跳转至

联邦化几乎免费,推理则不然:蛋白质表征工作流中 AI 联合科学家的权衡

文章背景与核心概要

由自然语言驱动的自主“联合科学家”(co-scientist)工作流在灵活性、推理能力与确定性、可重复性及可观测性之间面临着内在的张力。随着这些智能体在跨机构边界的协作日益频繁,理解联邦拓扑结构如何影响延迟和成本变得至关重要。

本文通过在生产级智能体科学平台上进行受控消融实验,系统地评估了上述权衡。作者利用蛋白质表征任务(通过路由工具从序列中确定蛋白质功能)作为基准,对比了不同的联邦拓扑结构、经典强化学习(RL)与大语言模型(LLM)驱动的框架、底层模型以及提示词工程的效果,并根据蛋白质的新颖性对结果进行了分层分析。


执行摘要

Natural language-driven autonomous "co-scientist" workflows face an inherent tension: they must balance flexibility and reasoning against determinism, reproducibility, and observability. As these agents increasingly communicate across institutional boundaries, understanding how federation topology affects latency and cost becomes critical.

由自然语言驱动的自主“联合科学家”工作流面临着一种内在的张力:它们必须在灵活性与推理能力以及确定性、可重复性与可观测性之间取得平衡。随着这些智能体在机构边界之间进行越来越多的通信,理解联邦拓扑结构如何影响延迟和成本变得至关重要。

This paper presents a systematic evaluation of these tradeoffs via controlled ablations on a production agentic scientific platform. Using a verifiable protein characterization task (routing through common tools to confidently determine a protein's function from its sequence), the authors compare federation topologies, classic reinforcement learning (RL) versus LLM-driven harnesses, underlying language models, and prompt expertise, while stratifying results by protein novelty.

本文通过在生产级智能体科学平台上进行受控消融实验,对这些权衡进行了系统性评估。作者使用了一项可验证的蛋白质表征任务(通过路由常用工具来准确确定蛋白质的序列功能),对比了联邦拓扑结构、经典强化学习(RL)与 LLM 驱动的框架、底层语言模型以及提示词专业性,并根据蛋白质的新颖性对结果进行了分层。


关键发现

  • Model Selection Trumps Architecture: The choice of Large Language Model (LLM) dominated prediction quality far more than federation topology or prompting strategies (e.g., Opus achieved ~92–94% accuracy vs. o4-mini at ~40–50%).
  • 模型选择优于架构: 大语言模型(LLM)的选择对预测质量的影响远超联邦拓扑结构或提示词策略(例如,Opus 达到了约 92–94% 的准确率,而 o4-mini 仅为 40–50%)。
  • Efficiency of Deterministic Policies: A Proximal Policy Optimization (PPO) policy achieved near-frontier accuracy (88%) with zero token cost, minimal latency, and perfect consistency, though it lacked an interpretable reasoning trace.
  • 确定性策略的效率: 近端策略优化(PPO)策略在零 Token 成本、极低延迟和完美一致性的前提下,实现了接近前沿水平的准确率(88%),尽管它缺乏可解释的推理轨迹。
  • The Cost of Expert Prompting: Expert-prompted LLMs reached peak accuracy, but came with high financial costs and reduced consistency. Prompt dependence was particularly pronounced for harder, more novel protein tasks.
  • 专家提示词的成本: 经过专家提示词优化的 LLM 达到了峰值准确率,但伴随着高昂的财务成本和降低的一致性。对于更困难、更新颖的蛋白质任务,对提示词的依赖性尤为明显。
  • Negligible Federation Overhead: Federation imposed an almost negligible performance penalty on scientific workflows.
  • 联邦化开销可忽略不计: 联邦化对科学工作流造成的性能损耗几乎可以忽略不计。

实践建议

  • For Routine, Verifiable Tasks: Utilize a cheap, deterministic policy to secure near-frontier accuracy alongside complete reproducibility.
  • 针对常规、可验证的任务: 利用廉价的确定性策略,在确保完全可重复性的同时获得接近前沿水平的准确率。
  • For Open-Ended Discovery: Reserve flexible, higher-cost LLM reasoning for exploratory scenarios where interpretability and qualitative reasoning traces are paramount.
  • 针对开放式发现任务: 将灵活且成本较高的 LLM 推理保留用于探索性场景,在这些场景中,可解释性和定性推理轨迹至关重要。

链接与资源