Unverified50% confidenceFactExact time
When there are only two options, the signal from using probability scores as the REINFORCE reward is identically zero.
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn a set of ordered-scale questions, using probability scores as the REINFORCE reward caused accuracy to drop sharply from 0.636 to 0.111.75% similarUnverifiedThe model's output tokens are zero, because it only returns probabilities rather than generating text.74% similarUnverifiedIn a four-choice scenario, even if the top-1 probability reaches 0.64, the normalized confidence may be only 0.26, which is why the plugin, comparing confidence against the 0.8 threshold, can never pass.73% similarUnverifiedIf a difficult problem fails across all samples (pass@k ≈ 0), the reward variance is zero, the gradient signal nearly vanishes, and the parameter update approaches zero.73% similarUnverifiedAn honest 0.8 probability is more useful than an inflated 0.99.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1164994API
curl https://kongchang.com/api/v1/knowledge/claims/1164994MCP
get_claim(id=1164994)