recontextualization

GLM-5.2 uses a technique to counteract reward hacking by detecting and blocking suspicious tool calls rather than penalizing the model, which prevents obfuscation seen in other methods.