Agentic Parallel Thinking for Deep Information Seeking

ACL ARR 2026 January Submission2340 Authors

Keywords: Information Seeking, Open-Domain Question Answering, LLM Agents, Parallel Thinking

Abstract:

Parallel thinking broadens exploration and complements deep information-seeking (IS) agents, but in this setting it is hindered by redundant from-scratch rollouts and context-limited answer generation that cannot reliably integrate long-horizon trajectories.
To address these issues, we propose ParallelMuse, a two-stage inference-only paradigm designed for deep IS agents. The first stage, Functionality-Specified Partial Rollout, partitions generated sequences into functional regions and performs uncertainty-guided path reuse and branching to enhance exploration efficiency. The second stage, Compressed Reasoning Aggregation, exploits reasoning redundancy to losslessly compress information relevant to answer derivation and synthesize a coherent final answer. Experiments across multiple open-source agents and benchmarks demonstrate up to 62% performance improvement with a 10--30% reduction in exploratory token consumption.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: open-domain QA, multihop QA, reasoning

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 2340