Prefix Sliding for efficient test-time scaling
Prefix Sliding drops stale reasoning tokens while retaining instructions and recent work, capping memory during long inference and reportedly making existing models 3× faster without training.
Prefix Sliding retains the instruction-and-tools prefix plus a window of the latest few thousand tokens, discarding older intermediate reasoning. This caps memory independently of reasoning length and made existing models **3× faster without training** while maintaining performance.
Builders operating long-running agents should distinguish durable context from transient reasoning. Keeping stable instructions and current work may be more efficient than retaining every intermediate token; reinforcement learning with the method enabled traces beyond **100,000 tokens**.
Prefix Sliding retains the instruction-and-tools prefix plus a window of the latest few thousand tokens, discarding older intermediate reasoning. This caps memory independently of reasoning length and made existing models **3× faster without training** while maintaining performance. Builders operating long-running agents should distinguish durable context from transient reasoning. Keeping stable instructions and current work may be more efficient than retaining every intermediate token; reinforcement learning with the method enabled traces beyond **100,000 tokens**. The claims come from a new paper, and the supplied material does not name the evaluated models, workloads, or deployment costs. Its ablations beat summarization and a vanilla sliding window, but broader generality remains open.
Prefix Sliding reframes long-context reasoning as a retention-policy problem: preserve stable instructions, tools, and recent work while dropping stale intermediate reasoning. Its reported speed and memory gains complement methods that reduce how much reasoning is generated, but the absent model, workload, and cost details prevent treating the 3× result as a general deployment expectation.