Unverified50% confidenceFactExact time
Under the retrieval condition, the GPT models often skip the report, answer based on their own higher-precision output files, and incorrectly round to an adjacent option, a failure mode the researchers call "evidence substitution".
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe root cause of poor knowledge base performance typically lies in unreasonable document chunking strategies, inappropriate vector model selection, or insufficient retrieval recall quantity, rather than issues with the model itself.70% similarUnverifiedIn the answer extraction process, parsing the answer required by the tree nodes from the model output itself involves model-related errors, further lowering end-to-end accuracy.70% similarUnverifiedIf measurement error is not properly addressed, it leads to systematic bias in the estimation of regression coefficients69% similarUnverifiedThe traces reveal failure modes that can still occur even with prompt constraints: source attribution errors, incomplete evidence coverage and unsafe assumptions.69% similarUnverifiedA common mistake in many RAG applications is treating "refusal to answer" as an exceptional case, causing the model to still tend to give an answer when evidence is insufficient.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1073395API
curl https://kongchang.com/api/v1/knowledge/claims/1073395MCP
get_claim(id=1073395)