Unverified50% confidenceFactExact time
When the KL ball is centered at the empirical distribution, distributions within the ball can only reweight historical records; when it is centered at a learned density model, the model may assign positive density to unobserved realizations, thereby allowing the adversary to increase the weight of these realizations.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedKL divergence is used to compare the difference in output probability distributions between the fine-tuned model and the original model given the same input; a larger value indicates greater deviation, and a KLD of 0.0244 means the fine-tuning has hardly changed the model's overall behavioral distribution.75% similarUnverifiedThe Glivenko–Cantelli theorem guarantees that as the sample size increases, the empirical distribution function converges uniformly to the true distribution function.70% similarUnverifiedThe robustness of a distribution essentially depends on how its tail behavior is modeled, and replacing the Gaussian distribution with a heavy-tailed distribution can improve a model's stability against outliers.69% similarUnverifiedKL divergence is an asymmetric metric measuring the difference between two probability distributions, and in reinforcement learning it is used to constrain the degree of deviation between the new and old policies (such as PPO's trust region).68% similarUnverifiedA pragmatic strategy for handling distribution shift is to set up an uncertainty-domain expansion mechanism for the decision model: when the model's overall confidence drops, automatically increase the proportion routed to the LLM and trigger retraining67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115656API
curl https://kongchang.com/api/v1/knowledge/claims/1115656MCP
get_claim(id=1115656)