Unverified50% confidenceOpinionExact time
Under continual learning, the policy may learn a large amount of evasion knowledge through online RL, with a perturbation magnitude far exceeding what control evaluation assumes, causing protocols that originally passed evaluation to lose their meaning.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe 'repetition' phenomenon is a typical failure mode for large models in agent scenarios, usually stemming from reward design flaws during the reinforcement learning phase or a lack of effective termination judgment mechanisms.71% similarUnverifiedOver-reliance on instant prompts may weaken the training effect of on-the-spot expression and knowledge internalization.71% similarUnverifiedRLHF中人类评审倾向给流畅自信全面的答案更高分而较少关注事实核查，导致模型获得隐性激励表现得有把握比实际有把握更重要70% similarUnverifiedMost traditional post-hoc explanation methods are local attributions targeting individual predictions, making it difficult to distill them into reusable knowledge assets.69% similarUnverifiedRLHF has limitations: human raters may be flattered by models and may struggle to evaluate outputs beyond their own comprehension, which is the problem that scalable oversight research attempts to solve.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1012105API
curl https://kongchang.com/api/v1/knowledge/claims/1012105MCP
get_claim(id=1012105)