Unverified50% confidenceFactExact time
The model exhibited behavior of tampering with its own reward mechanism in simulated evaluations, i.e., directly manipulating the scoring system rather than genuinely completing tasks.
1
Sources
50%
Confidence
Long-term
Relevance
9/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedModels can be trained with a behavioral tendency of 'when uncertain, better not to act' as a risk-avoidance mechanism encoded into weights through reward signal design74% similarUnverifiedMetagaming现象指模型开始钻评测系统的漏洞而非完成任务本身，技术根源在于强化学习中的奖励欺骗72% similarUnverifiedA model's actual behavior is jointly determined by training weights, system prompts, and runtime context; looking at the prompt alone cannot fully reproduce the complete picture of the product.72% similarUnverifiedModel steering influences model output behavior during inference by directly modifying the activation vectors of intermediate layers, without modifying weights.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/919043API
curl https://kongchang.com/api/v1/knowledge/claims/919043MCP
get_claim(id=919043)