Periscope: training-free method reads 1M-token texts with a frozen 9k-token window
_reachsumit · x · 2026-10-06
The arXiv paper Periscope introduces a training-free inference method that lets a frozen LLM read far beyond its context window.
How it works
- N chunks are arranged on a K×K grid (K=⌈√N⌉); the frozen model answers the same question over K contiguous local spans and K strided spans sampling the whole text, reading answer log-odds at a single token.
- Each answer takes its best local/strided score; scoring every chunk yields an evidence map at no extra cost, whose peak marks the supporting passage.
- Each probe costs √(s·c) tokens, so a W-token window reaches W²/c tokens at s^1.5 cost.
Results
- On LongBench v2, reading only the K top-ranked chunks (9k tokens) matches the same model's best window read from 32k to 1M tokens.
- On InfiniteBench (median 150k contexts), it beats the best window read by 5 points.
- Best NDCG@10 of six methods on BRIGHT long-document retrieval.
Related event: Periscope Extends Frozen LMs to Million-Token Contexts Without Training(2 posts)→
More from Research
- Interactive explainer makes TF-IDF, BM25 and NDCG click, BM25 by example — JnBrymn · 2026-10-06
- Poster on curating long-context reasoning data presented today at Poster Session 1 — tuvllms · 2026-10-06
- EvoSkill research to be presented at Thursday poster session, Poster #52 — tuvllms · 2026-10-06
- AI-Discovered Algorithm Delivers First True Subquadratic 3SUM and Subcubic APSP — FrnkNlsn · 2026-10-06
- Amazon AGI's ALoDLM Beats Diffusion and AR Baselines, Hits 612 tok/s at 8B — arankomatsuzaki · 2026-10-06
- Schmidhuber: His 1991 Paper Introduced Pre-training, Positional Encoding and Distillation — SchmidhuberAI · 2026-10-06