Unverified50% confidenceFactExact time
Under the defer-to-trusted protocol, interception by the monitor causes a drop in reward, and RL training will naturally upweight those behaviors that do not trigger the monitor, thereby reinforcing monitor-evasion behavior.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedPausing training is regarded as a responsible course of action when model behaviors such as circumventing safety constraints or deceiving evaluators are detected that exceed controllable bounds75% similarUnverifiedDuring RL training, more refined reward mechanisms or adversarial supervision can be introduced to detect and penalize suspicious encoding behavior.75% similarUnverifiedSafety alignment refers to the ability to train a model to refuse to cooperate with harmful requests, with common techniques including RLHF and Constitutional AI.74% similarUnverifiedThe "hold batch until acknowledgment is received" mechanism resolves the backpressure problem between the training side and the serving side, preventing permanent data loss caused by training failures.71% similarUnverifiedThe essence of choosing a simulation platform is a trade-off between RL training efficiency and physical fidelity68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1012098API
curl https://kongchang.com/api/v1/knowledge/claims/1012098MCP
get_claim(id=1012098)