The authors trained Qwen3-30B-A3B-Instruct-2507 on R2E-Gym and evaluated it on SWE-bench Verified and the Python subset of SWE-bench Pro.
Adaptive test weighting gives software agents denser, more useful rewards
CodeLift groups tests and patches by observed behavior, then emphasizes unresolved functionality and credits productive intermediate edits.
Chinese Tech
Junjielong Xu · Huachao Zhu · Alyssa Wu · Zhichao Hu · Pinjia He
The Chinese University of Hong Kong, Shenzhen · Tencent
Research Digest··2 min read
Xu and colleagues show that raw test pass-rate is a poorly calibrated reinforcement-learning reward for software engineering agents because test suites overrepresent some functions, difficulties and regression checks.
Why this paper
From Tencent and The Chinese University of Hong Kong, Shenzhen
In one line
CodeLift dynamically calibrates test pass-rate rewards to create an adaptive curriculum for training SWE agents.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§