On Wednesdays, We Ask Questions: Optimizing "Active Listening" in Automated Legal Triage and Referral
Researchers have developed a system called FETCH that uses a low-cost ensemble of LLMs to generate follow-up questions for legal triage. While these models are effective at classification, generating high-quality, plain-language questions proved challenging without more sophisticated models. The study found that prompt engineering alone was insufficient, and human evaluations diverged from LLM-as-judge assessments. Incorporating GPT-5 significantly improved the system's ability to elicit relevant information and enhance classification accuracy, though uneven fact elicitation across certain legal areas was noted. AI
IMPACT Enhances AI's role in legal services by improving information elicitation and classification accuracy.