Machine Learning 2026: Breakthroughs in Healthcare, Agents & Efficient Training

Developed at Penn Medicine, Percival is an AI model trained on roughly 400,000 CT scans from approximately 50,000 patients. Unlike models that must be directed to find specific conditions, Percival runs in the background and picks up patterns that might otherwise go unnoticed by human eyes [citation:1].

The model was found to uncover traces of atrial fibrillation and chronic kidney disease in scans ordered for other reasons. Compared to a similar model called TotalSegmentator, Percival flagged pancreatic cancer, cirrhosis, and bacterial pneumonia much more reliably [citation:1].

400K scans trained 50K patients Open source

"Percival can help identify unrecognized disease, predict future progression, and anticipate major cardiovascular and liver complications," said co-author Charles Kahn. The team hopes to test it clinically early next year [citation:1].

TrajOnco is a training-free, multi-agent large language model framework designed for scalable multi-cancer early detection. Using a chain-of-agents architecture with long-term memory, it performs temporal reasoning over sequential clinical events to generate patient-level summaries and predicted risk scores [citation:3].

Evaluated on de-identified Truveta EHR data across 15 cancer types, TrajOnco achieved AUROCs of 0.64–0.80 in zero-shot evaluation, performing comparably to supervised machine learning in a lung cancer benchmark. The multi-agent design also enabled effective reasoning with smaller-capacity models like GPT-4.1-mini [citation:3].

15 cancer types AUROC 0.64–0.80 Training-free

"These findings highlight the potential of multi-agent LLMs to execute interpretable temporal reasoning over longitudinal EHRs," the authors write, "advancing both scalable multi-cancer early detection and clinical insight generation" [citation:3].

Published in Nature Computational Science, PathSegmentor is a foundation model that segments 160 pathological categories via natural language prompts. It generalizes across datasets and clinical cohorts, enabling scalable and interpretable pathology image analysis [citation:4].

Unlike traditional segmentation models that require task-specific training, PathSegmentor accepts natural language descriptions of what to segment, making it adaptable to new tissue types and disease categories without retraining.

MIRACLE is an online continual learning framework that efficiently integrates single-cell multimodal data across batches, modalities, tissues, and diseases. As single-cell datasets grow in size and complexity, MIRACLE enables models to learn incrementally without forgetting previously acquired knowledge [citation:4].

The framework addresses a fundamental challenge in computational biology: how to build unified representations from data generated by different labs, technologies, and experimental conditions.

Step-by-step verifiers, also known as process reward models (PRMs), are a key ingredient for test-time scaling, but training them requires expensive step-level supervision. ThinkPRM builds data-efficient PRMs as verbalized step-wise reward models that verify every step by generating a verification chain-of-thought [citation:14].

Fine-tuned on orders of magnitude fewer process labels than discriminative PRMs, ThinkPRM outperforms LLM-as-a-Judge and discriminative verifiers using only 1% of the process labels in PRM800K. On ProcessBench, MATH-500, and AIME '24, it beats baselines under best-of-N selection and reward-guided search [citation:14].

1% of labels Outperforms full-label baselines

Test-time scaling (TTS) improves LLM reasoning by investing more compute at inference, but repeated sampling and aggregation grow costs linearly. Maze shifts some computation from repeated generation to prompt planning before generation [citation:10].

The framework models ordered combinations of few-shot examples as prompt trajectories in a latent geometric space, enabling joint optimization of example selection, ordering, and relevance to the current problem. The planner is decoupled from the target model, so it works with black-box deployments. On five benchmarks covering professional knowledge, scientific QA, and mathematical reasoning, Maze consistently outperforms prompt optimization methods under single-call conditions [citation:10].

"Reasoning capability expansion depends not only on the scale of compute investment," the authors write, "but also on how compute resources are allocated" [citation:10].

Efficient long-context inference requires compressing the KV cache under strict memory budgets without degrading output quality. ODM-KV formulates this as a constrained optimization problem: minimize total attention output distortion under a unified average bit-width budget, where each token's KV entries can be assigned different bit widths from 0 (evicted) to 16 (uncompressed) [citation:10].

Using first-order perturbation analysis, the method approximates each token's compression cost as the product of a risk score (sensitivity to perturbation) and quantization error. Experiments on LongBench and RULER with Llama-3.1-8B-Instruct and Qwen3-8B show that ODM-KV achieves accuracy comparable to uniform 4-bit quantization while using only 2 bits per entry on average [citation:10].

2-bit average Matches 4-bit accuracy Native mixed-precision

Caltech-led research demonstrates that a recurrent neural operator (RNO) can predict tipping points in nonstationary systems — where the underlying physics change over time — something standard machine learning cannot handle [citation:9].

In a simulated climate scenario where low ocean clouds rapidly disappear due to rising CO2, the RNO predicted the tipping point's timing with an error of about 11 days, more than two years ahead. It also predicted the tipping point in simulations of airflow over a tilting plane wing, where lift drops precipitously at stall [citation:9].

11-day error 2+ years ahead Climate & aerospace

"The system is drifting away over time, and the laws from physics are no longer holding at the tipping point," said Anima Anandkumar, Caltech's Bren Professor. "We can use the principles of physics on top of recurrent neural operators to make these predictions" [citation:9].

Researchers at the Chinese Academy of Sciences' Shenzhen Institute of Advanced Technology released MatBrain, the first lightweight dual-model collaborative agent for materials research. Instead of relying on a single massive model, MatBrain uses two smaller models working in tandem: Mat-R1 (30B parameters) for materials knowledge, reasoning, and judgment, and Mat-T1 (14B parameters) for calling databases, structure generation models, and computational software [citation:17].

MatBrain's comprehensive materials prediction accuracy is 66% higher than GPT-5 and DeepSeek-R1, while hardware deployment costs are reduced by over 95%. It has already run the full chain from hypothesis to experimental validation for electrocatalyst applications [citation:17].

66% accuracy gain 95% cost reduction Dual-model

Running ML inference on edge devices is increasingly common, but resource constraints make MLOps monitoring difficult. LSDD+ extends the Latent Space Drift Detector with more robust drift detection and a more advanced model adaptation step that can fully adapt to new concepts while returning to known concepts when necessary [citation:15].

The method detects concept drift by monitoring the latent space of an autoencoder, requiring fewer computational and memory resources than existing unsupervised drift detectors. It includes two new benchmark datasets — one for images and one for time series — and evaluates class separability across different autoencoder types [citation:15].