DERA: Enhancing Large Language Model Completions with Dialog-Enabled Resolving Agents
This work addresses the need for more reliable AI outputs in healthcare, though it is incremental as it builds on existing conversational LLM capabilities.
The paper tackles the problem of improving factual accuracy and completeness of large language model outputs in safety-critical healthcare applications by introducing DERA, a dialog-enabled resolving agent paradigm, which significantly outperforms base GPT-4 on medical conversation summarization and care plan generation tasks.
Large language models (LLMs) have emerged as valuable tools for many natural language understanding tasks. In safety-critical applications such as healthcare, the utility of these models is governed by their ability to generate outputs that are factually accurate and complete. In this work, we present dialog-enabled resolving agents (DERA). DERA is a paradigm made possible by the increased conversational abilities of LLMs, namely GPT-4. It provides a simple, interpretable forum for models to communicate feedback and iteratively improve output. We frame our dialog as a discussion between two agent types - a Researcher, who processes information and identifies crucial problem components, and a Decider, who has the autonomy to integrate the Researcher's information and makes judgments on the final output. We test DERA against three clinically-focused tasks. For medical conversation summarization and care plan generation, DERA shows significant improvement over the base GPT-4 performance in both human expert preference evaluations and quantitative metrics. In a new finding, we also show that GPT-4's performance (70%) on an open-ended version of the MedQA question-answering (QA) dataset (Jin et al. 2021, USMLE) is well above the passing level (60%), with DERA showing similar performance. We release the open-ended MEDQA dataset at https://github.com/curai/curai-research/tree/main/DERA.