AI

Periscope: Extending Frozen Language Models Beyond Their Context Window

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 08:18:25 UTC.

Briefings

Periscope is a training-free inference method that factorizes long text reading into local and strided spans to bypass context window limits. On LongBench v2, it matched the best window read across 32k to 1M tokens, while on InfiniteBench it led by 5 points at 150k token median contexts. The approach uses a K{times}K grid to query K local and strided spans, caching only one probe per call to reduce GPU memory requirements.

Key points

• Periscope factorizes long text reading into local and strided spans to bypass context window limits.
• On LongBench v2, Periscope matched the best window read across 32k to 1M tokens, while on InfiniteBench it led by 5 points at 150k token median contexts.

Context

This research addresses the limitation of standard language models losing accuracy when reading long texts beyond their context window. Periscope offers a training-free inference method that replaces long reads with a more efficient scoring mechanism using local and strided spans.

Limitations

Evidence is limited to two benchmark datasets (LongBench v2 and InfiniteBench) with specific token counts and model sizes. Results are from a single snapshot on 2026-10-01 and do not include performance on other benchmarks or real-world deployment scenarios.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 19/10/2026, 07:22:14 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeie5sfvb377d6mw4ibgggbqznlnqu73ubi7tlujca3nevajaak3biu