Research map / Large language models / Agents, code and math
RECAST: Expanding the Boundaries of LLMs' Complex Instruction Following with Multi-Constraint Data
Zhengkang Guo, Wenhao Liu, Mingchen Xie, Jingwen Xu, Zisu Huang, Muzhao Tian, Jianhan Xu, Yuanzhe Shen, Qi Qian, Muling Wu, Xiaohua Wang, Heda Wang and 4 more
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 |
|---|---|---|---|
| fine-tune on RECAST-30K instead of Tülu 3 Persona IF (SFT) | Qwen2.5-7B · RECAST-Test average | OSR (%): 26.83 -> 31.25 | Improved |
| fine-tune on RECAST-30K instead of Tülu 3 Persona IF (SFT) | Llama-3.1-8B · RECAST-Test average | OSR (%): 26.96 -> 30.54 | Improved |
| add RLVC reinforcement learning after SFT (ablation) | Llama-3.1-8B · RECAST-Test average | OSR (%): 30.54 -> 31.96 | Improved |
| add RLVC reinforcement learning after SFT (ablation) | Qwen2.5-7B · RECAST-Test average | OSR (%): 31.25 -> 32.33 | Improved |
| add RLVC after SFT (ablation) | Llama-3.1-8B and Qwen2.5-7B · RECAST-Test Level 1 OSR | OSR (%): Llama 18.50 -> 17.00; Qwen 14.50 -> 16.50 | Mixed |
| add RLVC after SFT (ablation) | Qwen2.5-7B · IFEval and FollowBench | accuracy / HSR: FollowBench 59.60 -> 63.23 | Improved |
| add RLVC after SFT (ablation) | Llama-3.1-8B · FollowBench | HSR: 57.10 -> 61.76 | Improved |
| fine-tune on RECAST-30K instead of Tülu 3 Persona IF | Qwen2.5-7B · IFEval | Pr.(L) accuracy: 73.94 -> 73.38 | No clear effect |
| SFT on RECAST-30K instead of inference-time DVR on Instruct model | Llama-3.1-8B · RECAST-Test RSR average | RSR (%): 7.0 -> 12.5 | Improved |
| SFT on RECAST-30K instead of inference-time DVR on Instruct model | Qwen2.5-7B · RECAST-Test RSR average | RSR (%): 6.5 -> 13.0 | Improved |
| fine-tune on RECAST-30K instead of Tülu 3 Persona IF | Qwen2.5-7B · GPQA+MuSR average | accuracy (%): 37.60 -> 34.97 | Hurt |
| add RLVC after SFT (ablation) | Qwen2.5-7B · GPQA+MuSR average | accuracy (%): 34.97 -> 38.39 | Improved |
Closest papers by meaning
- Generalizing Verifiable Instruction Following (NeurIPS 2025 · similarity 0.90)
- Can Large Language Models Understand Real-World Complex Instructions? (AAAI 2024 · similarity 0.90)
- Benchmarking Complex Instruction-Following with Multiple Constraints Composition (NeurIPS 2024 · similarity 0.90)
- Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Models (ICLR 2025 · similarity 0.90)
- Incentivizing Reasoning for Advanced Instruction-Following of Large Language Models (NeurIPS 2025 · similarity 0.89)
- Scaling Prompt Synthesis for Large Language Model Reasoning (ICML 2026 · similarity 0.89)
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.
