Polaris: A Safety-focused LLM Constellation Architecture for Healthcare

We develop Polaris, the first safety-focused LLM constellation for real-time patient-AI healthcare conversations. Unlike prior LLM works in healthcare focusing on tasks like question answering, our work specifically focuses on long multi-turn voice conversations. Our one-trillion parameter constellation system is composed of several multibillion parameter LLMs as co-operative agents: a stateful primary agent that focuses on driving an engaging conversation and several specialist support agents focused on healthcare tasks performed by nurses to increase safety and reduce hallucinations. We develop a sophisticated training protocol for iterative co-training of the agents that optimize for diverse objectives. We train our models on proprietary data, clinical care plans, healthcare regulatory documents, medical manuals, and other medical reasoning documents. We align our models to speak like medical professionals, using organic healthcare conversations and simulated ones between patient actors and experienced nurses. This allows our system to express unique capabilities such as rapport building, trust building, empathy and bedside manner. Finally, we present the first comprehensive clinician evaluation of an LLM system for healthcare. We recruited over 1100 U.S. licensed nurses and over 130 U.S. licensed physicians to perform end-to-end conversational evaluations of our system by posing as patients and rating the system on several measures. We demonstrate Polaris performs on par with human nurses on aggregate across dimensions such as medical safety, clinical readiness, conversational quality, and bedside manner. Additionally, we conduct a challenging task-based evaluation of the individual specialist support agents, where we demonstrate our LLM agents significantly outperform a much larger general-purpose LLM (GPT-4) as well as from its own medium-size class (LLaMA-2 70B).

Stochastic Code Generation

Large language models pre-trained for code generation can generate high-quality short code but often struggle with generating coherent long code and understanding higher-level or system-level specifications. This issue is also observed in language modeling for long text generation, and one proposed solution is the use of a latent stochastic process. This approach involves generating a document plan and then producing text that is consistent with it. In this study, we investigate whether this technique can be applied to code generation to improve coherence. We base our proposed encoder and decoder on the pre-trained GPT-2 based CodeParrot model and utilize the APPS dataset for training. We evaluate our results using the HumanEval benchmark and observe that the modified Time Control model performs similarly to CodeParrot on this evaluation.

Evaluation of Social Biases in Recent Large Pre-Trained Models

Large pre-trained language models are widely used in the community. These models are usually trained on unmoderated and unfiltered data from open sources like the Internet. Due to this, biases that we see in platforms online which are a reflection of those in society are in turn captured and learned by these models. These models are deployed in applications that affect millions of people and their inherent biases are harmful to the targeted social groups. In this work, we study the general trend in bias reduction as newer pre-trained models are released. Three recent models ( ELECTRA, DeBERTa, and DistilBERT) are chosen and evaluated against two bias benchmarks, StereoSet and CrowS-Pairs. They are compared to the baseline of BERT using the associated metrics. We explore whether as advancements are made and newer, faster, lighter models are released: are they being developed responsibly such that their inherent social biases have been reduced compared to their older counterparts? The results are compiled and we find that all the models under study do exhibit biases but have generally improved as compared to BERT.