CoreWeave, You.com and NVIDIA lift BrowseComp 22.9% in 3-week RL post-training sprint
RichardSocher · x · 2026-10-27
CoreWeave, NVIDIA and You.com ran a three-week sprint post-training the open-weight Nemotron 3.5 Lightning for agentic web search, combining SFT, RL and managed RL Rollouts.
- BrowseComp accuracy rose from 36.97% to 45.45% (a 22.94% relative gain), with 31.55% fewer tool calls per task.
- The core bottleneck in RL post-training is inference cadence: rollouts, policy updates and weight synchronization can stall rollout generation and idle accelerators. CoreWeave's RL Rollouts, built on NVIDIA Dynamo, cuts that synchronization overhead.
- The writeup details You.com's data strategy, CoreWeave's managed infrastructure, and lessons teams can reuse in their own post-training.
More from Infra
- PyTorch details 2.12x faster on-device streaming ASR on Arm SME2 via ExecuTorch — PyTorch · 2026-10-27
- Revenue per GPU can rise as token prices fall, argues Gavin Baker — GavinSBaker · 2026-10-27
- Anthropic S-1: only $19B total compute spend since inception, overbuild warning — bookwormengr · 2026-10-27
- Grid queues make power architecture an AI-factory decision; Exowatt and Redwood offer fixes — ingliguori · 2026-10-27
- Two CVSS 9.9 Wasmtime flaws allow sandbox escape, one ARM64-only since 2020 — basedjensen · 2026-10-27
- AI Infra Meme: While Countries Need TW-Scale Power for Datacenters, People Argue About Lightbulbs — mgostIH · 2026-10-27