ManuSearch: An Open Multi-Agent Framework for Democratizing Deep Search in LLMs

Forum topic · 小凯 · 2026-07-05

Summary

ManuSearch is a research paper (arXiv:2505.18105, May 2025) that proposes a transparent and open multi-agent framework for deep search in large language models, authored by Lisheng Huang, Yichen Liu, Jinhao Jiang, Rongxiang Zhang, Jiahao Yan, Junyi Li and colleagues. The work addresses long-standing challenges in agentic search—including efficiency, scalability, and user intent understanding—by moving beyond traditional pipelines that treat retrieval, ranking, and generation as disconnected stages. The forum discussion contextualizes ManuSearch within the evolution from BM25 and dense retrievers to generative retrieval and LLM-based search agents, where reasoning budget and action space (whether to retrieve, how many steps, which tools to call) become learnable decision variables. The post outlines a general four-stage methodology (problem formalization, system design, training, inference pipeline), lists core components such as retrievers, rerankers, planners, and memory/tool modules, and surveys evaluation practices across datasets like MS MARCO, BEIR, and Natural Questions with metrics including nDCG@10, Recall@k, and task success rate. It also highlights open problems: evaluation trustworthiness, latency and cost, hallucination and safety, and cross-lingual/multimodal extension, making the entry a useful index point within deep-research literature.

ManuSearch: Democratizing Deep Search in Large Language Models

Paper: arXiv:2505.18105 (May 2025) Authors: Lisheng Huang, Yichen Liu, Jinhao Jiang, Rongxiang Zhang, Jiahao Yan, Junyi Li, et al. (7 authors)

Key points

ManuSearch proposes a transparent and open multi-agent framework for deep search with large language models, aimed at making agentic deep search accessible rather than proprietary.
It targets the gap between traditional pipelines (retrieval → ranking → generation handled separately) and modern needs for natural-language interaction, multi-hop reasoning, and real-time knowledge.
The framework decomposes search agents into components such as retrievers, rerankers, planners, memory modules, and tool interfaces, combined serially or in parallel.
Typical learning strategies in this paradigm include supervised fine-tuning, contrastive learning, distillation, reinforcement learning (including process rewards), and bootstrapped data synthesis; inference may use single-pass, iterative, or parallel sub-query retrieval with early stopping and budget control.

Context within agentic search

Neural IR has evolved from BM25 through dense bi-encoder retrieval and cross-encoders to late-interaction, generative retrieval, and LLM agents. Each generation trades off efficiency, effectiveness, and maintainability. In the LLM era, the new variables are reasoning budget and action space: whether to retrieve, how many retrieval steps to take, and which tools to invoke. Agentic search extends RAG from one-shot retrieval to an iterative, verifiable, plannable process, shifting evaluation from static nDCG toward task success rate, citation accuracy, and multi-hop reasoning chain completeness.

Evaluation considerations

Datasets: MS MARCO, BEIR, Natural Questions, domain-specific corpora
Metrics: nDCG@10, MRR, Recall@k, Hit@k, human preference, task success rate, latency, token cost
Baselines: BM25, dense retrieval, cross-encoder reranking, retrieval-free LLMs, commercial search APIs

Quantitative results should be verified against the original PDF.

Open problems and outlook

Trustworthy evaluation; growing gap between offline metrics and online satisfaction (LLM-as-judge needs cross-validation with human assessment)
Latency and cost as hard constraints for production deployment
Hallucination and safety risks when agents operate on the open web
Cross-lingual and multimodal extension

Future directions include more efficient test-time compute allocation, deeper integration with knowledge graphs and structured databases, and causal/fairness constraints for recommendation scenarios.