Unverified50% confidenceFactExact time
This test bypassed the retrieval step and fed all candidate documents directly to the models, thus revealing the reasoning capability bottleneck of language models themselves rather than defects in retrieval algorithms.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/12/2026
First Seen
Valid until: 1/10/2027
Sources
Related Claims
UnverifiedWhen a model fails to correctly invoke retrieval tools or determines that a question does not require a search, it falls back to answering from training data, thereby exposing the knowledge cutoff issue.74% similarUnverifiedThe core idea of hard negative mining is to introduce samples near the classification boundary to drive refinement of the decision boundary, enabling the model to learn deep semantic features rather than surface-level keyword patterns.73% similarUnverifiedEven when the model is explicitly required to answer only based on retrieved documents, it still fills in gaps using parametric memory and generates seemingly well-formatted citations.71% similarUnverifiedThe ranking factors of traditional search can be partially reverse-engineered and restored through statistical analysis, but the decision-making process of large language models with billions of parameters is nearly impossible to audit at the same level of granularity.71% similarUnverifiedApplications should be tested with real data rather than idealized samples, because placeholder content masks layout issues, logic breaks, and edge cases that only appear during actual use.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1003225API
curl https://kongchang.com/api/v1/knowledge/claims/1003225MCP
get_claim(id=1003225)