Unverified50% confidenceFactExact time
Mechanism one: when labeled questions are relatively scarce, the Q^{-1/2} term dominates the excess risk, generic buffered ERM is optimal, and the marginal benefit of increasing the number of generations N is limited.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the paper, the amplification factor of model error for Q-value iteration under the risk-neutral setting is approximately 1/(1-γ), whereas under the recursive entropic risk setting, error amplification is more complex due to nonlinear exponential operations, resulting in an exponential dependence.72% similarUnverifiedMechanism three: when the number of questions is sufficiently large, the minimax excess risk reaches the identification lower bound of N^{-β}, which cannot be broken even with infinitely many questions, and this lower bound is independent of the number of questions.72% similarUnverifiedIn the ensemble setting, when the margin condition exponent β>1, the N^{-β/2} term may decay faster than N^{-1/2}, so the number of generations can bring more significant gains.70% similarUnverifiedThe lower bound for the submodel also applies to general ensembles, matching the upper bound of buffered ERM (ignoring logarithmic factors) in the question-scarce regime and part of the intermediate regime.70% similarUnverifiedThere is an optimal sweet-spot range for the heavy-tail degree of the weight spectrum, and deviating from the optimal value corresponds to degraded model quality or increased risk of overfitting.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1046165API
curl https://kongchang.com/api/v1/knowledge/claims/1046165MCP
get_claim(id=1046165)