ParallelMuse: Agentic Parallel Thinking for Deep Information Seeking

TL;DR

ParallelMuse introduces a two-stage paradigm that enhances deep information-seeking agents by improving exploration efficiency and integrating long-horizon reasoning, leading to significant performance gains and reduced token consumption.

Contribution

It proposes ParallelMuse, a novel two-stage framework that addresses efficiency and reasoning integration challenges in parallel thinking for deep IS agents.

Findings

Up to 62% performance improvement.

10-30% reduction in exploratory token consumption.

Effective across multiple agents and benchmarks.

Abstract

Parallel thinking expands exploration breadth, complementing the deep exploration of information-seeking (IS) agents to further enhance problem-solving capability. However, conventional parallel thinking faces two key challenges in this setting: inefficiency from repeatedly rolling out from scratch, and difficulty in integrating long-horizon reasoning trajectories during answer generation, as limited context capacity prevents full consideration of the reasoning process. To address these issues, we propose ParallelMuse, a two-stage 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.