Unverified50% confidenceFactExact time
The standard @1 objective (such as the common single-sample optimization in RLHF/PPO) only cares about the expected reward of a single sample, and tends to concentrate probability mass on a few high-confidence solutions, thereby compressing the exploration space.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn zero-shot classification, the scores of all candidates sum to 1; the fewer the candidates, the larger the probability space available to each, so the threshold parameter needs to be calibrated individually based on the specific number of categories.72% similarUnverifiedThe leverage compatibility assumption requires the sampling distribution to be aligned with the important directions of the target function in feature space, so as to avoid extreme sampling bias causing approximation error to explode.70% similarUnverifiedProbabilistic Focal Search (PFS) follows the heuristic-guided selection of FS with probability p, and with probability 1-p expands the node with the minimum f value in the OPEN list.70% similarUnverifiedThe more independent requirements included in a single prompt, the more the probability of the model omitting or incorrectly implementing certain requirements grows non-linearly; the MVP strategy improves reliability by narrowing the model's decision space.70% similarUnverifiedRLHF carries the risk of reward over-optimization, where models may learn to satisfy the surface preferences of human annotators rather than genuinely improving accuracy or safety.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130868API
curl https://kongchang.com/api/v1/knowledge/claims/1130868MCP
get_claim(id=1130868)