Unverified50% confidenceBenchmarkExact time
In the change test after removing the records, Jev's TVD was 49%, Gemini's 43%, luna's 30%, and Haiku's 27%; the noise floors for repeated questions were 3%, 12%, 9%, and 9% respectively.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedUnder the held-out evaluator, ASC's median error rate drops by 38%, the median improvement in supported sentences reaches 46%, and this holds in all test settings.61% similarUnverifiedWhen faced with 50 nearest-neighbor tools with a single correct tool present, Laya's accuracy drops sharply to 31%, while Jev maintains 98%.60% similarUnverifiedCounterfactual mDPO improved METEOR from 58.10% to 85.30%, reduced Self-BLEU-2 from 58.79% to 52.88%, and improved Distinct-3 from 6.68% to 7.81%.59% similarUnverifiedCode replay inherits errors from the extraction stage; in the cluster sufficiency test, gpt-oss-120b's pass rate actually dropped from 56% to 40%.59% 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.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1047727API
curl https://kongchang.com/api/v1/knowledge/claims/1047727MCP
get_claim(id=1047727)