文章背景与核心概要
在大语言模型(LLM)驱动的网页搜索智能体日益普及的背景下,研究人员发现深度研究任务中存在一种普遍的失效模式:惯性偏见(inertia bias)。也就是说,当智能体生成了某个查询、计划或中间结论后,在评估自身先前行为的后果时,其客观性会显著下降。这种由自主生成的行动历史所导致的判断扭曲,会进一步演变为系统级的噪声(工作层的搜索噪声和管理层的上下文噪声)。
为了应对并量化这一问题,本文作者推出了 IBIS 基准测试,并提出了 NIS-Agent 架构。该架构在最容易受到惯性偏见影响的两个决策点(网页分类筛选和最终答案验证)上应用了上下文隔离技术。在 GAIA、WebWalkerQA、BrowseComp 和 BrowseComp-zh 等多个基准测试中,NIS-Agent 不仅取得了极具竞争力的性能,还将 Token 成本降低了 33%。此外,通过训练一个 8B 模型使其内在具备抵抗惯性偏见的能力,该模型在相同的框架下甚至能够达到与 GPT-4o 媲美的深度研究性能。
从惯性走向客观:通过噪声隔离提升深度研究智能体的性能
作者: Xiangxin Zhang, Zhanwei Zhang, Zhihang Fu, Binbin Lin, Wenxiao Wang
主要学科: 计算机科学 > 人工智能 (cs.AI)
arXiv ID: arXiv:2608.23045 [cs.AI]
收录说明: EMNLP 2026 主会 (Main Conference)
📌 总结
Web search agents driven by Large Language Models (LLMs) often suffer from a recurring failure mode known as inertia bias: once an agent generates a query, plan, or intermediate conclusion, it becomes significantly less objective when evaluating the consequences of its own prior actions.
由大语言模型(LLM)驱动的网络搜索智能体经常面临一种反复出现的失效模式,被称为惯性偏见(inertia bias):一旦智能体生成了查询、计划或中间结论,它在评估自身先前行为的后果时就会变得明显不够客观。
To address and quantify this issue, the authors introduce the IBIS benchmark, demonstrating how self-authored action histories distort subsequent judgment and propagate into system-level noise (search noise at the worker level and contextual noise at the manager level). To combat this, they propose NIS-Agent, which applies context isolation during webpage triage and final-answer validation.
为了解决并量化这一问题,作者推出了 IBIS 基准测试,展示了自主编写的行动历史如何扭曲后续判断,并蔓延成系统级噪声(工作层的搜索噪声和管理层的上下文噪声)。为了应对这一挑战,他们提出了 NIS-Agent,在网页分类筛选(triage)和最终答案验证阶段应用上下文隔离。
Across benchmarks like GAIA, WebWalkerQA, BrowseComp, and BrowseComp-zh, NIS-Agent achieves competitive performance while cutting token costs by 33%. Furthermore, training an 8B model to be intrinsically resistant to inertia bias allows it to match GPT-4o's deep research performance using the same framework.
在 GAIA、WebWalkerQA、BrowseComp 和 BrowseComp-zh 等基准测试中,NIS-Agent 实现了具有竞争力的性能,同时将 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抗惯性偏见,它在相同的框架下能够匹配 GPT-4o 的深度研究性能。
📋 摘要
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.
由大语言模型(LLM)驱动的网络搜索智能体展现出了巨大的潜力,但深度研究任务暴露出了一种反复出现的失效模式:一旦智能体产生了一个查询、计划或中间结论,它在随后判断该行为的后果时就会变得不够客观。我们将这种现象称为惯性偏见。为了使其可量化,我们推出了 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先前自主行动的结果。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。我们进一步表明,这种偏见会演变为两种形式的系统级退化:工作层的搜索噪声和管理层的上下文噪声。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见影响的两个决策点(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 任务中,与基线相比,NIS-Agent 在降低 33% Token 成本的同时实现了竞争力的性能。我们进一步训练了一个 8B 模型,使其本质上更能抵抗惯性偏见;在相同的 NIS-Agent 框架下,它在深度研究基准测试上获得了与 GPT-4o 相当的平均性能。
🔗 链接与资源
- 查看 PDF: PDF 直链
- HTML 版本: arXiv HTML(实验性)
- 代码仓库: GitHub - PangSMPang/NIS-Agent
- 引用与工具:
- Google Scholar
- Semantic Scholar
- NASA ADS
📅 提交历史
- [v1] 2026年8月24日 周一 09:50:24 UTC
- [v2] 2026年8月27日 周四 02:49:14 UTC (当前版本)