Evaluating Stochasticity in Deep Research Agents

Haotian Zhai; Elias Stengel-Eskin; Pratik Patil; Liu Leqi

cs.AIFeb 26, 2026

Evaluating Stochasticity in Deep Research Agents

Authors: Haotian Zhai, Elias Stengel-Eskin, Pratik Patil, Liu Leqi

Organizations: University of Texas at Austin∗

Abstract

Deep Research Agents (DRAs) are promising agentic systems that gather and synthesize information to support research across domains such as financial decision-making, medical analysis, and scientific discovery. Despite recent improvements in research quality (e.g., outcome accuracy when ground truth is available), DRA system design often overlooks a critical barrier to real-world deployment: stochasticity. Under identical queries, repeated executions of DRAs can exhibit substantial variability in terms of research outcome, findings, and citations. In this paper, we formalize the study of stochasticity in DRAs by modeling them as information acquisition Markov Decision Processes. We introduce an evaluation framework that quantifies variance in the system and identify three sources of it: information acquisition, information compression, and inference. Through controlled experiments, we investigate how stochasticity from these modules across different decision steps influences the variance of DRA outputs. Our results show that reducing stochasticity can improve research output quality, with inference and early-stage stochasticity contributing the most to DRA output variance. Based on these findings, we propose strategies for mitigating stochasticity while maintaining output quality via structured output and ensemble-based query generation. Our experiments on DeepSearchQA show that our proposed mitigation methods reduce average stochasticity by 22% while maintaining high research quality.

Figures & tables

Figure 1: Overview of the evaluation pipeline. The process begins with a user question which triggers multiple independent Deep Research Agent (DRA) runs (in this example we use number of runs $k=2$ ). The resulting reports are decomposed into answers, findings, and citations, and then clustered. After clustering, each report’s answers, findings, and citations are mapped to binary vectors and normalized to compute the Total Variance (TV) as a measure of stochasticity. We expand findings as an example to illustrate the whole process. In reality, answers and citations are also processed in similar ways as in Section 4 .

$\lambda$	Step	Module	$\widehat{\text{TV}}(\widetilde{\mathbf{X}})$			$\sqrt{\widehat{\text{TV}}(\\|\mathbf{X}\\|_{0})}$		Mean Count ( $\mu$ )		Acc.
$\lambda$	Step	Module	Ans.	Find.	Cit.	Find.	Cit.	Find.	Cit.	Acc.
0.5	1	Query ( $\pi_{\text{query}}$ )	0.37	0.76	0.42	29.97	2.10	90.82	6.62	0.40
		Sum ( $\pi_{\text{sum}}$ )	0.58	0.84	0.52	32.01	2.17	90.84	6.46	0.40
		Update ( $\pi_{\text{update}}$ )	0.59	0.85	0.55	31.45	2.49	88.80	6.96	0.52
	2	Query	0.36	0.70	0.30	27.35	2.30	92.46	5.96	0.46
		Sum	0.36	0.65	0.38	30.85	2.60	88.14	6.48	0.44

Table 1: Full Ablation Results. A comprehensive breakdown of stochasticity metrics across varying temperatures (

\lambda

), temporal injection points (Steps), and specific modules. This table corresponds to the experimental decomposition analyzed in Section 5.1 .

Panel of Figure 2 — Figure 2: Comprehensive Analysis of Stochasticity Behavior. (a) Early-stage injections dominate propagation. (b) Strong positive correlations across answer, finding, and citation TV. (c) Higher sampling temperature increases total variance. (d) The Update module contributes the largest variance.

$\lambda$	Step	Module	$\widehat{\mathrm{TV}}(\mathbf{B})$	Acc.
0.5	Step 1	Query	0.76	0.40
1.0	Step 1	Query	0.80	0.36
0.5	Combined	Update	0.89	0.41
1.0	Combined	Update	0.88	0.56
0.5	Step 1	Sum	0.84	0.40
0.5	Step 3	Sum	0.52	0.34

Table 2: Examples showing that higher stochasticity does not monotonically improve accuracy. We select examples that have each one of

\lambda

, step, and module different.

Metric	Findings ( $\mathbf{B}$ )	Citations ( $\mathbf{C}$ )
$\widehat{\mathrm{TV}}(\widetilde{\mathbf{X}})$	0.76	0.44

Table 3: Average stochasticity of findings vs. citations across modules, steps, and temperatures.

Method	Acc.	$\widehat{\mathrm{TV}}(\mathbf{Y})$	$\widehat{\mathrm{TV}}(\mathbf{B})$	$\widehat{\mathrm{TV}}(\mathbf{C})$	Avg. TV
Baseline	0.24	0.62	0.83	0.62	0.69
Struc. Sum.	0.28	0.58	0.80	0.64	0.67
Struc. Update	0.32	0.52	0.75	0.58	0.62
Struc. Comb.	0.36	0.44	0.68	0.56	0.56
Quer. Int.	0.32	0.50	0.74	0.61	0.62
Comb.	0.36	0.38	0.61	0.43	0.47

Table 4: Comparison of Stochasticity Mitigation Strategies. The best performance in each category is highlighted in bold. Struc. Comb. applies structured summarization and reasoning simultaneously; Comb. integrates all three mitigation methods. Avg. TV represents the average value of the three-level stochasticity.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Metric	$T=0$	$\lambda=0.5$	$\lambda=1.0$
$\widehat{\mathrm{TV}}(\mathbf{Y})$	0.700	0.670	0.620
$\widehat{\mathrm{TV}}(\mathbf{B})$	0.828	0.836	0.834
$\widehat{\mathrm{TV}}(\mathbf{C})$	0.621	0.608	0.612
Accuracy	0.28	0.24	0.24

Appendix

Table 5: Stochasticity and Accuracy Metrics across Sampling Temperatures (

\lambda

) under API Setting.

Deep Research Agents

Deep research agents plan long-horizon searches, synthesize cited evidence, and produce reports for complex open-ended questions.

See all 115 papers

DuMate-DeepResearch: An Auditable Multi-Agent System with Recursive Search and Rubric-Grounded Reasoning
Jun 5, 2026Lingyong Yan, Can Xu, Yukun Zhao +13Deep Research Agents Agentic Search
DuMate Team, Baidu AI Cloud
Multi-Turn Evaluation of Deep Research Agents Under Process-Level Feedback
Jun 8, 2026Rishabh Sabharwal, Hongru Wang, Amos Storkey +1Deep Research Agents AI Agent Evaluation
School of Informatics, University of Edinburgh, United Kingdom · Huawei Technologies Co., Ltd., Edinburgh, United Kingdom
Why Your Deep Research Agent Fails? On Hallucination Evaluation in Full Research Trajectory
Jan 30, 2026Yuhao Zhan, Tianyu Fan, Linxuan Huang +2Deep Research Agents Hallucination Detection
Zhejiang University · The University of Hong Kong

Jun 5, 2026J/K move · Enter open · S save

DuMate-DeepResearch: An Auditable Multi-Agent System with Recursive Search and Rubric-Grounded Reasoning

Lingyong Yan, Can Xu, Yukun Zhao, Wenxuan Li+12

DuMate Team, Baidu AI Cloud

Deep Research Agents Agentic Search Multi-Agent Collaboration Multi-Agent Planning

Open paper AskPDF arXiv

Deep Research (DR) has emerged as a new agentic paradigm to tackle complex, open-ended research tasks, demanding systems that can iteratively frame problems, acquire evidence, verify sources, and synthesize long-form reports. In practice, however, current DR systems are constrained by four interrelated limitations: long-horizon planning over an underspecified scope, the bottleneck of decomposing and scheduling such tasks within a single agent, hallucination risk in long-form synthesis, and limited process auditability. This technical report presents DuMate-DeepResearch, a multi-agent DR framework built on the Qianfan Agent Foundry. The framework decouples the Agent Core, which handles task understanding, planning, and scheduling, from an extensible Tool Ecosystem for retrieval, evidence acquisition, and report rendering, making every intermediate decision and tool invocation explicitly traceable. Building on this infrastructure, DuMate-DeepResearch further introduces three mechanisms: (i) a graph-based dynamic planning strategy expands the research roadmap coarse-to-fine and continuously revises it through reflection, re-planning, backtracking, and parallel branching; (ii) a recursive two-level execution design delegates each complex search sub-task to an inner Search Agent that runs its own planning loop, isolating noisy retrieval and stabilizing long-horizon execution; (iii) a rubric-based test-time optimization mechanism dynamically generates task-specific quality criteria and uses them as live reasoning scaffolds for evidence-grounded synthesis and adaptive stopping. Across two deep research benchmarks, DuMate-DeepResearch establishes new state-of-the-art results: the best overall score (58.03%) on DeepResearch Bench, and the best overall score (61.95%) on DeepResearch Bench II while ranking first in information recall and analysis.

Abstract

Deep Research Agents

Explore similar work

DuMate-DeepResearch: An Auditable Multi-Agent System with Recursive Search and Rubric-Grounded Reasoning

Multi-Turn Evaluation of Deep Research Agents Under Process-Level Feedback

Why Your Deep Research Agent Fails? On Hallucination Evaluation in Full Research Trajectory