Topic
Failure modes in reinforcement learning with verifiable rewards where a model improves measured reward while degrading transfer, robustness, or true task success.
RLVR can train models to satisfy the verifier rather than the user objective. This topic tracks the gap between benchmark acceptance and deployed truth.
Verification EconomicsAgent InfrastructureTool-Agent Reward Hacking
A research paper on verifier failure under RLVR and tool agents. It defines the exploit tax: the cost paid when verifier-guided reasoning increases local acceptance faster than true transferred success.
Results are local, static, and index the same public surfaces exposed to crawlers.