SELD Detects Silent Evidence Loss in Context Compression
November 11, 2026
The SELD framework identifies cases where context compression removes critical evidence while leaving the model's final answer unchanged. It detects these risks by measuring shifts in a reader model's preferences among unsaid answers.
HOW THIS AFFECTS YOU
●
builderYou can use this to ensure your long-context compression doesn't make your model's answers unreliable.
●
researcherThis method reveals failures in context compression that standard answer-based evaluations miss.