Endoscopic procedures are essential for diagnosing and treating internal diseases, and multi-modal large language models (MLLMs) are increasingly applied to assist in endoscopy analysis. However, current benchmarks are limited, as they typically cover specific endoscopic scenarios and a small set of clinical tasks, failing to capture the real-world diversity of endoscopic scenarios and the full range of skills needed in clinical workflows. To address these issues, we introduce EndoBench, the first comprehensive benchmark specifically designed to assess MLLMs across the full spectrum of endoscopic practice with multi-dimensional capacities. EndoBench encompasses 4 distinct endoscopic scenarios, 12 specialized clinical tasks with 12 secondary subtasks, and 5 levels of visual prompting granularities, resulting in 6,832 rigorously validated VQA pairs from 21 diverse datasets. Our multi-dimensional evaluation framework mirrors the clinical workflow--spanning anatomical recognition, lesion analysis, spatial localization, and surgical operations--to holistically gauge the perceptual and diagnostic abilities of MLLMs in realistic scenarios. We benchmark 23 state-of-the-art models, including general-purpose, medical-specialized, and proprietary MLLMs, and establish human clinician performance as a reference standard. Our extensive experiments reveal: (1) proprietary MLLMs outperform open-source and medical-specialized models overall, but still trail human experts; (2) medical-domain supervised fine-tuning substantially boosts task-specific accuracy; and (3) model performance remains sensitive to prompt format and clinical task complexity. EndoBench establishes a new standard for evaluating and advancing MLLMs in endoscopy, highlighting both progress and persistent gaps between current models and expert clinical reasoning. We publicly release our benchmark and code.

Multimodal deep learning (MDL) has emerged as a transformative approach in computational pathology. By integrating complementary information from multiple data sources, MDL models have demonstrated superior predictive performance across diverse clinical tasks compared to unimodal models. However, the assumption that combining modalities inherently improves performance remains largely unexamined. We hypothesise that multimodal gains depend critically on the predictive quality of individual modalities, and that integrating weak modalities may introduce noise rather than complementary information. We test this hypothesis on a prostate cancer dataset with histopathology, radiology, and clinical data to predict time-to-biochemical recurrence. Our results confirm that combining high-performing modalities yield superior performance compared to unimodal approaches. However, integrating a poor-performing modality with other higher-performing modalities degrades predictive accuracy. These findings demonstrate that multimodal benefit requires selective, performance-guided integration rather than indiscriminate modality combination, with implications for MDL design across computational pathology and medical imaging.

LLMs are increasingly being used in healthcare. This promises to free physicians from drudgery, enabling better care to be delivered at scale. But the use of LLMs in this space also brings risks; for example, such models may worsen existing biases. How can we spot when LLMs are (spuriously) relying on patient race to inform predictions? In this work we assess the degree to which Sparse Autoencoders (SAEs) can reveal (and control) associations the model has made between race and stigmatizing concepts. We first identify SAE latents in Gemma-2 models which appear to correlate with Black individuals. We find that this latent activates on reasonable input sequences (e.g., "African American") but also problematic words like "incarceration". We then show that we can use this latent to steer models to generate outputs about Black patients, and further that this can induce problematic associations in model outputs as a result. For example, activating the Black latent increases the risk assigned to the probability that a patient will become "belligerent". We evaluate the degree to which such steering via latents might be useful for mitigating bias. We find that this offers improvements in simple settings, but is less successful for more realistic and complex clinical tasks. Overall, our results suggest that: SAEs may offer a useful tool in clinical applications of LLMs to identify problematic reliance on demographics but mitigating bias via SAE steering appears to be of marginal utility for realistic tasks.

Deploying large language models (LLMs) in clinical settings faces critical trade-offs: cloud LLMs, with their extensive parameters and superior performance, pose risks to sensitive clinical data privacy, while local LLMs preserve privacy but often fail at complex clinical interpretation tasks. We propose MedOrchestra, a hybrid framework where a cloud LLM decomposes complex clinical tasks into manageable subtasks and prompt generation, while a local LLM executes these subtasks in a privacy-preserving manner. Without accessing clinical data, the cloud LLM generates and validates subtask prompts using clinical guidelines and synthetic test cases. The local LLM executes subtasks locally and synthesizes outputs generated by the cloud LLM. We evaluate MedOrchestra on pancreatic cancer staging using 100 radiology reports under NCCN guidelines. On free-text reports, MedOrchestra achieves 70.21% accuracy, outperforming local model baselines (without guideline: 48.94%, with guideline: 56.59%) and board-certified clinicians (gastroenterologists: 59.57%, surgeons: 65.96%, radiologists: 55.32%). On structured reports, MedOrchestra reaches 85.42% accuracy, showing clear superiority across all settings.

The advent of Large Language Models (LLMs) offers potential solutions to address problems such as shortage of medical resources and low diagnostic consistency in psychiatric clinical practice. Despite this potential, a robust and comprehensive benchmarking framework to assess the efficacy of LLMs in authentic psychiatric clinical environments is absent. This has impeded the advancement of specialized LLMs tailored to psychiatric applications. In response to this gap, by incorporating clinical demands in psychiatry and clinical data, we proposed a benchmarking system, PsychBench, to evaluate the practical performance of LLMs in psychiatric clinical settings. We conducted a comprehensive quantitative evaluation of 16 LLMs using PsychBench, and investigated the impact of prompt design, chain-of-thought reasoning, input text length, and domain-specific knowledge fine-tuning on model performance. Through detailed error analysis, we identified strengths and potential limitations of the existing models and suggested directions for improvement. Subsequently, a clinical reader study involving 60 psychiatrists of varying seniority was conducted to further explore the practical benefits of existing LLMs as supportive tools for psychiatrists of varying seniority. Through the quantitative and reader evaluation, we show that while existing models demonstrate significant potential, they are not yet adequate as decision-making tools in psychiatric clinical practice. The reader study further indicates that, as an auxiliary tool, LLM could provide particularly notable support for junior psychiatrists, effectively enhancing their work efficiency and overall clinical quality. To promote research in this area, we will make the dataset and evaluation framework publicly available, with the hope of advancing the application of LLMs in psychiatric clinical settings.

We know from prior work that LLMs encode social biases, and that this manifests in clinical tasks. In this work we adopt tools from mechanistic interpretability to unveil sociodemographic representations and biases within LLMs in the context of healthcare. Specifically, we ask: Can we identify activations within LLMs that encode sociodemographic information (e.g., gender, race)? We find that gender information is highly localized in middle MLP layers and can be reliably manipulated at inference time via patching. Such interventions can surgically alter generated clinical vignettes for specific conditions, and also influence downstream clinical predictions which correlate with gender, e.g., patient risk of depression. We find that representation of patient race is somewhat more distributed, but can also be intervened upon, to a degree. To our knowledge, this is the first application of mechanistic interpretability methods to LLMs for healthcare.

Medical language models (MLMs) have become pivotal in advancing medical natural language processing. However, prior models that rely on pre-training or supervised fine-tuning often exhibit low data efficiency and limited practicality in real-world clinical applications. While OpenAI's o1 highlights test-time scaling in mathematics, attempts to replicate this approach in medicine typically distill responses from GPT-series models to open-source models, focusing primarily on multiple-choice tasks. This strategy, though straightforward, neglects critical concerns like data privacy and realistic deployment in clinical settings. In this work, we present a deployable, small-scale medical reasoning system, MedS3, designed for long-chain reasoning in clinical tasks using a self-evolution paradigm. Starting with a seed dataset of around 8,000 instances spanning five domains and 16 datasets, we prompt a base policy model to perform Monte Carlo Tree Search (MCTS) to construct rule-verifiable reasoning chains. Each reasoning step is assigned an evolution rollout value, allowing verified trajectories to train the policy model and the process reward model (PRM). During inference, the policy model generates multiple responses, and the reward model selects the one with a newly proposed PRM-guided Vote-Sum (P-VS) strategy. Experiments on eleven evaluation datasets demonstrate that MedS3 outperforms not only the prior strongest medical model by 6.59, but also 32B-level general reasoning models by 8.71 points. Code and data are available at https://github.com/pixas/MedSSS.

Large language models with a transformer-based encoder/decoder architecture, such as T5, have become standard platforms for supervised tasks. To bring these technologies to the clinical domain, recent work has trained new or adapted existing models to clinical data. However, the evaluation of these clinical T5 models and comparison to other models has been limited. Are the clinical T5 models better choices than FLAN-tuned generic T5 models? Do they generalize better to new clinical domains that differ from the training sets? We comprehensively evaluate these models across several clinical tasks and domains. We find that clinical T5 models provide marginal improvements over existing models, and perform worse when evaluated on different domains. Our results inform future choices in developing clinical LLMs.

Language models, such as [1, 2, 3] have emerged to be an efficient tool in the domain of EHR tasks. These models, extensively trained on diverse sources of clinical data, such as physician notes and longitudinal medical codes, have demonstrated remarkable effectiveness in predicting clinical outcomes. Despite their capabilities, measuring and reducing the uncertainties of these models in EHR tasks is crucial for ensuring patient safety, as clinicians can avoid interventions that the model indicates are uncertain and potentially hazardous. In addition, quantifying the uncertainties in clinical tasks can enhance the reliability of AI-driven medical decision-making systems [4]. To address this challenge, leveraging the transparency of model parameters, we utilize established uncertainty metrics and propose to combine them with ensembling and multi-tasking approaches to effectively quantify and mitigate uncertainties in EHR tasks for these white-box language models. Recently, large language models have embarked on demonstrating their utility in clinical-related tasks, including EHR prediction tasks [5], analyzing radiology report examinations [6] and medical reasoning [7]. However, the encapsulation of modern Large Language Models, typically offered as API services with restricted access to internal model parameters and prediction probabilities, impedes the direct application of traditional uncertainty quantification methods. To overcome this limitation, We redefine uncertainty quantification as a post-hoc approach by analyzing the distribution of answers generated repeatedly from our designed prompts for clinical prediction tasks. Inspired by the effectiveness of our proposed methods in reducing model uncertainty for white-box LMs, we adapted and applied ensembling and multi-tasking methods to the black-box settings.

The success of multi-task learning can depend heavily on which tasks are grouped together. Naively grouping all tasks or a random set of tasks can result in negative transfer, with the multi-task models performing worse than single-task models. Though many efforts have been made to identify task groupings and to measure the relatedness among different tasks, it remains a challenging research topic to define a metric to identify the best task grouping out of a pool of many potential task combinations. We propose a metric of task relatedness based on task difficulty measured by pointwise V-usable information (PVI). PVI is a recently proposed metric to estimate how much usable information a dataset contains given a model. We hypothesize that tasks with not statistically different PVI estimates are similar enough to benefit from the joint learning process. We conduct comprehensive experiments to evaluate the feasibility of this metric for task grouping on 15 NLP datasets in the general, biomedical, and clinical domains. We compare the results of the joint learners against single learners, existing baseline methods, and recent large language models, including Llama 2 and GPT-4. The results show that by grouping tasks with similar PVI estimates, the joint learners yielded competitive results with fewer total parameters, with consistent performance across domains.