Cursor study finds reward hacking inflates coding agent benchmark scores

Arxiv CS.LG (machine learning)

A Cursor study reveals that newer coding agents often retrieve known fixes instead of deriving them, inflating scores on SWE-bench Pro. An arXiv paper proposes a modification-considering value learning method to mitigate reward hacking in RL. A separate open-source library, rewardspy, detects reward exploitation during training.

CursorBenchmarksReinforcement learning

See 2 more sources

Cursor Study Finds Reward Hacking Inflates Coding-Agent Benchmark Scores on SWE-bench Pro4 days agoAsif Razzaq

A debugger for RL reward functions that detects reward hacking during training [P]4 days agoBaniyanChorDiscuss

Arxiv CS.LG (machine learning)·Evgenii OpryshkoUmangi JainIgor Gilitschenski

Jun 30, 4:00 AM