Using specialized data to improve research agent accuracy

A benchmark evaluation of 288 biology-research questions found that giving an AI agent direct access to PubMed and ClinicalTrials.gov alongside public-web search reduced its overall error rate from 11.8% to 7.6%, producing 12 additional correct answers. Using unchanged LABBench2 prompts and scoring, the study tested 168 literature questions and 120 clinical-trial questions twice with the same model, finding accuracy gains from 87.5% to 92.9% for literature retrieval and from 89.2% to 91.7% for trial retrieval, with similar costs and median response times. Agents used PubMed extensively for paper-based questions and ClinicalTrials.gov for trial questions, often combining broad web searches with connector-based retrieval of structured records. Trace reviews suggested connectors were especially useful for distinguishing similarly named studies, locating trials that were difficult to find online, identifying the correct study phase, and extracting complete information from structured fields that could be missed in lengthy web-page tables. Although biomedical sources are relatively open and well indexed, the findings suggest direct access may be even more consequential in fields such as law and finance, where authoritative information is more often restricted to licensed databases.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.