Unverified50% confidenceFactExact time
In tw-exam-bench, the answer-key consistency rate for 31,312 groups of duplicated questions reached 99.91%, and the only 27 discrepancies were verified and attributed one by one.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
Unverified@servasyy_ai manually annotated 300 judgments and found that all 255 answers with confidence above 90% were correct, while the middle zone had too few samples to draw a conclusion.73% similarUnverifiedOn Track B, the flat RAG baseline can detect 76%–97% of true contradictions, but produces false positives on 16%–43% of surface-similar safe drafts (the specific ratio depends on the backend).69% similarUnverifiedModels that only read the call transcripts passed 29 of the 31 contradictory cases.69% similarUnverifiedOnly 38.5% of GPT-OSS 120b's capitulations adopted the decoy answer, with the remainder retreating to vagueness, providing a safer failure mode.68% similarUnverified相比纯 CoT 推理，ReAct 在需要信息检索的问答任务上错误率降低约 34%67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1038581API
curl https://kongchang.com/api/v1/knowledge/claims/1038581MCP
get_claim(id=1038581)