TL;DR
EXPERTQA is a new high-quality dataset designed to bridge the gap between general-purpose AI evaluation and the rigorous needs of domain experts. By involving 484 experts across 32 fields (like medicine and law), the researchers created a benchmark of 2,177 technical questions and expert-verified, attributed answers, revealing that current LLMs still struggle significantly with factuality and reliable sourcing in high-stakes scenarios.
What Problem Does This Paper Address?
Large Language Models (LLMs) are increasingly used by professionals for high-stakes tasks like medical diagnosis or legal research. However, these models are prone to "hallucinations"—generating plausible-sounding but false information. While "attribution" (providing citations) is a common defense against this, existing benchmarks for attribution often use general web queries that don't reflect the complexity of expert information-seeking.
The authors identify a critical gap: we don't fully understand how LLMs perform when faced with "long-tail," niche, or highly technical queries where the cost of being wrong is high.
Methods
The researchers employed an "expert-in-the-loop" framework consisting of several stages:
- Question Curation: Experts (verified via education and 3+ years of experience) wrote 5 questions each from their professional lives, ranging from simple directed questions to complex hypothetical scenarios.
- Response Generation: The team sampled answers from six different systems, including vanilla GPT-4, BingChat, and various Retrieval-Augmented Generation (RAG) setups (using both Google Search and the Sphere corpus).
- Expert Evaluation: The same experts evaluated the generated responses at the claim level. They judged claims based on informativeness, factuality, citeworthiness, support (whether the provided evidence actually backed the claim), and source reliability.
- Expert Revision: Finally, experts edited the claims to ensure they were factually correct and supported by trustworthy sources, creating a "gold standard" dataset.
Datasets
The primary contribution is EXPERTQA, which contains:
- 2,177 questions spanning 32 diverse fields.
- Expert-verified answers with claim-level annotations.
- Attributions for each claim, evaluated for quality.
The distribution of these questions is heavily weighted toward high-stakes "Professional Schools" (like Healthcare/Medicine) and "Engineering and Technology," as seen in the field distribution plots.
Key Insights and Ideas
- Retrieval Source is King: The quality of an AI's answer depends heavily on where it looks. Systems using Google Search results were rated as more reliable than those using static dumps like Sphere, which often contained unreliable blog comments or forum posts.
- The "Hallucination" Paradox: Purely generative systems (like vanilla GPT-4) were often rated as more "useful" but frequently hallucinated URLs. Conversely, RAG systems provided better citations but sometimes missed citations for claims that clearly needed them.
- Automated Metrics are Blind: The study found that common automated metrics for attribution (like AutoAIS) have high precision but low recall. They often flag a claim as "unsupported" simply because they lack the implicit world knowledge that a human expert possesses (e.g., knowing that "calcium carbonate is an alkali").
- High-Stakes Vulnerability: Medicine and law showed a higher percentage of incomplete attributions (35% and 31% respectively) and a higher reliance on unreliable sources compared to other fields.
Key Results
- Dataset Scale: 2,177 questions across 32 fields, with an average of 5.79 claims per answer.
- Human-Expert Agreement: The authors performed a manual audit and found >85% agreement with the experts' labels, validating the dataset's quality.
- Finetuning Gains: The researchers finetuned open-source models (Llama-2, Vicuna) on EXPERTQA. They found that while finetuning significantly improved performance on long-form QA metrics (like ROUGE and QAFactEval), there is still a massive gap between these models and the expert-revised gold standard.
- Factuality: Only about half of the AI-generated claims were labeled as "definitely correct" by experts, highlighting the persistent risk of misinformation.
Limitations and Future Work
The authors acknowledge several limitations:
- Atomicity: Claims were evaluated at the sentence level. Breaking these down into even smaller "atomic" facts might provide more precision but would be significantly more expensive.
- Tokenization Errors: Standard tools sometimes struggled to split sentences correctly when tables or lists were present in the AI's answer.
- Subjectivity: While they measured agreement, some expert judgments (like "informativeness") are inherently subjective.
- Language: The dataset is currently limited to English-speaking experts primarily from the West and Africa.
Future work will focus on expanding the dataset to rarer fields and developing better automated metrics that can match the nuanced reasoning of human experts.
Summary updated