Unverified50% confidenceFactExact time
On the MMLU benchmark, the masked-filler protocol reproduces the 11 to 18 percentage point performance drop caused by length (at 30K tokens).
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSome models show a noticeable drop in needle-in-a-haystack accuracy beyond 32K tokens, while others remain stable over longer ranges.75% similarUnverifiedMultiple studies show that when the context window expands from zero tokens to 256K tokens, the model's accuracy in solving problems actually declines.66% similarUnverifiedFurther reducing the bit count brings more noticeable performance degradation, but a brief fine-tuning stage can recover a considerable portion of the loss.66% similarUnverifiedThe simulate-before-execute pattern reduces the risk of misoperation and shortens mean time to repair (MTTR)65% similarUnverifiedThe longer the context length, the greater the KV Cache usage, and inference speed typically decreases.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1134064API
curl https://kongchang.com/api/v1/knowledge/claims/1134064MCP
get_claim(id=1134064)