Research map / Large language models / Instruction and fine-tuning
OptimSyn: Influence-Guided Rubrics Optimization for Synthetic Data Generation
Zhiting Fan, Ruizhe Chen, Tianxiang Hu, Ru Peng, Zenan Huang, Haokai Xu, Yixin Chen, Jian Wu, Junbo Zhao, Zuozhu Liu
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 |
|---|---|---|---|
| Add influence-guided RL rubric optimization (OptimSyn) vs no SFT base | Qwen3-8B-Base · MMLU-pro (History) | accuracy: 22.83 -> 56.96 | Improved |
| Add influence-guided RL rubric optimization (OptimSyn) vs no SFT base | Qwen3-8B-Base · MedQA | accuracy: 51.45 -> 58.75 | Improved |
| Add influence-guided RL rubric optimization (OptimSyn) vs no SFT base | Qwen3-8B-Base · SuperGPQA (Health) | accuracy: 28.06 -> 38.28 | Improved |
| Add OptimSyn data vs Qwen3-8B-Instruct | Qwen3-8B-Base · BBH | accuracy: 85.21 -> 75.65 | Hurt |
| Add OptimSyn data vs Qwen3-8B-Instruct | Qwen3-8B-Base · MMLU-pro (Health) | accuracy: 60.39 -> 56.97 | Hurt |
| Add OptimSyn data vs Qwen3-8B-Instruct | Qwen3-8B-Base · SuperGPQA (Health) | accuracy: 37.16 -> 38.28 | Improved |
| Use OptimSyn instead of best open SFT corpus (Wildchat) | Qwen3-8B-Base · MMLU-pro (History) | accuracy: 52.76 -> 56.96 | Improved |
| Use OptimSyn instead of Medical-R1-Distill | Qwen3-8B-Base · HealthBench | accuracy: 80.38 -> 74.82 | Hurt |
| Increase rubric rollouts per seed G from 5 to 10 to 15 (ablation) | OptimSyn · downstream benchmarks | accuracy | Improved |
| Select higher-influence-score training subsets (ablation) | Qwen3-8B-Base SFT · test accuracy (2K/4K subsets) | accuracy: R^2=0.57 for 2K and R^2=0.54 for 4K | Improved |
Closest papers by meaning
- It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs (NeurIPS 2026 · similarity 0.86)
- Towards a Theoretical Understanding of Synthetic Data in LLM Post-Training: A Reverse-Bottleneck Perspective (ICLR 2025 · similarity 0.86)
- Deep Research as Rubric (NeurIPS 2026 · similarity 0.86)
- OptMATH: A Scalable Bidirectional Data Synthesis Framework for Optimization Modeling (ICML 2025 · similarity 0.86)
- DataGen: Unified Synthetic Dataset Generation via Large Language Models (ICLR 2025 · similarity 0.85)
- How to Synthesize Text Data without Model Collapse? (ICML 2025 · similarity 0.85)
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.
