Unverified50% confidenceOpinionExact time
Current mainstream alignment methods include reinforcement learning from human feedback (RLHF), interpretability research and adversarial testing, but none has fundamentally solved the alignment problem at the superintelligence level.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/1/2026
First Seen
Valid until: 1/30/2027
Sources
Related Entities
Related Claims
UnverifiedRLHF 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.77% 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73% similarUnverifiedIn modern models that have undergone alignment training (RLHF, etc.), the effect of politeness on answer quality has been greatly reduced, as models are optimized to provide high-quality responses to reasonable requests regardless of phrasing.71% similarUnverifiedThe essence of RLHF optimization is to produce outputs that appear safe to human evaluators, rather than the model genuinely understanding why a given request is harmful.71% similarUnverifiedUnder 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.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1123752API
curl https://kongchang.com/api/v1/knowledge/claims/1123752MCP
get_claim(id=1123752)