Unverified50% confidenceFactExact time
Prototype compression is lossy and discards details of the local distribution such as variance, skewness, and multimodality.
1
Sources
50%
Confidence
Long-term
Relevance
11/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGradient weighting on high-similarity negatives persistently harms downstream generalization, i.e., excessively pushing the model away from certain samples disrupts the overall structure of the embedding space, so retrieval performance drops rather than rises (termed 'gradient distortion').74% similarUnverifiedSummary compression is lossy; when an LLM generates a summary it selectively retains information, and technical details, specific numerical values, and edge cases have a relatively high probability of being discarded.72% similarUnverifiedOn-device deployment is constrained by compute power, storage and model size, and typically requires compressing the model, which may further reduce accuracy.72% similarUnverifiedThe cost of model compression is degraded capability in complex reasoning, long-context understanding and multilingual generalization.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1153878API
curl https://kongchang.com/api/v1/knowledge/claims/1153878MCP
get_claim(id=1153878)