Sliding reasoning windows make long test-time scaling 3x faster

Muennighoff et al. show that discarding stale intermediate tokens caps memory use while preserving performance, and RL training with the method extends reasoning beyond 100k tokens.

Top University

Niklas Muennighoff · Zhengyang Wang · Zeyi Chen · Weijia Shi · Binyuan Hui · John Yang · +12 more

Stanford University · University of California, Santa Barbara · Prime Intellect · University of Washington

Research Digest·27 August 2026·2 min read

Thread:Context Engineering for Agents

In one line

Prefix Sliding makes existing models 3x faster while maintaining performance during test-time scaling by keeping only the prefix and a recent window.

Before you read this

Two ideas this paper builds on

test-time scalingAllocating extra computation during inference to improve output quality.90 papers →reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.239 papers →

Read it as

Prefix Sliding is a technique for cutting memory use during . Instead of keeping the whole reasoning trace, it keeps the prompt prefix plus a sliding window over the most recent tokens, dropping intermediate steps. Because those intermediate tokens receive little attention once the model has moved on, the memory cost becomes constant no matter how long the model thinks.

Without retraining, this gives existing models a 3x speedup at the same performance. With , it enables rollouts beyond 100,000 tokens and better final performance. Ablations show it beats summarizing old tokens or using a sliding window alone.

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

What it released

Code

Concepts

test-time scaling prefix sliding context window reinforcement learning token efficiency long-context processing