Unverified50% confidenceFactExact time
On a set of ordered-scale questions, using probability scores as the REINFORCE reward caused accuracy to drop sharply from 0.636 to 0.111.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/13/2026
First Seen
Valid until: 2/11/2027
Sources
Related Entities
Related Claims
UnverifiedWhen the single-run error probability p=0.10, the error rate of strict majority voting is 0.02800 for N=3, 0.00856 for N=5, and 0.000147 for N=10.77% similarUnverifiedFor an invoice classification model with 97% accuracy, when a reviewer examines 300 invoices per day, only about 9 will be wrong, and errors are buried among correct items, causing the reviewer's rejection rate to approach zero.76% 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.75% similarUnverifiedIn the synthetic experiment, "always output RISK_LEVEL: 1" achieves an ordinary accuracy as high as 0.922, but an average reward of −0.000, showing that class-balanced rewards can prevent laziness.75% similarUnverifiedWhen there are only two options, the signal from using probability scores as the REINFORCE reward is identically zero.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1164995API
curl https://kongchang.com/api/v1/knowledge/claims/1164995MCP
get_claim(id=1164995)