Unverified50% confidenceFactExact time
The experiment applied a bias of -2 to approximately 50 tokens via llama.cpp's --logit-bias parameter.
1
Sources
50%
Confidence
Long-term
Relevance
9/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn a RAG pipeline, a user query that was originally 50 tokens may expand into a complete prompt containing 5,000–10,000 tokens.66% similarUnverified高推理强度模式下单次请求的 Token 消耗可能是低强度模式的 5-20 倍63% similarUnverifiedIn traditional LLM inference, the approach of pre-allocating contiguous GPU memory for the KV cache of each request often results in actual GPU memory utilization of less than 50%.60% similarUnverifiedIn MoE models, dense layer parameters typically account for less than 20% of total parameters, but every token must pass through them, so they require higher-precision protection.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/954323API
curl https://kongchang.com/api/v1/knowledge/claims/954323MCP
get_claim(id=954323)