The authors audited GRPO reinforcement-learning runs to track how unit-test rewards affect pass rates, patch sizes and behavior on already-correct programs.
Rewarding necessary edits makes debugging patches smaller without reducing correctness
PreciseCoder supplements unit-test rewards with semantically verified edit precision and recall, discouraging unnecessary changes while preserving pass rates.
Big Tech
Yunkai Zhan · Miaosen Chai · Shangshang Wang · Jike Zhong · Yuqing Yang · Deqing Fu · +2 more
University of Chicago · University of Southern California · Google · University of California, Berkeley
Research Digest··3 min read
Zhan and colleagues show that reinforcement learning based only on unit-test outcomes can improve code correctness while encouraging larger patches and invented bugs.
Why this paper
From Google and 3 others
In one line
PreciseCoder reduces average edit lines to 3 per bug while keeping pass rates on par with standard training.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ✓Compute or model size stated (params 4B, 9B, and 27B)
- ✓Limitations stated by the authors (2 noted)
- ✓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.
§