跳转至

文章背景与核心概要

大语言模型驱动的网络搜索智能体在执行复杂任务时，经常面临一种被称为“惯性偏差（Inertia Bias）”的顽固失效模式：一旦智能体自主生成了某个查询、计划或中间结论，它就很难再对该行动的后续后果进行客观评估。为了系统性地量化和解决这一问题，本文作者引入了 IBIS 基准测试，并提出了 NIS-Agent 框架，通过在网页分类筛选和最终答案验证两个关键决策点引入上下文隔离，有效削减了系统级降级和搜索/上下文噪声。

该研究不仅成功将 Token 成本降低了 33%，同时保持了极具竞争力的性能；此外，通过训练一个 8B 参数的模型使其具备内在的抗惯性偏差能力，该框架在深度研究基准测试中达到了与 GPT-4o 媲美的平均性能表现，为未来构建更具鲁棒性的自主研究智能体提供了重要的技术路径。

From Inertia to Objectivity: Improving Deep Research Agents with Noise Isolation

arXiv ID: 2608.23045 [cs.AI]
Authors: Xiangxin Zhang, Zhanwei Zhang, Zhihang Fu, Binbin Lin, Wenxiao Wang
Submitted: August 24, 2026 (EMNLP 2026)

arXiv ID: 2608.23045 [cs.AI]
Authors: Xiangxin Zhang, Zhanwei Zhang, Zhihang Fu, Binbin Lin, Wenxiao Wang
Submitted: August 24, 2026 (EMNLP 2026)

📌 Executive Summary

大语言模型（LLM）驱动的网络搜索智能体常常受到惯性偏差（inertia bias）的困扰——这是一种反复出现的失效模式：一旦智能体编写了某个查询、计划或结论，它在后续就很难客观地判断该查询、计划或结论的后果。为了分析和解决这个问题，作者引入了 IBIS 基准测试，以隔离搜索观测结果并衡量自我署名导致的认知失真（self-authorship distortion）。

论文表明，这种偏差会级联导致系统级的性能退化，从而在工作者（worker）层面引发搜索噪声，并在管理者（manager）层面引发上下文噪声。为了缓解这一问题，研究人员提出了 NIS-Agent，它在网页分类筛选（webpage triage）和最终答案验证（final-answer validation）阶段应用了上下文隔离。NIS-Agent 在保持竞争力的同时，成功将 Token 成本降低了 33%。此外，通过训练一个 8B 模型使其本质上能够抵御惯性偏差，该框架实现了可与 GPT-4o 相媲美的深度研究性能。

📌 Executive Summary

Large Language Model (LLM)-powered web search agents often suffer from inertia bias—a recurring failure mode where an agent struggles to objectively judge the consequences of a query, plan, or conclusion once it has authored it. To analyze and solve this, the authors introduce the IBIS benchmark to isolate search observations and measure self-authorship distortion.

The paper demonstrates that this bias cascades into system-level degradation, causing search noise at the worker level and contextual noise at the manager level. To mitigate this, the researchers propose NIS-Agent, which applies context isolation during webpage triage and final-answer validation. NIS-Agent successfully cuts token costs by 33% while maintaining competitive performance. Furthermore, by training an 8B model to resist inertia bias intrinsically, the framework achieves deep-research performance comparable to GPT-4o.

📋 Bibliographic & Paper Details

📋 Bibliographic & Paper Details

🔍 Abstract & Key Findings

摘要： 大语言模型（LLM）驱动的网络搜索智能体展现出巨大的潜力，但深度研究任务暴露出一种反复出现的失效模式：一旦智能体生成了某个查询、计划或中间结论，它在随后评估该行动的后果时就会变得不够客观。我们将这种现象称为惯性偏差（inertia bias）。为了使其可量化，我们引入了 IBIS 基准测试，该测试在控制搜索观测结果的同时，改变模型是否在评估其自身先前行动的结果。我们发现，当模型“拥有”先前的搜索步骤时，其表现会显著变差，这表明自主编写的行动历史会系统性地扭曲后续的判断。我们进一步证明，这种偏差会演变成两种形式的系统级退化：工作者层面的搜索噪声和管理者层面的上下文噪声。为了解决这个问题，我们提出了 NIS-Agent，它在最容易受到惯性偏差影响的两个决策点——网页分类筛选和最终答案验证——应用了上下文隔离。在 GAIA、WebWalkerQA、BrowseComp 和 BrowseComp-zh 上，NIS-Agent 在保持竞争力的同时，与基线相比将 Token 成本降低了 33%。我们进一步训练了一个 8B 模型，使其对惯性偏差具有更强的内在抵抗力；在相同的 NIS-Agent 框架下，它在深度研究基准测试上取得了与 GPT-4o 相当的平均性能。

Abstract: Web search agents powered by Large Language Models (LLMs) show strong promise, but deep research tasks expose a recurring failure mode: once an agent has produced a query, plan, or intermediate conclusion, it becomes less objective when later judging the consequences of that same action. We term this phenomenon inertia bias. To make it measurable, we introduce the IBIS benchmark, which controls the search observations while varying whether the model is evaluating the outcome of its own prior action. We find that models are substantially worse when they "own" the preceding search step, showing that self-authored action history can systematically distort subsequent judgment. We further show that this bias propagates into two forms of system-level degradation: search noise at the worker level and contextual noise at the manager level. To address this problem, we propose NIS-Agent, which applies context isolation at the two decision points most vulnerable to inertia bias: webpage triage and final-answer validation. Across GAIA, WebWalkerQA, BrowseComp, and BrowseComp-zh, NIS-Agent achieves competitive performance while reducing token cost by 33% compared to our baseline. We further train an 8B model to be intrinsically more resistant to inertia bias; under the same NIS-Agent framework, it attains average performance comparable to GPT-4o on deep research benchmarks.

🛠️ Associated Resources & Tools

🛠️ Associated Resources & Tools