Unverified50% confidenceOpinionExact time
Once the reward model can read the chain of thought, the model will learn to write reasoning text that appears to make the reward model give a high score, rather than actually solving the problem, which is highly consistent with classic reward hacking.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
The Thinking Leakage Problem in LLM RL Training
hackernewshackernews10/16/2026
Related Entities
Related Claims
UnverifiedIf rewards target only final outcomes, the model tends to take shortcuts and memorize solutions for specific scenarios.76% similarUnverifiedSpeculative Reward Hacking is viewed as a manifestation of Goodhart's Law at inference time—the model actively infers and overfits to an imagined reward signal during inference.75% similarUnverifiedThe Process Reward Model (PRM) applies the reward signal step by step to each reasoning link, which can alleviate the reward hacking risk caused by traditional RLHF giving rewards only at the final answer.72% similarUnverified推理模型在生成最终答案前会先产生大量内部思维链（Chain-of-Thought）token，这些 token 不直接暴露给用户72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1044677API
curl https://kongchang.com/api/v1/knowledge/claims/1044677MCP
get_claim(id=1044677)