EMNLP Oral: training-free Nexus Sampling cuts KV-cache memory 10x; OpenResearcher tops BrowseComp-Plus at 54.8%
yuz9yuz · x · 2026-10-27
Lambda and collaborators published two papers at EMNLP 2026 in Budapest, including an Oral:
- Nexus Sampling: a training-free KV-cache eviction method for long-context and agentic LLM inference. It replaces the standard per-step direct-attention-score + deterministic top-K template with Nexus scoring (an iterative walk over direct attention that surfaces bridge tokens) plus weighted reservoir sampling. At 80% eviction it matches dense attention within 1% on LongBench, beats top-K baselines on retrieval-heavy tasks, and shrinks per-sequence cache memory up to 10x.
- OpenResearcher: a fully open pipeline for long-horizon deep research trajectory synthesis, enabling scalable, reproducible training of deep research agents; its data is already adopted by NVIDIA's Nemotron family including Nemotron 3 Ultra. A perf/params scatter shows OpenResearcher leading BrowseComp-Plus at 54.8%, ahead of much larger models like Claude-Opus-4 (36.8%), GPT-4.1 (36.4%), and Gemini-2.5-Pro (29.5%).
More from Infra
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27
- Running a local 9B as a coding agent: brief the big model to plan, let the small model execute — Electrical_Ninja3805 · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27