Federated learning across clients running different frozen vision encoders, whose coordinate systems cannot be averaged or aligned. Each client sends only the Gram matrix of its class prototypes — a coordinate-invariant summary — and the server recovers a global quotient geometry in a single round, with no alignment maps and no representation training. Across six benchmarks under heterogeneous CNN and transformer encoders it gave the best accuracy of all methods, +9.4 points over no collaboration, at one to two orders of magnitude lower communication.

When a text-to-video model animates the same modern artwork under different seeds, the disagreement is signal, not noise — yet standard uncertainty quantification collapses it to one dispersion number. The first study of the structure of generative uncertainty for modern-art animation: 13 source-blind and reference-aware estimators, a distributional profile, and an artwork-level protocol, on a new corpus of 250 artworks rendered by Wan2.1-14B under four seeds (1,000 videos). It classifies seed-set topology at 0.98 balanced accuracy (chance 0.25) and isolates outlier configurations at AUROC 1.00 where a scalar reaches only 0.35.

Predicts spatial gene expression from standard H&E images with a Transformer that adds two context pathways in every layer: a rotation- and translation-invariant local pathway over neighboring spots and a whole-slide summary. Across six HEST-1k organ cohorts it improves mean Pearson correlation over the strongest baseline by 9.7–11.6% for 10- to 100-gene panels and leave-one-organ-out transfer by 6.0%.

Argues that enterprise AI agents should not simply retain everything: obsolete policies or market regimes cause negative transfer, but may become relevant again. Proposes reversible forgetting with active, dormant, and retired memory states, instantiated as a hysteretic memory controller that gates retirement by policy and tests reactivation in shadow mode, illustrated on regime-dependent financial knowledge.

The first Mixture-of-Recursions framework for gene- and pathway-level omics learning: biological structure drives embedding smoothing, a structural attention bias, and a graph-aware router that sets each token’s recursion depth. Across eight single-cell and multi-omics benchmarks under one five-fold protocol it improved macro-F1 by 8.2 points over the strongest biology-agnostic MoR baseline while using 75% fewer parameters and up to 58% fewer FLOPs than a non-recursive Transformer.

A training-free framework that audits whether frozen pathology foundation models encode biologically meaningful signal, pairing histology with spatial transcriptomics (HEST-1k, 240 samples across breast, skin, and brain) and using TabPFN as a standardized probe. Across five encoders, UNI led pathway decodability at 0.303, but organ-level shift cost 62% of the pathway signal and section identity stayed linearly decodable up to 214× chance — while benign augmentations retained ≥95% performance.

Asks whether a medical VLM is properly unsure when it cannot really see the image. Three locally run VLMs, including a medically specialized one, are tested on VQA-RAD and PathVQA with intact, blank, and mismatched images. On VQA-RAD, removing or swapping the image lowers accuracy while confidence barely moves, nearly doubling calibration error (ECE 0.23→0.43); about half of high-confidence answers turn wrong and 48% of answers do not change when the image is swapped. The medically specialized model is the most overconfident, and a believable wrong image is harder to detect than a blank one.

Poses an open question: why should explicit latent-dynamics structure buy compute efficiency? On a Yahoo Finance benchmark of 319 equities (2018–2026), a compact latent world model matched or exceeded a Transformer’s directional accuracy on raw prices while training in a fraction of the time — yet no existing theory explains the gap or predicts when structured world models should be preferred over unconstrained attention.

Predicts which children go without health insurance from how their household is put together — family type, multigenerational living, parents’ own coverage and citizenship — using U.S. census microdata.

AgriMorph treats a field robot’s physical shape as a decision made while it works, not a fixed hardware property. Separate latent models of the terrain and of the robot’s own state let a learned world model roll out candidate futures under each available body, so shapes are compared before slip, stall or crop contact happens and changed only when the predicted gain beats the transformation cost. An ensemble-disagreement confidence gate blocks shape changes made on weak sensing, and one shape-conditioned controller carries over to every body without retraining.

A walk-forward backtesting framework benchmarks classical, neural, and world-model predictors on a global equity dataset, isolating when engineered features outperform raw price inputs across model families.

Accepted for poster presentation at the UChicago and Caltech AI + Science Conference (Oct 19–21, 2026), which convenes core-AI and domain-science researchers around AI-enabled scientific discovery: an autonomous, symptom-grounded vision-language agent for crop-disease diagnosis spanning 335 crops and ~839K images.

Selected for poster presentation at the AI Scientist Summer Workshop hosted by Microsoft Research (Aug 4, 2026): an autonomous, symptom-grounded vision-language agent for crop-disease diagnosis, evaluated on the largest plant-disease image–symptom dataset to date with stepwise diagnostic reasoning that generalizes to unseen crops.

The first unified benchmark for pathway-guided cancer therapy-response modeling (BINN, GraphPath, PATH), over 2,622 TCGA patients across 5 cohorts; the best model reached 0.92 AUROC on prostate targeted-therapy prediction at 11% class prevalence.

A cross-year benchmark for rice-disease recognition on Bangladeshi field images: training on 2021–2025 and testing only on held-out 2026 data cost 43.91 macro-F1 points across all eight model/strategy configurations, with only 48.28% of validation performance retained. The better fine-tuning strategy proved architecture-dependent, and the most temporally robust model was not the one validation ranked first.

A multi-agent iterative-refinement system (generators DeepSeek R1 + Med-PaLM; evaluators LLaMA 3.1 + Phi-4) aligned to AMA ethics and a five-tier safety assessment. Cut ethical violations by 89% with a 92% risk-downgrade rate over 900 clinical queries.

A reinforcement-learning intervention engine that models intervention timing and selection as a sequential decision problem for syndemic health management.

A first-authored analysis interpreting transformer architectures through physics and dynamical-systems theory to explain their internal computation dynamics.

An agentic, pathway-grounded tree-of-thought inference pipeline that structures LLM reasoning over biological pathway knowledge for mechanistic therapeutic reasoning in oncology.

Benchmarked LLaMA on multi-class healthcare test-result prediction, assessing reliability for responsible medical-AI deployment.

Baseline benchmarks and reproducible metrics for protein structural-ensemble prediction in computational structural biology.

An LLM-powered tool to estimate and analyze carbon footprint from operational data.

A privacy-preserving federated-learning framework for multi-modal 2D/3D dental imaging with cross-modal prompt alignment, hierarchical optimization, and Byzantine-resilient aggregation with differential privacy.

The poster prototype of the multi-agent medical-LLM safety framework (DeepSeek R1 / Med-PaLM with LLaMA 3.1 / Phi-4 evaluators), later extended into the arXiv study above.