Unverified50% confidenceOpinionExact time
BCSM's optimization geometry focuses more on relative ranking than absolute probability, helping the student explore reasonable outputs beyond the teacher's truncation.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedReverse KL divergence tends to make the student focus on a few high-probability modes of the teacher (mode-seeking), and when the teacher has a sharper, more trustworthy distribution because it carries the Reminder, it is more suitable as the correction target for failed rollouts.72% similarUnverifiedToken-probability-based supervision (such as KL divergence) is simple, but the teacher's output probability distribution often only reflects relative preferences near the decision boundary and underuses the rich internal semantic representations.70% similarUnverifiedSDPO's direct-matching self-teacher distribution policy may destroy existing successful behaviors on correct rollouts.66% similarUnverifiedIn structure learning tasks, a carefully designed estimation bias with a controllable direction is sometimes more useful than pursuing unbiased or uniformly small errors.66% similarUnverifiedCompared with traditional Chernoff bounds, the betting method does not depend on pre-known variance or tail distribution parameters, learns adaptively from the data, and the resulting bounds are usually tighter and naturally satisfy anytime validity.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1133926API
curl https://kongchang.com/api/v1/knowledge/claims/1133926MCP
get_claim(id=1133926)