Med-PaLM
A medical-domain language model aligned from Flan-PaLM using instruction prompt tuning.
Med-PaLM is a large language model for clinical and medical question answering, created by Google researchers and described in a paper posted in December 2022. It was built by aligning the 540-billion-parameter Flan-PaLM to medicine with instruction prompt tuning, and evaluated on the MultiMedQA benchmark alongside human clinicians.
- Built by instruction prompt tuning on Flan-PaLM, a 540-billion-parameter model
- Introduced MultiMedQA, combining six existing datasets with new HealthSearchQA
- Flan-PaLM scored 67.6% on MedQA, over 17% above prior state-of-the-art
- Human evaluation axes: factuality, precision, possible harm and bias
- Encouraging results, but still inferior to clinicians
What it is
Med-PaLM is a language model aimed at medical and clinical use, reported by a team of Google researchers in a paper posted in December 2022. It starts from Flan-PaLM, the instruction-tuned form of PaLM, a model with 540 billion parameters. The team adapted it to medicine using instruction prompt tuning, a method that aligns a model to a new domain with only a handful of example prompts. The work also introduced MultiMedQA, a benchmark built from six existing question-answering datasets plus HealthSearchQA, a new set of free-response medical questions taken from online searches.
How it came about
The model appears in the paper Large Language Models Encode Clinical Knowledge, whose author list runs to 30 people including Karan Singhal, Shekoofeh Azizi, Tao Tu, Blaise Aguera y Arcas, Alan Karthikesalingam and Vivek Natarajan. The preprint was posted on 26 December 2022. It grew out of an effort to assess clinical knowledge more broadly than existing automated benchmarks allowed, combining several question-answering datasets and adding a human evaluation framework.
Inside the model
Med-PaLM is produced by instruction prompt tuning, described as a parameter-efficient way of aligning a language model to a new domain using a few exemplars. The base is Flan-PaLM, itself an instruction-tuned version of PaLM, a 540-billion-parameter model. The team evaluated PaLM and Flan-PaLM on MultiMedQA using a combination of prompting strategies before applying the medical alignment step. Their human evaluation framework scored answers on axes including factuality, precision, possible harm and bias. The paper reports that comprehension, recall of medical knowledge and medical reasoning all improved with model scale and with instruction prompt tuning.
What it can do
Flan-PaLM reached state-of-the-art accuracy on every multiple-choice dataset in MultiMedQA, including 67.6% on MedQA, which draws on US Medical License Exam questions, more than 17 percentage points above the previous best result reported there. Human evaluation of Flan-PaLM answers exposed important gaps, which Med-PaLM was designed to address. The paper says Med-PaLM performed encouragingly but still fell short of clinicians. It also stresses that today's models have notable limitations.
How to use it
The work is presented as research. The paper describes the model and its evaluation but does not set out a public release of weights, an API or a licence.
Safety and testing
The team built human evaluation around factuality, precision, possible harm and bias, and used it to surface weaknesses that automated benchmarks missed. Their stated aim is to reinforce both evaluation frameworks and method development for safe, helpful models in clinical settings. No incidents are described.
Reception and debate
The paper reports that Med-PaLM performed encouragingly on the human evaluation axes, while remaining inferior to clinicians. It frames the results as showing both the potential of large language models in medicine and the limitations of current systems.
Why it matters
The work introduced MultiMedQA, combining six existing question-answering datasets with the new HealthSearchQA set of free-response medical questions from online searches. It also proposed a human evaluation framework covering factuality, precision, possible harm and bias, and showed that comprehension, knowledge recall and medical reasoning improve with scale and instruction prompt tuning.
Questions people ask about Med-PaLM
What is Med-PaLM?
A medical version of the Flan-PaLM language model, created by Google researchers and described in a December 2022 paper. It was aligned to the medical domain using instruction prompt tuning with a small number of example prompts, and tested on clinical and consumer medical questions.
How was it adapted for medicine?
Through instruction prompt tuning, a parameter-efficient technique that aligns a model to a new domain using only a few exemplars. The starting point was Flan-PaLM, the instruction-tuned variant of the 540-billion-parameter PaLM model.
How well does it perform?
Its parent Flan-PaLM reached 67.6% accuracy on MedQA, drawn from US Medical License Exam questions, over 17% above the prior best result reported. Human reviewers found Med-PaLM encouraging but still below the level of clinicians.
How were the answers judged?
A human evaluation framework scored responses on several axes, including factuality, precision, possible harm and bias. This was introduced because automated benchmarks alone did not capture the quality of model answers across a breadth of medical tasks.
Can I use it?
The work is presented as research. The paper describes the model and its evaluation but does not set out a public release of weights, an API or a licence.
Good reading and sources
There is no English Wikipedia article on Med-PaLM yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2212.13138. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.