The authors first conduct a diagnostic study on SWE-bench Verified to identify critical behaviors for repair success.
Training coding agents to locate, vary, and verify boosts repair reliability.
Weighted supervised fine-tuning and mutation-based reinforcement learning improve pass@8 from 46.7% to 60.7% on SWE-bench Verified.
Big Tech
Muhammad Ahmed Mohsin · Myeongsoo Kim · Kangrui Ruan · Shweta Garg · Varun Kumar · Murali Krishna Ramanathan
Stanford University · AWS AI Labs
Research Digest··3 min read
The authors diagnose three teachable behaviors—location diversity, edit diversity, and verification—that determine the success of autonomous coding agents.
Why this paper
From AWS AI Labs and Stanford University
In one line
Autonomous coding agents can learn fault location, edit variation, and patch verification, lifting held-out SWE-bench pass@1 from 31.9% to 43.0%.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks (3 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§