Periscope Extends Frozen LM Context via Grid-Based Factorization
October 1, 2026
Periscope is a training-free inference method that extends frozen LLMs beyond their context window by arranging text chunks on a grid. It evaluates local and strided spans to create an evidence map, allowing for efficient document retrieval and reasoning.
HOW THIS AFFECTS YOU
●
builderYou can implement long-context reasoning on frozen models without retraining or expensive KV caching.
●
researcherThis provides a new way to factorize the reading process for decision-making tasks.