Unverified50% confidenceBenchmarkExact time
On the selected models and languages, ReciLex augmentation lowered TER from 77.14 to 75.12.
1
Sources
50%
Confidence
Long-term
Relevance
11/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn orchestration mode the local model scored 78/78 and 77/78; in single large-requirement mode it only got 76, 74, 71 (out of 78)58% similarUnverifiedResQAdam reduces optimizer state memory by 71% to 74% relative to AdamW.57% similarUnverifiedPerforming a rank-1 edit on a single feedforward layer halves the loop rate from 19.4% to 9.4% under T=0.7 sampling.56% similarUnverifiedRunning as BLM, the open-source model Qwen3.8-27B raised its task solve rate on the τ²-bench telecom benchmark from 29% to 96%, with the number of model calls reduced by 93%.56% similarUnverifiedOn the Tajik legal retrieval task, the unified model achieved a recall@10 of 92.2, essentially on par with the starting model's 92.4, but recall@1 dropped by 4.4 percentage points.55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1182875API
curl https://kongchang.com/api/v1/knowledge/claims/1182875MCP
get_claim(id=1182875)