Unverified50% confidenceBenchmarkExact time
In 4,580 controlled integration experiments conducted on the gpt-oss-120b endpoint (stratified sampling from 458 committees), the sample estimate was -0.09 percentage points, and the design-weighted gain was +0.94 percentage points (95% confidence interval [-1.31, +3.50]), while reducing input by 18.72%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/11/2026
First Seen
Valid until: 2/9/2027
Sources
Related Entities
Related Claims
Unverifiedgpt-oss-120b spent about 25 times the cost per correct answer on the distractor counting task, while its accuracy was 7.5 percentage points lower (not significant).78% similarUnverifiedBy improving the sampling process, while keeping the same weights, the nominal maximum layer-evaluation budget required for about 96% accuracy dropped from 60,000 (100 attempts × 50 steps × 12 layers) to 1,200 (5 attempts × 20 steps × 12 layers), roughly a 50-fold reduction.75% similarUnverifiedAblation analysis shows that fine-tuning on the segment paradigm yields a +5.29-point improvement on the test set, while replacing the first 4,500 transcript words with 2,000 retrieved words yields only a +1.55-point improvement.74% similarUnverifiedUnder a group-disjoint source protocol, the dual-view stacking model achieves an average accuracy of 95.29% with a standard deviation of 1.62 percentage points (15 test folds).73% similarUnverifiedIn a controlled experiment with eight paired generation blocks, the iterative program achieved success rates of 46.1% and 31.6% on unseen lengths n=13 and 15, respectively, while independently generated programs achieved only 19.5% and 3.1%.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1158198API
curl https://kongchang.com/api/v1/knowledge/claims/1158198MCP
get_claim(id=1158198)