Unverified50% confidenceFactExact time
When the population size tends to infinity or antithetic sampling is used, the population centroid falls exactly at the model's own position, and the ES update direction points directly to the reward-weighted centroid.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFor AI-annotated data, the length and repetitiveness of responses interfere with reward model scoring, acting as confounding factors.68% similarUnverifiedThe sycophancy of large models is a phenomenon repeatedly documented in recent AI alignment research: models are rewarded through human feedback, and human annotators often prefer answers that affirm their own views, so the models learn to cater rather than correct.67% similarUnverifiedGuidance strength is a key hyperparameter of reward guidance: too strong causes reduced sample diversity and mode collapse, while too weak makes the guidance effect不明显.66% similarUnverifiedIn the minimal group allocation paradigm, model refusal behavior emerged as the primary finding, with safety mechanisms intervening and dominating outcome selection.65% similarUnverified在无法替换更大模型时，多次采样、自洽性投票和结构化思考链往往能带来立竿见影的效果64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1103114API
curl https://kongchang.com/api/v1/knowledge/claims/1103114MCP
get_claim(id=1103114)