#Deep Research Agents#Stochasticity#AI Reliability#Information Processing#Scientific Computing#Machine Learning#Research Quality#Variance Reduction
📌 Key Takeaways
Researchers identified stochasticity as a critical barrier to real-world deployment of Deep Research Agents
Three sources of stochasticity were identified: information acquisition, information compression, and inference
Reducing stochasticity improves research output quality, particularly in inference and early stages
Proposed mitigation strategies achieved 22% reduction in stochasticity while maintaining quality
📖 Full Retelling
Researchers Haotian Zhai, Elias Stengel-Eskin, Pratik Patil, and Liu Leqi published a groundbreaking paper on arXiv on February 26, 2026, addressing the critical issue of stochasticity in Deep Research Agents (DRAs) that creates a significant barrier to their real-world deployment across domains like financial decision-making, medical analysis, and scientific discovery. Deep Research Agents are sophisticated AI systems designed to gather and synthesize information to support research across various fields, yet the researchers discovered that these systems often produce inconsistent results when given identical queries, exhibiting substantial variability in research outcomes, findings, and citations. Through their formalized study modeling DRAs as information acquisition Markov Decision Processes, the team introduced an evaluation framework to quantify variance in these systems and identify the root causes of their unpredictable behavior. The researchers conducted controlled experiments to investigate how stochasticity from different modules across various decision steps influences the overall variance of DRA outputs, ultimately developing strategies to mitigate this randomness while maintaining high-quality research outcomes.
🏷️ Themes
Artificial Intelligence, Research Reliability, System Variability
Field that uses computers and mathematical models to analyze and solve scientific problems
Computational science, also known as scientific computing, technical computing or scientific computation (SC), is a division of science, and more specifically the computer sciences, which uses advanced computing capabilities to understand and solve complex physical problems in science. While this ty...
Stochastic (; from Ancient Greek στόχος (stókhos) 'aim, guess') is the property of being well-described by a random probability distribution. Stochasticity and randomness are technically distinct concepts. Stochasticity refers to a modeling approach, while randomness describes phenomena.
In cognitive psychology, information processing is an approach to the goal of understanding human thinking that treats cognition as essentially computational in nature, with the mind being the software and the brain being the hardware. It arose in the 1940s and 1950s, after World War II. The informa...
No entity connections available yet for this article.
Original Source
--> Computer Science > Artificial Intelligence arXiv:2602.23271 [Submitted on 26 Feb 2026] Title: Evaluating Stochasticity in Deep Research Agents Authors: Haotian Zhai , Elias Stengel-Eskin , Pratik Patil , Liu Leqi View a PDF of the paper titled Evaluating Stochasticity in Deep Research Agents, by Haotian Zhai and 3 other authors View PDF HTML Abstract: Deep Research Agents 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. Subjects: Artificial Intelligence (cs.AI) Cite as: arXiv:2602.23271 [cs.AI] (or arXiv:2602.23271v1 [cs.AI] for this version) https://doi.o...
