Research map / Large language models / Alignment and preferences
SWE-IF: Aligning Code Evaluation with Human Preference
Ming Zhong, Xiang Zhou, Ting-Yun Chang, Qingze Wang, Nan Xu, Xiance Si, Dan Garrette, Shyam Upadhyay, Jeremiah Zhe Liu, Jiawei Han, Benoit Schillings, Jiao Sun
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 more non-functional instructions (1 to 5), single-turn | all evaluated LLMs · Big-SWE-IF | functional regression rate (%): 2.48 -> 5.76 | Hurt |
| add more non-functional instructions (1 to 5), multi-turn | all evaluated LLMs · Big-SWE-IF | functional regression rate (%): 3.18 -> 9.31 | Hurt |
| use multi-turn editing instead of single-turn generation | all evaluated LLMs · Big-SWE-IF / Live-SWE-IF | functional regression: 5.76 -> 9.31 (5 instructions, Big-SWE-IF) | Hurt |
| use multi-turn editing instead of single-turn generation | all evaluated LLMs · Big-SWE-IF | task-level IF score: 3% to 4.5% advantage | Improved |
| use multi-turn editing instead of single-turn generation | all evaluated LLMs · Live-SWE-IF | task-level IF score: around 8% advantage | Improved |
| add more instructions (1 to 5) | Claude 4 Opus, single-turn · Big-SWE-IF | task-level IF score: 88.77 -> 46.75 | Hurt |
| add more instructions (1 to 5) | Kimi K2, single-turn · Live-SWE-IF | task-level IF score: 62.75 -> 11.94 | Hurt |
| add more instructions (1 to 5) | o4 mini, multi-turn · Live-SWE-IF | functional regression rate (%): 3.63 -> 15.92 | Hurt |
| add more instructions (1 to 5) | Claude 4 Opus, single-turn · Big-SWE-IF | functional regression rate (%): -0.86 -> -2.08 | No clear effect |
| mix IF with functional score (alpha=0.4) rather than Func alone | LMArena Elo correlation · Big-SWE-IF | Pearson correlation with LMArena Elo: alpha=0.4 optimum | Improved |
| mix IF with functional score (alpha=0.7) rather than either alone | LMArena Elo correlation · Big-SWE-IF | Spearman correlation with LMArena Elo: alpha=0.7 optimum | Improved |
| use pure IF instead of pure Func | LMArena Elo correlation · Big-SWE-IF | Spearman correlation with LMArena Elo: over 0.1 points higher | Improved |
Closest papers by meaning
- SelfCodeAlign: Self-Alignment for Code Generation (NeurIPS 2024 · similarity 0.89)
- Copilot Arena: A Platform for Code LLM Evaluation in the Wild (ICML 2025 · similarity 0.88)
- Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Models (ICLR 2025 · similarity 0.88)
- DINGO: Towards Diverse and Fine-Grained Instruction-Following Evaluation (AAAI 2024 · similarity 0.88)
- Training Language Models to Generate Quality Code with Program Analysis Feedback (NeurIPS 2025 · similarity 0.88)
- VERSE: Verification-based Self-Play for Code Instructions (AAAI 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.
