Agentic Search Loops for Iterative Evidence Gathering

36 concepts26 papers2 proposals

Standard Retrieval-Augmented Generation (RAG) follows a linear path: a user asks a question, the system retrieves a fixed set of documents, and the model generates an answer. While effective for simple queries, this "single-shot" approach often fails when faced with complex, multi-hop questions where the answer requires connecting disparate pieces of information.

To address this, a new paradigm has emerged that transforms Large Language Models (LLMs) from passive readers into agentic searchers. Instead of a single retrieval step, these systems employ an iterative reasoning-action loop, mimicking the way a human researcher explores a topic by searching, reading, and refining their queries based on what they find.

From Passive Retrieval to Active Research

The core limitation of traditional retrieval is the "alignment gap" between the retriever and the LLM. As Chen et al. argue in their work on the C-3PO framework, single-shot retrieval often misses the nuanced evidence needed for complex reasoning. By shifting to an iterative process, the system can review intermediate results and decide whether to continue searching, effectively mimicking human-like research behavior.

This shift is characterized by a move from simple keyword matching to multi-turn interactions. Mo et al. describe this as an agentic architecture where retrieval is treated as a multi-step planning task. The agent decomposes a high-level goal into a sequence of search operations, executing them based on feedback from the environment.

The Anatomy of an Agentic Search Loop

An agentic search loop typically consists of four recurring stages:

  1. Planning and Decomposition: The model breaks a complex query into smaller, searchable sub-questions. For example, Shi et al. propose EXSEARCH, which uses a self-incentivized loop to decompose queries into discrete actions like sub-query generation and evidence extraction.
  2. Execution and Retrieval: The agent uses tools to interact with external databases, the web, or structured knowledge graphs. In the CiteAgent system, Press et al. demonstrate how an agent can iteratively query databases and perform full-text analysis to verify scientific claims.
  3. Reasoning and Evaluation: The model assesses the retrieved information. Does it answer the sub-question? Is there a logic gap? Yang et al. introduce a "critic" module in their GraphSearch workflow that specifically looks for these gaps in drafted responses, triggering new searches if the evidence is insufficient.
  4. Refinement and Pivoting: If the information is incomplete, the model uses the new context to refine its next query. Guo et al. describe this in the DioR framework as "contextual pivoting," where keywords from initial snippets are reused to inform subsequent, more targeted searches.

The power of iterative search extends beyond text. In long-video understanding, Ma et al. show that an agent can treat a video as a document, scanning broad summaries and then "zooming in" on specific frames to resolve details. Similarly, Yu et al. demonstrate with CogPlanner that an agent can dynamically decide whether to search for text or images to answer multimodal questions.

This iterative approach is particularly vital for specialized domains. In legal research, the PAKTON framework uses a multi-agent loop (Archivist, Researcher, and Interrogator) to resolve ambiguities in long contracts, as shown by Raptopoulos et al.. For robotics, Anwar et al. use an iterative retrieval process to help robots reason over spatio-temporal memories, narrowing down search parameters when initial results are ambiguous.

Why Iteration Outperforms Scale

The primary advantage of agentic search is its efficiency under constraints. While traditional RAG systems often suffer when the retrieval budget (the number of documents retrieved) is low, Yang et al. find that agentic workflows maintain high performance by intelligently navigating to critical evidence. By isolating essential facts through targeted follow-up queries, these systems avoid the "information overload" and noise that come with large, single-batch retrievals.

Furthermore, this active navigation allows models to overcome their training cutoffs. As Zhang et al. demonstrate with the Vision Search Assistant, an agent that can iteratively decompose images into searchable queries can accurately explain events that occurred long after the model was trained, simply by knowing how to search and verify information in the wild.

Key Foundational Work

GraphSearch: An Agentic Deep Searching Workflow for Graph Retrieval-Augmented Generation (2025) — contributed 3 concept elements to this direction.
Layered Query Retrieval: An Adaptive Framework for Retrieval-Augmented Generation in Complex Question Answering for Large Language Models (2024) — contributed 2 concept elements to this direction.
PledgeTracker: A System for Monitoring the Fulfilment of Pledges (2025) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_iterative_agentic_retrieval_df034ff6,
  title = {iterative agentic retrieval},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/iterative-agentic-retrieval/rdr_df034ff6891c4fbd94db8c8d9c4c4790},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_iterative_agentic_retrieval_df034ff6_v1,
  title = {iterative agentic retrieval},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/iterative-agentic-retrieval/rdr_df034ff6891c4fbd94db8c8d9c4c4790/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake