LectureRechercheRadarCadre d'investissement
Connexion / Inscription
← Retour à la lecture
SujetRédaction DisconfirmAIMis à jour 2026-10-27 19:05

NVIDIA

2 étapes1 pièces3 Sources

【背景】NVIDIA研究科学家Dhruv Nathawani在2026年AI Engineer大会上展示了一条基于Wikidata知识图谱的合成数据管线,旨在训练大模型调用搜索工具而非凭记忆作答。该管线通过随机游走获取实体路径并转化为多步搜索谜题,经Tavily实时搜索生成轨迹后用于监督微调。报告显示,5万个种子数据经严格过滤保留7千条高质量记录,使模型在BrowseComp基准上的表现从0%提升至30%。然而,Nathawani也坦承验证环节的困难:即使源自Wikidata的谜题,其答案也可能因事实过时而失效,导致生成器与验证器同时给出错误结果。这一案例揭示了合成数据在解决“幻觉”与保持“时效性”之间的张力,表明单一模型评分或静态源图无法确立数据集质量,需结合确定性检查、人工复核与动态事实证据。

Comment on en est arrivé là

  1. 2026-10-27

    NVIDIA研究员Dhruv Nathawani在AI Engineer大会展示Wikidata搜索训练管线,报告BrowseComp得分从0%升至30%。

    youtube.com ↗
  2. 2026-10-27

    NVIDIA NeMo Data Designer发布开发笔记,详述四阶段管线:从5万Wikidata种子到7千SFT轨迹,使用MiniMax-M2进行Rollout。

    docs.nvidia.com ↗

Ce qui reste non résolu

  • 如何解决Wikidata等静态知识源的事实过时问题,以确保训练数据的长期有效性?
  • 除了BrowseComp,该管线在其他复杂推理基准上的泛化能力如何?

Sources

  • youtube.comYouTube: NVIDIA Researcher Talk↗
  • docs.nvidia.com
Connexion / Inscription
LectureRechercheRadarCadre d'investissement
NVIDIA Docs: Search Agent SFT Data
↗
  • ai.engineerAI Engineer: Speaker Profile↗
  • Couverture dans ce sujet

    • NVIDIA Researcher Dhruv Nathawani: Constructing Search Puzzles from Wikidata Graph Paths to Train Models to Use Search Tools Instead of Answering from Memory2026-10-27Actualités

      At the AI Engineer World's Fair, NVIDIA research scientist Dhruv Nathawani presented a synthetic data pipeline that extracts paths via random walks on the Wikidata entity graph, transforms those paths into puzzles requiring multi-step searches to solve, and then uses search tools to execute the full trajectory for supervised fine-tuning. He reported that after generating and filtering 50,000 seeds, 7,000 training records were retained, moving the model’s performance on the BrowseComp benchmark from 0% to 30%. He also acknowledged difficulties in verification: for a puzzle about the ownership of U.S. Steel, both the generator and verifier provided outdated answers.