IntrAgent/IntraBench

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

Fengbo Ma, Zixin Rao, Xiaoting Li, Zhetao Chen, Hongyue Sun, Yiping Zhao, Xianyan Chen, Zhen Xiang
University of Georgia, Athens, GA, USA

Key Contributions

IntrAgent introduces IntraView, a new task for content-grounded information retrieval from a provided scientific paper, and proposes an LLM agent that mimics human literature reading through structure-aware section ranking and iterative evidence gathering. To evaluate this setting, the work presents IntraBench, a 315-instance benchmark across five STEM domains on which IntrAgent improves average cross-domain accuracy by 13.2% over strong RAG and literature-agent baselines.

• Defines IntraView, a new task for faithful information retrieval from a provided scientific paper rather than from external search.
• Introduces IntrAgent, a specialized LLM agent that follows a human-like workflow: identify promising sections, extract evidence, and stop when support is sufficient.
• Develops two core mechanisms for the task: hierarchy-preserving section ranking and iterative reading with information sufficiency checking.
• Builds IntraBench, the first benchmark for this setting, covering 315 instances across five impactful STEM domains.

IntraView Task

IntraView is formulated as a content question answering problem over a full scientific paper. Given a literature document C and a research-driven query Q, the system must return an answer A that is accurate, concise, and explicitly grounded in the provided paper.

Compared with standard content QA, the task is harder because scientific papers are long, structurally complex, and filled with domain-specific terminology. The relevant evidence may appear anywhere in the document, may require cross-referencing multiple sections, and may sometimes be absent entirely, making hallucination control central to the task.

IntrAgent Method

Stage 1: Section Ranking

IntrAgent first parses section titles and preserves the paper hierarchy so the model can reason over the document as a structured artifact rather than a flat list of chunks. The LLM then ranks sections by likely relevance to the question, producing a reordered reading path.

This hierarchy-aware step is designed to better align scientific questions with the parts of a paper most likely to contain supporting evidence.

Stage 2: Iterative Reading

The agent reads the ranked sections sequentially, extracts anchored details such as terminology, measurements, results, and comparisons, and stores them in short-term memory for answer synthesis.

After each reading step, IntrAgent performs an explicit information sufficiency check. If the evidence is still incomplete, it continues reading; otherwise it stops and synthesizes a grounded answer.

Figure 1. Overview of the IntrAgent pipeline containing two stages: Section Ranking (top) reorders the paper’s sections by relevance to the Research Question Q, while Iterative Reading (bottom) steps through ranked sections, extracting information until gathered information is sufficient.

IntraBench Benchmark

To evaluate IntraView, the paper introduces IntraBench, the first benchmark specifically designed for literature-grounded information retrieval. It contains 315 test instances derived from expert-authored questions paired with research papers.

The benchmark spans five high-impact domains and is intended to capture technical depth, conceptual complexity, and domain-specific phrasing encountered in real literature review workflows.

Domains: Physics, Earth Science, Public Health, Engineering, Material Science

Questions are organized around four research-oriented categories described in the paper: study subject and experimental setup, data characteristics and collection, technical approach and details, and conclusions and results. Evaluation is performed through LLM-grounded multiple-choice mapping to handle synonyms, abbreviations, and scientific terminology variation.

Figure 2. An example of IntrAgent executing a question-paper pair from IntraBench. For an input question Q regarding paper C, vanilla RAG fails to extract the correct chunk, resulting in an incorrect answer. In contrast, IntrAgent ranks the sections through reasoning and retrieves the correct details that pass the sufficiency check in the first iteration, leading to a correct answer. Details for section ranking and sufficiency check are also presented.

Benchmark Papers Used in IntraBench

Domain	Title
Public Health - Infectious-disease Modeling	Mathematical modeling and analysis of COVID-19: A study of new variant Omicron
Public Health - Infectious-disease Modeling	COVID-19 pandemic in India: a mathematical model study
Public Health - Infectious-disease Modeling	A mathematical COVID-19 model considering asymptomatic and symptomatic classes with waning immunity
Public Health - Infectious-disease Modeling	Mathematical modeling and analysis of COVID-19 pandemic in Nigeria
Public Health - Infectious-disease Modeling	Mathematical modeling of COVID-19 transmission dynamics with a case study of Wuhan
Physics - Surface Enhanced Raman Spectroscopy	Quantification of Analyte Concentration in the Single Molecule Regime Using Convolutional Neural Networks
Physics - Surface Enhanced Raman Spectroscopy	Machine learning enabled multiplex detection of periodontal pathogens by surface-enhanced Raman spectroscopy
Physics - Surface Enhanced Raman Spectroscopy	Rapid Detection of SARS-CoV-2 Variants Using an Angiotensin-Converting Enzyme 2-Based Surface-Enhanced Raman Spectroscopy Sensor Enhanced by CoVari Deep Learning Algorithms
Physics - Surface Enhanced Raman Spectroscopy	Rapid Detection of SARS-CoV-2 RNA in Human Nasopharyngeal Specimens Using Surface-Enhanced Raman Spectroscopy and Deep Learning Algorithms
Physics - Surface Enhanced Raman Spectroscopy	Quantitative detection of α1-acid glycoprotein (AGP) level in blood plasma using SERS and CNN transfer learning approach
Earth Science - Remote Sensing	Annual maps of global artificial impervious area (GAIA) between 1985 and 2018
Earth Science - Remote Sensing	Annual dynamics of global land cover and its long-term changes from 1982 to 2015
Earth Science - Remote Sensing	Finer resolution observation and monitoring of global land cover: first mapping results with Landsat TM and ETM+ data
Earth Science - Remote Sensing	Mapping Essential Urban Land Use Categories in Beijing with a Fast Area of Interest (AOI)-Based Method
Earth Science - Remote Sensing	Mapping essential urban land use categories in China (EULUC-China): preliminary results for 2018
Engineering - Human Factor	A data analytic end-to-end framework for the automated quantification of ergonomic risk factors across multiple tasks using a single wearable sensor
Engineering - Human Factor	Assessing human situation awareness reliability considering fatigue and mood using EEG data: A Bayesian neural network-Bayesian network approach
Engineering - Human Factor	Automatic driver cognitive fatigue detection based on upper body posture variations
Engineering - Human Factor	Enhancing Data Privacy in Human Factors Studies with Federated Learning
Engineering - Human Factor	Worker’s physical fatigue classification using neural networks
Material Science - Additive Manufacturing	Autonomous optimization of process parameters and in-situ anomaly detection in aerosol jet printing by an integrated machine learning approach
Material Science - Additive Manufacturing	Geometrical defect detection for additive manufacturing with machine learning models
Material Science - Additive Manufacturing	Layer-Wise Modeling and Anomaly Detection for LaserBased Additive Manufacturing
Material Science - Additive Manufacturing	Online droplet anomaly detection from streaming videos in inkjet printing
Material Science - Additive Manufacturing	Toward the digital twin of additive manufacturing- Integrating thermal simulations, sensing, and analytics to detect process faults

Experiments and Results

The experiments compare IntrAgent against a broad set of RAG-based retrieval systems and literature-oriented agents, including vanilla RAG variants, contextual RAG, DRAGIN, R2AG, LongRAG, LUMOS, PaperQA2, Agentic-Hybrid-RAG, and SciMaster.

On IntraBench, IntrAgent sets a new state of the art across all five domains and seven backbone LLMs. Reported average accuracies include 70.0% with GPT-4o, 75.8% with GPT-4.1, 74.4% with DeepSeek-R1, 73.4% with o3, 73.8% with o4-mini, 75.9% with Gemini-2.5 Pro, and 68.8% with Llama-3.1-70B.

The paper attributes these gains to two main design choices: hierarchy-aware section ranking and the sufficiency check that stops reading once evidence is complete. In contrast, flat RAG pipelines often inject irrelevant chunks, while literature agents designed for online search degrade into static retrieval pipelines when constrained to a provided paper.

Cross-Domain Accuracy on IntraBench

Group	Method	GPT-4o	GPT-4.1	DS-R1	o3	o4-mini	Gemini-2.5 Pro	Llama-3.1-70B
RAG	Vanilla RAG all-MiniLM-L6-v2	60.3	61.2	64.3	60.4	61.5	61.8	59.2
RAG	Vanilla RAG E5-mistral-7b-instruct	59.4	64.2	63.8	60.3	61.4	59.9	60.5
RAG	Vanilla RAG GritLM-7B	60.4	63.2	63.2	59.7	58.4	58.4	61.4
RAG	Context. RAG E5-mistral-7b-instruct	60.7	63.8	62.8	59.1	58.3	58.9	58.9
RAG	Context. RAG GritLM-7B	60.8	62.8	61.6	58.4	60.7	61.6	59.2
RAG	DRAGIN	42.5	44.6	46.9	44.0	46.9	45.9	45.4
RAG	R²AG	59.4	59.5	61.5	56.6	55.3	55.6	56.1
RAG	LongRAG	62.1	64.7	65.5	57.0	58.3	57.1	57.4
Agent	LUMOS	50.2	52.1	55.4	55.2	56.4	54.9	54.4
Agent	PaperQA2	47.7	48.9	54.0	51.8	49.2	51.2	53.8
Agent	Agentic-Hybrid-RAG	59.8	60.2	62.3	57.5	57.8	57.2	56.6
Agent	SciMaster	59.0	57.6	63.3	57.2	58.1	57.2	57.0
Agent	IntrAgent (Ours)	70.0	75.8	74.4	73.4	73.8	75.9	68.8

Resources

• Paper: coming soon
• GitHub: https://github.com/FengboMa/IntrAgent
• Hugging Face: https://huggingface.co/datasets/IntrAgent/IntraBench

Citation

@inproceedings{ma2026intragent,
  title={IntrAgent: An {LLM} Agent for Content-Grounded Information Retrieval through Literature Review},
  author={Fengbo Ma and Zixin Rao and Xiaoting Li and Zhetao Chen and Hongyue Sun and Xianyan Chen and Yiping Zhao and Zhen Xiang},
  booktitle={The 64th Annual Meeting of the Association for Computational Linguistics},
  year={2026},
}