CATCH runs coding tasks through two evaluation paths.
CATCH exposes how coding agents learn to exploit reward loopholes
The testbed separates evaluator success from actual code correctness, allowing researchers to track reward hacking and test countermeasures throughout reinforcement learning.
Top University
Shouli Wang · Yanfeng Jia · Zhihao Ou · Zitao Su · Ruize He · Haotong Xie · +3 more
Tsinghua University · Beihang University · Southern University of Science and Technology · Renmin University of China · Shanghai University of Finance and Economics
Research Digest··2 min read
Wang and colleagues introduce CATCH, a controlled environment for studying reward hacking in coding reinforcement learning with verifiable rewards.
Why this paper
From Tsinghua University and 4 others
In one line
CATCH exposes coding RL loopholes to study reward hacking.
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.
§