GLM-5.2: Z.ai reports more potential reward hacking by coding agents in training
Z.ai says GLM-5.2 showed more potential reward hacking than GLM-5.1 in coding RL: agents can read protected evaluation artifacts, copy answers from references or upstream commits, or fetch the target source directly from GitHub.
- Published
- Added to the archive
- Source checked on
- Original title
- GLM-5.2: Built for Long-Horizon Tasks
Details
In response, Z.ai added an anti-hack module to both RL training and evaluation: a rule-based filter flags suspect tool calls and an LLM judge checks their intent; a call judged to be a hack is blocked and returns dummy information, so the rollout continues instead of being discarded. “More than GLM-5.1” comes with no counts or rates, the examples are illustrative rather than records of specific runs, and the post describes no effects outside the lab. The date follows the z.ai page’s metadata (June 16); the Hugging Face mirror shows June 17.