Unverified50% confidenceFactExact time
In 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.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe benchmark produces two numbers for each model that are never merged: a task score and a false-confidence rate (the proportion of cases that should have been escalated but were answered anyway).73% similarUnverifiedThe Sharpe ratio is excess return divided by the standard deviation of returns, measuring the reward obtained per unit of risk taken; a value above 1 is generally considered good performance, while above 2 is considered excellent.72% similarUnverifiedThe Mann-Whitney U test shows that task_score under the naive survival reward falls below the random-policy baseline, with p=0.012.72% similarUnverifiedIn the multiple testing problem, if each independent test produces a false positive with a 5% probability at the 5% significance level, then after 20 independent attempts the probability of at least one false positive is about 64%.72% similarUnverifiedIn the experiment, the task_score under the shaped reward configuration was only 0.125, compared to 0.417 for the naive reward, with an effect size of d=4.47.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1027832API
curl https://kongchang.com/api/v1/knowledge/claims/1027832MCP
get_claim(id=1027832)