To Search or Not to Search: Aligning the Decision Boundary of Deep Search Agents via Causal Intervention

TL;DR

This work formalizes the problem of when a deep search agent should stop querying and start answering as a dynamic decision boundary governed by the agent's latent knowledge state. It introduces a causal intervention framework to diagnose over-search and under-search, then builds DAS to align the agent's policy using counterfactual preferences via Direct Preference Optimization. Across three QA datasets, decision boundary errors are widespread and outcome-focused RL alone trades accuracy for efficiency, while DAS achieves notable gains in both final accuracy and runtime efficiency by calibrating the boundary. The approach offers a practical pathway to more reliable and economical autonomous search systems, with code and data publicly available. This work advances the design of boundary-aware reasoning for web-scale information seeking and multi-hop QA tasks.

Abstract

Deep search agents, which autonomously iterate through multi-turn web-based reasoning, represent a promising paradigm for complex information-seeking tasks. However, current agents suffer from critical inefficiency: they conduct excessive searches as they cannot accurately judge when to stop searching and start answering. This stems from outcome-centric training that prioritize final results over the search process itself. We identify the root cause as misaligned decision boundaries, the threshold determining when accumulated information suffices to answer. This causes over-search (redundant searching despite sufficient knowledge) and under-search (premature termination yielding incorrect answers). To address these errors, we propose a comprehensive framework comprising two key components. First, we introduce causal intervention-based diagnosis that identifies boundary errors by comparing factual and counterfactual trajectories at each decision point. Second, we develop Decision Boundary Alignment for Deep Search agents (DAS), which constructs preference datasets from causal feedback and aligns policies via preference optimization. Experiments on public datasets demonstrate that decision boundary errors are pervasive across state-of-the-art agents. Our DAS method effectively calibrates these boundaries, mitigating both over-search and under-search to achieve substantial gains in accuracy and efficiency. Our code and data are publicly available at: https://github.com/Applied-Machine-Learning-Lab/WWW2026_DAS.

Figures (9)

• Figure 1: Trade-off between search depth and accuracy on HotpotQA, illustrating the diminishing gains.
• Figure 2: The overall framework of DAS. (a) Formal definition of over-search and under-search as the two primary failure modes at the agent's decision boundary. (b) Causal Intervention for decision errors diagnosis within an agent's trajectory through the use of counterfactual instruction. (c) Decision Boundary Alignment for Search Agent based on a preference dataset from identified errors.
• Figure 3: Distribution of decision boundary errors (OSR and USR) across reasoning steps on NQ, HotpotQA, and 2WikiMultihopQA (2Wiki) datasets.
• Figure 4: Entropy distributions for correct vs. incorrect decisions at the Search Step (left) and Answer Step (right).
• Figure 5: ROC curves illustrating the effectiveness of using entropy to classify correct vs. incorrect decisions for the Search Step (a) and Answer Step (b).