AI evidence extraction

PaperQA: Retrieval-Augmented Generative Agent for Scientific Research

Jakub Lála, Odhran O’Donoghue, Aleksandar Shtedritski, Sam Cox, Samuel G. Rodriques, Andrew Dickson White · arXiv (Cornell University) · 2023

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

2/4

Quality (LMQS)

Evidence

Citations

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.48550/arxiv.2312.07559

Methodology & findings

Study design

Computational system development with evaluation on multiple-choice question-answering benchmarks.

Main result

PaperQA outperforms competing models on scientific question answering tasks. On the modified PubMedQA dataset (closed-book), "PaperQA beats GPT-4 by 30 points (57.9% to 86.3%)". On the new LitQA dataset, "PaperQA outperforms all models tested and commercial tools, and is comparable to human experts on LitQA on performance and time". The system achieved 69.5% accuracy on LitQA compared to 66.8% for human experts, with "the lowest rate of incorrectly answered questions out of all tools, which rivals that of humans". Additionally, "no hallucinated citations were produced through PaperQA" across 237 citations tested, compared to 40-60% hallucination rates in baseline LLMs.

Research paradigm

Empirical-computational with pragmatic application focus

Author conclusions

The authors conclude: "We introduced PaperQA, a Retrieval-Augmented Generative (RAG) agent that can answer scientific questions better than other LLMs and commercial products. We found PaperQA to be more cost-efficient than humans, while still retaining its accuracy on par with human researchers." They further state: "We measured the hallucination rate of citations for recent LLMs to be between 40-60%, whereas we were not able to find a single hallucinated citation in PaperQA's responses." The authors express their hope that "this open-source implementation of a scientific question-answering system illuminates the design of future RAG agents and tools that reduce hallucinations in LLMs."

Risk of bias

Selection bias in LitQA dataset: only biomedical questions included, limiting generalizability; Human evaluator bias: only 5 biomedical researchers evaluated questions; limited diversity in human annotators; Publication bias in retrieved papers: papers accessible through public APIs may not represent full literature; LLM parametric knowledge bias: models may have latent knowledge from training data that influences performance beyond retrieval; Multiple comparisons issue: many ablations and benchmarks tested without correction; Commercial tool evaluation bias: single runs for some commercial tools limit reliability of comparisons; PDF parsing failure rates not fully characterized: acknowledged in Appendix C but impact unclear; Question design bias: LitQA questions were assembled by researchers with potential implicit biases in question formulation; Limited domain scope: Evaluation limited to biomedical domain only; Search engine selection bias: Results may depend on choice of Google Scholar as primary search engine; Multiple-choice format bias: Ablation shows significant performance drop without MC options, suggesting potential answer-keyword exploitation; Paper availability bias: Questions can only be answered if papers are accessible through available APIs; LLM parametric knowledge bias: Models may have latent knowledge despite claimed post-cutoff questions; Limited human evaluation (only 5 biomedical researchers); Knowledge cutoff effects in underlying LLMs may bias results; Selection bias in LitQA dataset composition (curated by experts); PDF parsing failures may introduce systematic retrieval errors; Search engine API differences (Google Scholar vs Semantic Scholar) may bias results

Limitations

The authors acknowledge several limitations: "We have an underlying assumption from this that the information in the underlying papers is correct, which may not hold." Additionally, "Our models and benchmarks are affected by the changing nature of science and the availability of scientific literature: some of the questions in LitQA may have new correct answers or become invalid over time." Furthermore, "While recent work has been conducted on prompt optimization, the complex setting of multiple agents with individual prompts is unsolved
Specifically, the task becomes a non-trivial, bi-level optimization problem
Consequently, discerning the impact of manual prompt adjustments becomes difficult
Thus, it is unlikely our prompts are optimal and it is difficult to assess which pieces of the prompts are necessary."

Open questions raised

Authors identify the need for better prompt optimization in multi-agent settings, noting this is an unsolved 'non-trivial, bi-level optimization problem.' They also highlight the importance of developing methods to assess paper quality beyond citation count and journal name. Future directions include broader domain coverage beyond biomedical sciences and improved handling of evolving scientific knowledge.
Authors identify the need for better prompt optimization in multi-agent settings as an unsolved problem. They note that "while recent work has been conducted on prompt optimization, the complex setting of multiple agents with individual prompts is unsolved." They also suggest future work on reducing hallucinations in LLMs through improved RAG agent design.
Need for better prompt optimization in multi-agent systems
Development of more sophisticated quality signals beyond journal name and citation count
Temporal validity of benchmarks in rapidly evolving scientific domains
Integration of real-time calculations and up-to-date information for scientific tasks

Data: LitQA: 50 multiple-choice questions from recent biomedical literature (after September 2021); PubMedQA (modified to remove context); MedQA; BioASQ; Abstract Retrieval Metric dataset: 500 questions derived from PubMed searches; LitQA dataset (50 multiple-choice questions from biomedical literature, post-September 2021); LitQA - 50 multiple-choice questions from biomedical literature (after September 2021); PubMedQA (modified, context removed); MedQA (USMLE questions in English); BioASQ (yes/no questions); Abstract retrieval evaluation dataset - 500 questions from 20,000 PubMed keyword searchesCode: Not explicitly provided in the paper, though the authors mention 'open-source implementation' in conclusion but no GitHub link given; Implementation within LangChain's agent framework (open-source mentioned but no specific repository link provided); Open-source implementation mentioned as available, implemented within LangChain's agent framework, but no explicit GitHub URL provided in the paperExtracted from: pdfAgreement 41%

Explore related topics

RAG for Research Literature Discovery Benchmarking