Stanford's Prefix Sliding speeds up long reasoning by 3x
omarsar0 · x · 2026-08-31
Stanford researchers introduced Prefix Sliding, a new method to optimize test-time compute for long reasoning chains (CoT).
- Problem: Long reasoning traces require full attention, keeping the entire history in memory. This makes the hardest problems (which need the most thinking) the most expensive to run.
- Insight: Intermediate tokens steadily lose importance as the model continues reasoning.
- Solution: Prefix Sliding retains the prefix (instructions and available tools) and a sliding window of the last few thousand tokens, discarding everything in between. This caps total memory usage regardless of reasoning length.
- Results: Achieves a 3x speedup on existing models without any training, while matching full-attention performance.
More from Infra
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16
- Google Raises 2026 Capital Spending to Up to $205 Billion — emmanuelvivier · 2026-10-16
- Google open-sources AX, a Kubernetes-style orchestrator for billions of AI agent tasks — thisdudelikesAI · 2026-10-16