IntrAgent: An LLM Agent for Content-Grounded Information Retrieval through Literature Review

Technical Summary: IntrAgent: An LLM Agent for Content-Grounded Information Retrieval through Literature Review

1. Problem Statement

Scientific research requires the precise extraction of metadata, experimental parameters, and contextual knowledge from complex literature. Existing approaches face three primary hurdles:

• Structural Complexity: Scientific papers are not flat text; they possess hierarchical structures (sections, subsections) that are crucial for locating information.
• Limitations of RAG: Standard Retrieval-Augmented Generation (RAG) relies on semantic similarity of text chunks. This often fails in scientific contexts because it ignores document structure and struggles to align domain-specific terminology with relevant content.
• Hallucination and Scope: General-purpose scientific agents often search external resources, whereas researchers frequently need to retrieve information strictly grounded in a specific provided paper. Standard LLMs also struggle to manage the overwhelming volume of a full paper, often leading to "hallucinations" when information is missing.

To address these gaps, the authors define a new task: IntraView (INformation reTRieval through literAture reVIEW), which requires extracting and synthesizing information faithfully grounded in a provided document.

2. Methodology: The IntrAgent Framework

The authors propose IntrAgent, an LLM-based agent designed using "mindset bionics" to mimic how humans read technical literature. It operates via a two-stage pipeline:

Stage 1: Section Ranking (Structure-Aware Reasoning)
Instead of treating the paper as a collection of independent chunks, IntrAgent:

• Hierarchy Preservation: Converts the paper into a Markdown format and constructs a "section tree" to capture the parent-child relationships between headings.
• Reasoning-Based Ranking: Uses an LLM to reason about which sections (based on their titles and hierarchical position) are most likely to contain the answer to a specific query. It then reorders the sections by estimated relevance.

Stage 2: Iterative Reading (Progressive Accumulation)
The agent traverses the reordered sections sequentially:

• Section Detail Extraction: The agent examines the current section to extract specific scientific details (numerical data, terminology, etc.) and stores them in short-term memory.
• Information Sufficiency Check: After each extraction, the agent performs a reasoning step to ask: "Is the gathered information sufficient to answer the query?" If YES, it terminates; if NO, it moves to the next most relevant section. This prevents premature termination and reduces hallucination by ensuring answers are substantiated.
• Final Synthesis: Once sufficiency is met, the agent synthesizes the accumulated details into a concise, grounded answer.

3. Key Contributions

• New Task (IntraView): A specialized task for fine-grained, content-grounded information retrieval from full scientific papers.
• New Benchmark (IntraBench): A rigorous evaluation set consisting of 315 expert-authored test instances across five STEM domains: Physics, Earth Science, Public Health, Engineering, and Material Science.
• Novel Agent Architecture: The introduction of the two-stage (Ranking + Iterative Reading) pipeline with hierarchy preservation and sufficiency checking.
• Robustness Mechanisms: Implementation of a "confidence-based mechanism" (conservative, balanced, and aggressive reading styles) to allow users to balance accuracy against computational overhead.

4. Results

The authors evaluated IntrAgent against state-of-the-art RAG systems and literature-oriented agents (e.g., PaperQA2, SciMaster) using seven backbone LLMs (including GPT-4o, DeepSeek-R1, and Gemini-2.5 Pro).

• Superior Accuracy: IntrAgent achieved an average cross-domain accuracy 13.2% higher than the strongest baselines.
• Ablation Insights:
  • Removing Hierarchy Preservation significantly dropped accuracy, proving structural knowledge is vital.
  • Removing the Sufficiency Check led to massive accuracy drops and increased hallucinations (e.g., the agent fabricating values when the check was disabled).
• Robustness: The agent maintained high performance even when section headings were intentionally corrupted (e.g., "Shakespearean" or "noisy" styles) and showed no significant performance degradation regardless of paper length.

5. Significance

IntrAgent represents a shift from "retrieve-then-generate" (RAG) to "reason-then-read" (Agentic) workflows for scientific discovery. By mimicking human cognitive processes—prioritizing structure and iteratively verifying information—it provides a more reliable and faithful method for automating literature reviews. This has profound implications for accelerating research workflows in highly specialized scientific disciplines where precision is non-negotiable.