Unverified50% confidenceFactExact time
On the MedAbstain benchmark, the median appropriate abstention rate of the base models was only 8.3%, and after fine-tuning it further dropped to a median of 4.2%.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
11/9/2026
First Seen
Valid until: 11/23/2026
Sources
Related Entities
Related Claims
UnverifiedCompared to the average of the two graphs, the conditioned response graph reduces mean squared prediction error by 53.6% on Gemma-2-9B.75% similarUnverifiedAt the 0.81 threshold, the proportion of cached answers that "fit the question" is 5.8 to 10.7 percentage points lower than fresh answers; converted to all responses, the overall quality loss is about 2 to 3 percentage points; Haiku 5.5 has the largest gap (10.7 percentage points), with a cached-answer pass rate of only 67.9%.74% similarUnverifiedFlips are concentrated in the regions where the model's own confidence is low: the lowest confidence tertile has a flip rate of 54%, while the highest tertile has only 8%.74% similarUnverifiedIn the placebo test, 90% of the synthetic control method estimates narrowed to between -7% and +9%, with RMSE dropping to 4.9 percentage points, more than one-third lower than DiD.72% similarUnverifiedComparing model outputs with reference answers using the difflib ratio, among the 44 passes the highest ratio was only 0.73, with most below 0.3, indicating that the model is not reciting answers.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1149934API
curl https://kongchang.com/api/v1/knowledge/claims/1149934MCP
get_claim(id=1149934)