Source-linked AI summary
From Inertia to Objectivity: Improving Deep Research Agents with Noise Isolation
Xiangxin Zhang, Zhanwei Zhang, Zhihang Fu, Binbin Lin, Wenxiao Wang
TL;DR
Deep-research agents can lose objectivity toward their own action history, a failure the paper measures with IBIS and addresses with NIS-Agent’s context isolation. NIS-Agent isolates webpage triage and final-answer validation, while an 8B trained model reaches performance comparable to GPT-4o; the approach reduces, but does not eliminate, self-reinforcing errors.
Problem
LLMs become less objective when judging the consequences of their own queries, plans, or intermediate conclusions, creating inertia bias in deep research.
Method
NIS-Agent applies context-isolated filtering for worker webpage triage and isolation-based stepwise validation for manager final-answer checking.
Results
An 8B open-source model trained to resist inertia bias achieves deep-research performance comparable to GPT-4o.
Takeaways & Limitations
Context isolation mitigates inertia bias at the decision points most vulnerable to self-authored action-history effects.
Takeaways & Limitations
NIS-Agent can still discard relevant pages or endorse flawed conclusions, so it reduces rather than eliminates self-reinforcing errors.
Abstract
from arXiv · showhide
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 \textbf{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.
1 Introduction
Deep-research agents can become anchored to their own queries, plans, and conclusions, producing search and contextual noise. The paper introduces IBIS to measure this inertia bias and NIS-Agent to isolate vulnerable judgments.
- Failure mode: Inertia bias makes LLMs less objective when judging the consequences of their own queries, plans, or intermediate conclusions.IBIS isolates the effect of action history on information-seeking decisions.
- IBIS benchmark: Agentic Mode fails significantly more often than Observer Mode on identical search results, demonstrating that self-authored action history distorts judgment.Agentic Mode includes the model’s own query history, whereas Observer Mode presents the same results as external information.
- System effects: Inertia bias amplifies search noise at the worker level and contextual noise at the manager level.Workers may continue browsing superficially relevant pages, while managers can be misled by accumulated reasoning context.
- Proposed approach: NIS-Agent applies context isolation to webpage filtering and final-answer validation to restore objectivity at key decision points.The filter rejects irrelevant results before browsing, while validation checks local inferences under isolated context.
- Results: 33% lower token cost accompanies NIS-Agent’s performance enhancement over the smolagents DR baseline.The framework also reports state-of-the-art performance among open-source frameworks on GAIA and WebWalkerQA with stronger model configurations.
- Results: An 8B open-source model trained for resistance to inertia bias achieves deep-research performance comparable to GPT-4o.The model is trained with supervised fine-tuning and GRPO.
2 Related Work
The related work distinguishes inertia bias from sycophancy and confirmation bias and situates deep research within multi-step internet search supported by manager–worker agents.
- Cognitive biases: Inertia bias is triggered by a model’s own prior action, unlike sycophancy’s external framing or confirmation bias’s prior belief.The distinction concerns the source of influence rather than the general presence of anchoring.
- Deep research agents: Deep research consists of multi-step internet searches for complex tasks, often using manager–worker communication patterns.Managers decompose tasks and workers gather information for subtasks.
- Existing methods: Existing manager-focused methods reduce noise using external experience databases, including action-knowledge bases, teacher guidance, and summarized sub-workflows.These methods include KnowAgent, Agent KB, and Agent Workflow Memory.
3 Inertia Bias in Information Seeking
The paper operationalizes inertia bias as path dependence on an agent’s own search action and measures it with controlled IBIS conditions that separate action history from search-result content.
- Benchmark: IBIS is a diagnostic benchmark designed to isolate the impact of action history on information-seeking decisions.It targets query generation, search execution, and the choice between browsing and query revision.
- Inertia Bias as an Agentic Failure Mechanism: Inertia bias is the tendency to select URLs from an unproductive query even when returned summaries indicate that re-searching is correct.This behavior reflects path dependence on the agent’s own prior action.
- Benchmark construction: The benchmark labels search outcomes as Should Re-search, Should Visit Page, or Should Return Answer based on retrieved abstracts.Samples are retained only when three annotators agree on the category.
- Evaluation: IBIS evaluates both general next-step reasoning and irrational adherence to the current search path on the Should Re-search subset.The subset directly measures failures to reject factually insufficient results.
- Evaluation: Agentic Mode includes the model’s self-issued search history, while Observer Mode presents identical results without that ownership context.The controlled comparison decouples action-history influence from content reasoning.
4 NIS-Agent
NIS-Agent isolates the two decisions most vulnerable to inertia bias: worker webpage triage and manager final-answer validation. These modules reduce noise by separating judgments from the execution history that produced them.
- Framework: NIS-Agent targets worker-level search noise during webpage triage and manager-level contextual noise during final-answer validation.It intervenes at these two decision points without redesigning the standard multi-agent architecture.
- SearchAgent: The context-isolated filter screens candidate webpages before access using a compact task-relevant decision context.It can trigger adaptive query rewriting when relevance is insufficient.
- SearchAgent: The filter separates relevance judgments from the SearchAgent’s full execution trajectory to reduce rationalization of an unproductive search path.This isolation supports abandoning the current path when abstracts are insufficiently relevant.
- ManagerAgent: The isolation-based validation module audits proposed answers after the ManagerAgent first marks them ready.It replaces validation that reuses the entire reasoning trace and remains exposed to inertia bias.
5 NIS-8B
NIS-8B is trained through a two-stage pipeline combining supervised imitation with reinforcement learning for inertia-resistant decision-making. The reinforcement stage focuses on single web-search decisions and assigns reward at the point where inertia bias manifests.
- Training pipeline: NIS-8B uses a two-stage training pipeline built on Qwen3-8B.The model is named NIS-8B after this training process.
- Supervised fine-tuning: The SFT stage teaches tool-call competence and warms up the task distribution using teacher-generated IBIS examples with gold actions.It draws on a public tool-use corpus and synthesizes decision examples following the IBIS construction protocol.
- Reinforcement learning: GRPO follows SFT so the model can internalize anti-inertia behavior through self-exploration and feedback.The authors motivate this continuation because imitation alone cannot correct inertia-biased decisions during rollouts.
- Decision setting: Each rollout begins after the agent’s first web_search observation and contains one decision: revise the query, browse a candidate page, or terminate.This concentrates training on the decision point where inertia bias appears.
- Reward design: R = 0.1 · Rformat + 0.9 · Raction combines protocol adherence with correctness against a pre-computed reference rubric.The reward is concentrated at the decision point, providing denser credit assignment than evaluating the full trajectory’s final answer.
6 Experiment
Experiments evaluate NIS-Agent across deep-research benchmarks, inertia-bias settings, efficiency, ablations, and transfer to reasoning tasks. The results show competitive or leading performance, lower token use, component-level gains, and broader applicability.
- Experimental Setup: NIS-Agent is evaluated on GAIA, WebWalkerQA, BrowseComp, BrowseComp-zh, and AIME 2024/2025, using multiple model configurations.GAIA uses its full 165-query validation set; the other browsing benchmarks use 200 sampled examples, and AIME reports average accuracy across five runs.
- Diagnosing Inertia Bias with IBIS: 15% to 30% performance improvement occurs when models switch from Agentic Mode to Observer Mode on the IBIS Should Re-search subset.The modes share identical search results but differ in whether the model’s prior query is presented as self-initiated; NIS-8B shows a minimal mode gap.
- Main Results in Deep Research: 81.21% on GAIA and 75.00% on WebWalkerQA make NIS-Agent state-of-the-art among open-source frameworks with DeepSeek-V4-Pro.With Claude-3.7-Sonnet, NIS-Agent reaches 72.73% on GAIA and 68.50% on WebWalkerQA; gains over smolagents DR with GPT-4.1 and GPT-4o reach approximately 8 to 11 percentage points.
- Main Results in Deep Research: 25.0 on BrowseComp and 45.9 on BrowseComp-zh substantially outperform the compared open-source baselines.OpenAI DR remains stronger on BrowseComp overall, while NIS-Agent achieves the best result among listed systems on BrowseComp-zh.
- Efficiency: 33% lower average total token usage per query on GAIA with GPT-4o is achieved relative to smolagents DR.The reduction mainly comes from cutting input tokens.
- Ablation Study: 61.82 GAIA accuracy results after adding both validation stages, compared with 56.96 without them and 58.79 after Stage 1 alone.The ablation indicates distinct contributions from the context-isolated filter and the two-stage validation process; replacing standard CoT validation also improves both GPT-4o and GPT-4.1 on AIME.
7 Conclusions
The paper identifies and quantifies inertia bias, then proposes NIS-Agent to isolate judgments vulnerable to action-history ownership. It also trains an 8B model with greater intrinsic resistance to the bias.
- Inertia bias is defined as LLMs reinforcing prior choices regardless of objective correctness and is quantified with the IBIS benchmark.
- NIS-Agent isolates judgments most vulnerable to action-history ownership to mitigate inertia bias in deep research.
- An 8B open-source model trained to resist inertia bias attains performance comparable to GPT-4o.
8 Limitations
The paper’s limitations concern its black-box treatment of inertia bias, uncertain generalization across agentic domains, and risks from incorrect isolation or validation.
- The study does not investigate the mechanistic origins of inertia bias within the underlying language models.It leaves attention patterns, pretraining distributions, reinforcement-learning artifacts, and internal representations unexplored.
- The evidence covers information seeking and multi-step mathematical reasoning rather than all agentic workflows.Coding, embodied, and planning agents remain future-work settings.
- NIS-Agent can still fail if its filter discards relevant pages or its validation module endorses flawed intermediate conclusions.The authors characterize it as reducing rather than eliminating self-reinforcing errors.
9 Ethics Statement
The paper describes publicly available, pre-anonymized datasets and states that IBIS contains neither identifying information nor offensive content. It also documents the search pipeline’s behavior when prior action history is included.
- The study’s datasets are publicly available and pre-anonymized, and IBIS contains no identifying information or offensive content.The authors state that they manually verified the data for ethical and privacy compliance.
- LLMs were used only for language polishing and grammatical improvements, while researchers authored and verified the core content.
- The pipeline ideally stops when all retrieved results are irrelevant, but full action history can make the model continue selecting irrelevant URLs.
- The search procedure consists of query generation followed by result retrieval and an ideal stop decision when results are irrelevant.
A.2 Quantitative Analysis of Inertia Bias.
The quantitative analysis measures how inertia bias produces search and contextual noise in GAIA and examines controlled evidence for action-history effects. It also distinguishes inertia bias from adjacent bias categories and evaluates the IBIS design.
- Quantitative Analysis: The GAIA analysis used smolagents with GPT-4.1 and manually measured errors induced by search noise and contextual noise.
- Impact of Search Noise: 24.0% of total searches were affected by search noise, while 13.5% of sub-task conclusions were adversely affected despite possible later recovery.
- Impact of Contextual Noise: Contextual noise caused incorrect answers through misinterpretation of complete observations or premature termination with partial observations.
- Conceptual Distinctions: Inertia bias differs from sycophancy and confirmation bias because it anchors the agent to its own action trajectory rather than external preferences or propositions.
- IBIS Benchmark: IBIS holds task text and search results fixed while varying whether the model owns the prior search action, isolating action-history ownership.
- IBIS Benchmark: 91.9% of observed mode flips followed the inertia direction: correct re-searching in Observer Mode but path persistence in Agentic Mode.
B.1 Details of IBIS Experiments
The IBIS experiments compare models’ next-step decisions under controlled evaluation modes and analyze how context isolation affects relevance filtering. Results show distinct model tendencies, validate Direct Mode, and report lower runtime for NIS-Agent than smolagents DR.
- IBIS Results: 4.08% accuracy in Agentic Mode was achieved by GPT-4o on Should Re-search, reflecting its strong preference for visiting pages after initial queries.Claude-3.7-Sonnet was more inclined to call web_search to refine results.
- IBIS Results: Agentic Mode outperformed Observer Mode on Should Visit Page, indicating that self-authored query history increased the tendency to call visit_page.The passage attributes this pattern to commitment to the generated query rather than superior reasoning.
- Context-Isolated Evaluation: Direct Mode matched Agentic Mode’s Should Visit Page accuracy while reducing irrelevant webpage intake relative to Agentic Mode.Direct Mode isolates the current query and candidate webpages from action history.
- Runtime: NIS-Agent’s average runtime was lower than smolagents DR, with increasingly pronounced savings as task difficulty increased.Figure 9 reports average runtime per task.
C.4.4 Qualitative Failure Case Analysis
The failure analysis identifies incorrect filtering of weakly signaled pages as a recurring isolation-specific error, while many other failures arise from reasoning, search depth, page length, or evidence-access demands. The section also describes rubric-based training and controlled evaluation modes for improving next-action judgments.
- Qualitative Failure Cases: Weakly relevant abstracts can cause the context-isolated filter to discard pages whose decisive evidence appears only in the page body.A cited example involves a figure buried in a company annual report while the snippet shows only its title and publication date.
- Qualitative Failure Cases: A single filtering error does not always cause failure because iterative searching can recover the same evidence from another page.The ManagerAgent can also override unreasonable validation suggestions because it retains full context.
- Qualitative Failure Cases: Most remaining failures are largely independent of context isolation and include insufficient backbone reasoning, exhausted search-hop budgets, failed localization on long pages, and target-page dependencies.These categories describe boundaries beyond the isolation-specific error pattern.
- NIS-8B Training: Stage 1 creates an offline reference rubric containing the correct next-action category, acceptable evidence or queries, disqualifying failure modes, and grading instructions.The rubric is generated once per prompt and stored as fixed metadata.
- NIS-8B Training: Stage 2 scores each rollout using only the question, prior query, rubric, and chosen action, without exposing the cheap judge to raw search results.The action class is primary, while parameter quality supplies a ternary secondary judgment.