PulseAugur
实时 23:31:30
English(EN) 1/ DSGym: A Holistic Framework for Evaluating and Training Data Science Agents

Together AI 发布 8 项新的 LLM 推理和训练系统进展

Together AI 发布了一系列研究论文,详细介绍了 LLM 推理和训练系统的进展。其中包括通过 Batch-Aware Expert Routing (OEA) 优化专家混合 (MoE) 模型的方法,以及使用 Ulysses 实现内存高效的上下文并行。该公司还展示了 Aurora,一个用于自适应推测训练的统一系统,以及 V1,它统一了并行推理器的生成和自我验证。其他创新包括用于通过演示学习推理的 RARO、用于 AI 驱动的科学发现的 TTT-Discover、用于程序感知代理推理的 ThunderAgent,以及用于评估和训练数据科学代理的 DSGymAI

影响 这些进展旨在提高 LLM 的效率、推理能力和代理工作流程,从而可能加速 AI 驱动的发现和复杂任务的执行。

排序理由 Together AI 发布了多篇详细介绍新 LLM 推理和训练技术的论文。

在 X — Together (inference / OSS) 阅读 →

AI 生成摘要 · Google Gemini · 来自 8 个来源。 我们如何撰写摘要 →

Together AI 发布 8 项新的 LLM 推理和训练系统进展

报道来源 [8]

  1. X — Together (inference / OSS) TIER_1 English(EN) · togethercompute ·

    8/ Opportunistic Expert Activation: Batch-Aware Expert Routing for Faster Decode Without Retraining (OEA)

    8/ Opportunistic Expert Activation: Batch-Aware Expert Routing for Faster Decode Without Retraining (OEA) https://t.co/dw33plIoxW

  2. X — Together (inference / OSS) TIER_1 English(EN) · togethercompute ·

    7/ Untied Ulysses: Memory-Efficient Context Parallelism via Headwise Chunking

    7/ Untied Ulysses: Memory-Efficient Context Parallelism via Headwise Chunking https://t.co/LgGqu8vl97

  3. X — Together (inference / OSS) TIER_1 English(EN) · togethercompute ·

    6/ When RL Meets Adaptive Speculative Training: A Unified Training-Serving System (Aurora)

    6/ When RL Meets Adaptive Speculative Training: A Unified Training-Serving System (Aurora) https://t.co/fvLuHrqDbX

  4. X — Together (inference / OSS) TIER_1 English(EN) · togethercompute ·

    5/ V1: Unifying Generation and Self-Verification for Parallel Reasoners

    5/ V1: Unifying Generation and Self-Verification for Parallel Reasoners https://t.co/X1zUsS7gY8

  5. X — Together (inference / OSS) TIER_1 English(EN) · togethercompute ·

    4/ Escaping the Verifier: Learning to Reason via Demonstrations (RARO)

    4/ Escaping the Verifier: Learning to Reason via Demonstrations (RARO) https://t.co/gQZCEav8Nb

  6. X — Together (inference / OSS) TIER_1 English(EN) · togethercompute ·

    3/ Learning to Discover at Test Time (TTT-Discover)

    3/ Learning to Discover at Test Time (TTT-Discover) https://t.co/pKeadv4DHl

  7. X — Together (inference / OSS) TIER_1 English(EN) · togethercompute ·

    2/ ThunderAgent: A Simple, Fast and Program-Aware Agentic Inference System

    2/ ThunderAgent: A Simple, Fast and Program-Aware Agentic Inference System https://t.co/7I1Yf5s8B8

  8. X — Together (inference / OSS) TIER_1 English(EN) · togethercompute ·

    1/ DSGym: A Holistic Framework for Evaluating and Training Data Science Agents

    1/ DSGym: A Holistic Framework for Evaluating and Training Data Science Agents https://t.co/jV4uMB1g48