarks | average accuracy: 31.8 -> 40.7 | Improved |
| Apply STAT-Syn before GRPO | Qwen2.5-3B · Average over 9 math benchmarks | average accuracy: 46.0 -> 48.4 | Improved |
| Add GRPO on top of STAT-Syn SFT | Llama-3.2-3B-Instruct · Average over 9 math benchmarks | average accuracy: 37.2 -> 41.3 | Improved |
| Continual STAT with skill profile from MATH-perturb-hard | Llama-3.2-3B-Instruct STAT models · MATH-perturb-hard | accuracy: 3-4% gain | Improved |
| Use embedding-based selection/synthesis instead of skill-based (ablation) | Llama-3.2-3B-Instruct · MATH^2 | accuracy: 21.9 -> 19.5 | Hurt |
Closest papers by meaning
- Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning (ICLR 2025 · similarity 0.87)
- Can Models Learn Skill Composition from Examples? (NeurIPS 2024 · similarity 0.87)
- MASS: Mathematical Data Selection via Skill Graphs for Pretraining Large Language Models (ICML 2025 · similarity 0.86)
- An Emulator for Fine-tuning Large Language Models using Small Language Models (ICLR 2024 · similarity 0.86)
- The Best Instruction-Tuning Data are Those That Fit (NeurIPS 2025 · similarity 0.86)
- Advancing Mathematical Reasoning in Language Models: The Impact of Problem-Solving Data, Data Synthesis Methods, and Training Stages (ICLR 2025 · similarity 0.86)
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.
