Unverified50% confidenceFactExact time
The only helpful placement of orthogonal constraints is to use symmetric positive-definite factors on the query and key projections, but its tiny gain vanishes as the model grows.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRetrieval augmentation and tool use can only mitigate factual errors and cannot solve the problem of judging importance; even if the model retrieves relevant reports, it may still mistakenly analogize an individual case as a systemic event.70% similarUnverifiedThe limitation of frozen-feature probes is that they can only show that information is present and readable, and cannot prove that the model actually uses this information in downstream reasoning.70% similarUnverifiedTool-calling judge models have a systematic flaw: when no source for a value can be found in the input, the judge is more inclined to reject it even if the value is completely correct, a phenomenon named the "derivation gap".69% similarUnverifiedThe limitations of vector retrieval include that retrieval quality depends on the semantic understanding capability of the embedding model, and it cannot guarantee the recall of all key information, posing a risk of memory omission.68% similarVerifiedThe self-attention mechanism uses dot-product operations among three matrices—Query, Key, and Value—to overcome the parallelization bottleneck caused by the sequential computation of RNNs/LSTMs.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1129389API
curl https://kongchang.com/api/v1/knowledge/claims/1129389MCP
get_claim(id=1129389)