PaSaMaster: AI system outperforms GPT-5.2 at 1% cost with zero hallucination
New self-evolving retrieval system beats GPT-5.2 by 30% while costing 99% less.
A team of researchers from multiple institutions have unveiled PaSaMaster, a self-evolving agentic literature retrieval system designed to overcome the dual challenges of source authenticity and deep comprehension in academic search. Traditional keyword-based retrieval fails to capture complex research intents, while generative LLMs suffer from high costs and hallucination rates up to 37.79%. PaSaMaster transforms retrieval from a one-shot query-document matching into an iterative process of intent analysis, retrieval, and ranking. It uses a frontier LLM only for high-level intent understanding, while large-scale retrieval and relevance scoring are handled by customized corpora and lightweight models, dramatically cutting costs. The system produces relevance-scored paper rankings with evidence-grounded recommendations.
Evaluated on the PaSaMaster Benchmark across 38 scientific disciplines, the system exposed severe inaccuracies in traditional keyword retrieval (improving F1-score by 15.6X) and the unreliability of generative LLMs. Remarkably, PaSaMaster outperforms GPT-5.2 by 30.0% at just 1% of the computational cost while ensuring zero source hallucination. This represents a major leap for researchers needing accurate, cost-effective literature search without fabricated references. The system is open-sourced, with code and data available online.
- PaSaMaster uses a two-tier architecture: a frontier LLM for intent understanding and lightweight models for retrieval/scoring, cutting compute costs by 99%.
- It outperforms GPT-5.2 by 30% on a 38-discipline benchmark while eliminating all source hallucination.
- The system improves F1-score by 15.6x over traditional keyword-based retrieval methods.
Why It Matters
PaSaMaster makes high-quality, hallucination-free literature search affordable, potentially accelerating research across all scientific fields.