6arXiv cs.CL (Computation and Language)·
Self-supervised pretraining drives representational convergence in medical image encoders more than clinical supervision
A controlled study across 18 image and 7 text encoders spanning 7M to 27B parameters finds that representational convergence in medical foundation models is primarily driven by the self-supervised pretraining objective, not by clinical supervision or model scale. Using 650,982 chest radiographs across six datasets, the authors show matched self-supervised encoders align at 40.4% versus 21.1% for label-supervised and 3.3% for image-text models, and that convergence does not grow with parameter count. Despite modest convergence, a linear classifier transfers across encoders to five held-out hospitals retaining ~85% of within-encoder performance, suggesting interoperability must be deliberately designed through objective choice rather than assumed from scale.
Related guides (1)
Related events (8)
Domain-specific transformer embeddings for detecting dosing errors in clinical trial protocols
Researchers from CaresAI evaluated biomedical transformer models (ClinicalBERT, PubMedBERT, BioBERT, MedCPT) for detecting dosing errors in clinical trial protocols, combining text embeddings with structured metadata and classical ML classifiers. BioBERT achieved the best single-encoder performance at ROC-AUC 0.794, while gradient boosting and SVM ensembles reached 0.821–0.853. The study finds that domain alignment of the encoder matters more than stacking multiple embeddings, and demonstrates a practical NLP pipeline for clinical trial safety monitoring.
Training framework reduces calibration error 60%+ in Medical VQA multimodal LLMs
A new arXiv preprint proposes a finetuning framework to improve verbalized uncertainty calibration in multimodal LLMs applied to Medical Visual Question Answering. The composite loss function combines Brier-style calibration, anchor regularization, contrastive image-text alignment, and KL-based stabilization, evaluated on MedGemma 4B IT and Qwen2-VL 7B Instruct across three medical VQA benchmarks. The method reduces calibration error by 60% or more and improves discrimination by 26% or more while preserving predictive accuracy, outperforming prompting-, sampling-, and training-based baselines.
LLM-guided MAP-Elites evolution improves medical decision pipelines at inference time
Researchers propose using LLM-guided MAP-Elites evolutionary search as an inference-time alternative to fine-tuning for adapting LLMs to clinical workflows, formulating triage, consultation, and image classification as evolutionary searches over executable artifacts. Across three medical settings, evolved programs substantially outperform manually designed baselines: triage accuracy improves from 77.3% to 87.1% and emergency recall from 0.60 to 0.97, with gains also shown on MIMIC-ESI, iCRAFTMD, and PneumoniaMNIST. The approach works across Llama-3, Qwen-3.5, and Gemma-4 backbones and produces interpretable program-level mechanisms rather than superficial prompt changes.
Visual pretraining on document images outperforms text-only pretraining for language intelligence
A new arXiv preprint argues that training language models directly on visual representations of documents (figures, equations, page layouts) consistently outperforms text-only pretraining on the same underlying corpora. The authors conduct a systematic study of unsupervised visual pretraining paradigms across multiple backbones and benchmarks, framing visual pretraining as a scalable alternative to the dominant text-extraction pipeline. The result challenges a foundational assumption in LLM pretraining and has implications for how future foundation models are trained on visually rich sources like PDFs and web pages.
Normal Guidance: Bell-Curve Regularization for Attention-Based MIL in 3D Medical Imaging
This paper addresses weakly supervised classification of 3D medical images where only volume-level binary labels are available. The authors identify that a simple center-focused baseline outperforms attention-based and transformer-based multiple instance learning (MIL) at slice-level classification across brain, thoracic, and abdominal CT datasets. They propose Normal Guidance, a regularization technique that constrains learned attention distributions to follow a bell-shaped curve, achieving superior slice-level localization over state-of-the-art MIL methods across datasets totaling over 4 million 2D slices.
Representation-Conditioned Diffusion Models for Controllable Image Generation
This paper explores conditioning diffusion models on representations from pre-trained self-supervised models as an alternative to text prompts or semantic maps, which require large annotated datasets. The self-conditioning mechanism improves unconditional image generation quality and provides a controllable representation space. The authors identify directions of variation in this space and demonstrate smoothness and disentanglement properties, suggesting potential for fine-grained generative control without heavy annotation overhead.
MC Dropout uncertainty estimation masks sub-region calibration failures in brain tumour segmentation
A preprint from arXiv evaluates Monte Carlo Dropout for voxel-level uncertainty estimation in glioma segmentation on 126 BraTS21 patients, comparing a pretrained SegResNet and a locally trained UNet-Res. While global uncertainty-error alignment is strong (AUROC ~0.97), the study finds that UNet-Res exhibits near-zero entropy and an ECE of 0.915 on the enhancing tumour sub-region despite a Dice of only 0.714, a severe miscalibration invisible to standard Dice and AUROC metrics. The paper argues that sub-region-specific calibration assessment is necessary for clinical safety and cannot be replaced by aggregate metrics alone.
Foundation Model for Wearable Health Data Pretrained on 1 Trillion Minutes from 5 Million Participants
Researchers propose a large-scale foundation model for wearable health data, pretrained on over one trillion minutes of unlabeled sensor signals from five million participants. The model demonstrates systematic performance improvements across 35 health prediction tasks spanning cardiovascular, metabolic, sleep, and mental health domains, with joint scaling of model capacity and data volume. A 'classroom' of LLM agents autonomously searches downstream predictive head configurations, and the resulting embeddings are integrated into a Personal Health Agent validated by 1,860 clinician ratings. The work establishes label-efficient few-shot learning and generative capabilities for daily health metric estimation.