Unverified50% confidenceFactExact time
In standard ES, subtracting the mean reward from the rewards (mean subtraction) reduces variance; from the centroid perspective, it switches the displacement reference point from the model's current position to the sampled population centroid, with direction and relative magnitude unchanged.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFor any finite population, the mean-subtracted ES update is exactly equal to the centroid displacement multiplied by a scalar factor that depends on the mean reward.78% similarUnverifiedThe endpoint metric used in the evaluation is the difference of 'model minus uniform control group', where a positive value represents performing worse than the control.71% similarUnverifiedThe validation loss on preference data measures how well the model fits the preferred responses, whereas downstream benchmarks examine generation quality; the two are not equivalent, and the model may reduce validation loss while degrading on downstream tasks.70% similarUnverifiedModule residual anchoring lowered the hit rate under natural questioning from 28.7% to 22.0% while also reducing the measured behavioral drift, a trade-off of exchanging recall for preservation of non-target behavior.68% similarUnverifiedGradient descent computes the partial derivatives of the loss function with respect to model parameters and updates the parameters in the direction opposite to the gradient to approach the minimum of the loss function.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1103113API
curl https://kongchang.com/api/v1/knowledge/claims/1103113MCP
get_claim(id=1103113)