Towards Self-Evolving Agentic Literature Retrieval

Abstract: As LLMs reshape scientific research, literature retrieval faces a twofold challenge: ensuring source authenticity while maintaining a deep comprehension of academic search intents. While reliable, traditional keyword-centric search fails to capture complex research intents. Frontier LLMs can handle complex research intents, but their high cost and tendency to hallucinate remain key limitations. Here we introduce PaSaMaster, a self-evolving agentic literature retrieval system that produces relevance-scored paper rankings with evidence-grounded recommendations through iterative intent analysis, retrieval, and ranking. It is built on three key designs. First, it transforms literature retrieval from a one shot query--document matching problem into a search process that evolves over time, using ranked evidence to reveal gaps, refine intents, and guide follow-up searches. Second, it prevents hallucinated sources by treating retrieval as intent--paper relevance ranking rather than generation. Finally, PaSaMaster improves cost efficiency by separating planning from retrieval: a frontier LLM is used only for intent understanding, while large scale retrieval and relevance scoring are delegated to customized corpora and lightweight models. Evaluated on the PaSaMaster Benchmark across 38 scientific disciplines, our system exposes the severe inaccuracy and incompleteness of traditional keyword retrieval (improving F1-score by 15.6X) and the unreliability of generative LLMs (which exhibit hallucination rates up to 37.79%). Remarkably, PaSaMaster outperforms GPT-5.2 by 30.0% at a mere 1% of the computational cost while ensuring zero source hallucination: https://github.com/sjtu-sai-agents/PaSaMaster

• The paper introduces a self-evolving agentic system that employs iterative intent refinement with Navigator and Librarian agents for adaptive literature search.
• It ensures evidence-grounded relevance ranking and zero source hallucination using a three-tier repository and domain-specific scorers.
• The cost-efficient design decouples high-level intent analysis from scalable retrieval, achieving 30% higher F1-score at 1% query cost compared to leading methods.

Self-Evolving Agentic Paradigm for Literature Retrieval: An Expert Summary of PaSaMaster

Background and Motivation

Automated scientific literature retrieval is a critical bottleneck in the research workflow, particularly as the growth of the scientific corpus vastly outpaces human cognitive bandwidth. While scalable retrieval remains essential, persistent trade-offs hamper current approaches: traditional keyword-based and semantic retrieval guarantee source authenticity but collapse complex intents into shallow representations, while LLM-powered generative methods improve semantic comprehension but hallucinate sources and do not scale efficiently. Recent systems employing agentic retrieval models have attempted to bridge this gap, but rigid, fixed-intent pipelines still limit adaptive search. There is a consequential need for systems that realize deep intent understanding, maintain strict source verifiability, and do not incur unsustainable computational costs.

Core Contributions of PaSaMaster

The paper introduces PaSaMaster, a self-evolving agentic literature retrieval architecture that systematically addresses the limitations of prior paradigms. Its hallmark innovations are:

1. Self-Evolving Search Process: Literature retrieval is operationalized as an iterative, closed-loop process where initial output informs subsequent intent refinement and search direction. This enables dynamic coverage gap identification, constraint disambiguation, and intent re-specification, realized via the Navigator (planning) and Librarian (execution) agent modules.
2. Hallucination-Free, Evidence-Grounded Relevance Ranking: Retrieval is not dependent on generative synthesis from parametric LLMs. Instead, evidential grounding is anchored on a three-tier agent-native repository—metadata, abstract-level, and evidence chunk—ensuring only papers verifiably present in indexed corpora are returned. Paper selection and relevance scoring are strictly evidence-based, enforced by lightweight, domain-knowledge distilled Scorers.
3. Cost-Efficient Planning-Retrieval Separation: High-cost, high-reasoning LLMs are exclusively engaged for complex, adaptive intent understanding (Navigator) and not for large-scale retrieval, evidence extraction, or relevance attribution, which are offloaded to scalable, parallel Librarian agents and optimized Scorers. This minimizes computational overhead without compromising the granularity of reasoning or evidence evaluation.

PaSaMaster-Bench: Benchmarking Complex Literature Retrieval

To evaluate intent interpretability, evidentiary precision, and cost trade-offs at scale, the paper introduces PaSaMaster-Bench, a multidisciplinary benchmark dataset. It features 244 expert-curated tasks spanning 38 scientific disciplines, each comprising multi-constraint natural language queries with strict ground-truth criteria and checklist-style verification. Importantly, PaSaMaster-Bench is constructed to probe full academic research intent, not simple keyword expansion, distinguishing it from prior benchmarks.

Empirical Results

PaSaMaster demonstrates clear empirical superiority over established retrieval paradigms. Key findings:

• Retrieval Quality: On PaSaMaster-Bench, PaSaMaster achieves NDCG@20 of 37.93, Recall@20 of 31.84, Precision@20 of 22.19, and F1-score@20 of 21.69. Compared with Google Scholar, F1-score improves 15.6× (from 1.39 to 21.69), and outperforming the strongest generative LLM baseline (GLM-5; F1-score 18.18) by 19.3% and the best fixed-pipeline agent (Google Scholar Labs; F1-score 18.87) by 14.9%.
• Source Hallucination: PaSaMaster achieves zero hallucination, while generative LLMs such as MiniMax-M2.7, Kimi-K2.5, and Gemini-3.1-pro exhibit rates of 35–38%. Even GPT-5.2 is susceptible (11.8% hallucinated sources). This robustly substantiates the benefit of explicit evidence-constrained retrieval versus parametric citation generation.
• Cost Efficiency: Query cost is drastically reduced—PaSaMaster operates at $0.05 per query, or 1% of GPT-5.2’s$6.06/query budget, achieving a 30% higher F1-score. There is no discernable quality sacrifice despite resource savings, validating the planning–retrieval decoupling strategy.
• Cross-Disciplinary Consistency: Performance gains persist across all 38 evaluated scientific areas, indicating strong generality and domain-independence.

Implications and Future Directions

PaSaMaster epitomizes a paradigm shift from static, shallow, or generative retrieval methods towards adaptive agentic architectures explicitly structured for iterative intent revision, evidence-grounded verification, and cost-scaled deployment. This approach carries meaningful practical and theoretical implications:

• Research Acceleration: The agentic framework offers a robust mechanism for AI-driven literature discovery at the scale required by modern scientific practice, directly impacting hypothesis generation, systematic review, and exploration in high-velocity knowledge domains.
• AI Reliability: Eliminating source hallucination is non-negotiable in scientific research; the approach demonstrated here sets a new reliability baseline for AI-augmented search.
• Scalable Intellectual Automation: By decoupling high-level abstraction and low-level search, PaSaMaster presents a scalable blueprint for further agentic systems targeting other information-dense workflows, such as automated review or meta-analysis synthesis.
• Foundation for Next-Generation Retrieval Systems: The self-evolving, agentic paradigm may extend naturally to support richer forms of intent expression (e.g., conditional logic, dynamic knowledge graphs), more granular evidence integration (full-text extraction, citation context summarization), or scientific discovery with evolving constraints and feedback via human–AI interaction.

Conclusion

PaSaMaster offers a systematically validated architecture for cost-effective, hallucination-free, adaptive literature retrieval that attains high precision and recall in complex, multi-constraint scientific search scenarios. Positioned as the prototype of self-evolving agentic retrieval, it addresses longstanding limitations of both traditional and LLM-based paradigms, and lays the groundwork for practical, reliable, and efficient AI instrumentation across scientific domains.

(2605.14306)