Dylan Patel

Insertion Language Models Get a Principled Foundation via Continuous-Time Markov Chains

This paper derives a diffusion-style denoising objective for Insertion Language Models (ILMs) by modeling the noising process as a continuous-time Markov chain (CTMC) over variable-length sequences — providing a principled, first-principles unification of previously ad-hoc insertion-based generation methods. Prior ILM formulations are shown to be special cases of this generalized CTMC denoising framework. Empirically, the approach matches left-to-right and masked diffusion models on language modeling benchmarks while outperforming both on a synthetic planning task, and offers greater sampling flexibility than existing ILMs.

The Daily Dose: Workflow-Integrated Large Language Model Automation for Clinical Summarization and Trial Identification in Radiation Oncology

OH/IR stars in the inner Galactic bulge: I. Colour, CO line, and stellar light curve characteristics

GWTC-5.0: An Introduction to Version 5.0 of the Gravitational-Wave Transient Catalog

Study of nuclear effects on charm production in light-ion collisions

First Ab Initio DDT Mechanism Explains Why Type Ia Supernovae Are Standard Candles

This paper presents the first globally self-consistent, laboratory-validated mechanism for the deflagration-to-detonation transition (tDDT) in near-Chandrasekhar mass white dwarf Type Ia supernova progenitors, replacing decades of ad hoc detonation-initiation assumptions. Using 3D hydrodynamical simulations, the authors show that the turbulence-driven Chapman-Jouguet criterion acts as a physical attractor: despite varying central ignition density by a factor of six and using qualitatively distinct ignition topologies, all models converge to nearly identical peak-luminosity spectra matched to SN 1999aa. This convergent behavior provides the first physically motivated explanation for the empirical standardizability of SNe Ia used in cosmology.

Single-Channel Bladder Pressure Signals Are Sufficient for Automated Urological Event Classification via Two-Stage MLP

A two-stage MLP framework trained on wavelet-derived features from vesical pressure (Pves) signals alone achieves clinically meaningful accuracy for classifying three urological event types — voiding contraction, detrusor overactivity, and abdominal artifacts — without requiring the abdominal pressure channel standard in conventional urodynamics. The pipeline processes 0.8-second segments, aggregates features at the event level, and was externally validated across three independent UDS datasets, demonstrating cross-dataset generalizability. This work establishes a pathway for automated interpretation of catheter-free, home-based bladder monitoring devices that capture only Pves.

Thermo-VL: Injecting Thermal Infrared Perception into Frozen Vision-Language Models via Prompt-Conditioned Fusion

Thermo-VL extends a frozen Molmo-7B VLM to handle thermal infrared inputs without retraining the base model, using a trainable thermal encoder and a text-guided dual-attention fusion module that conditions thermal features on both language prompts and RGB context. A gated residual mechanism injects thermal evidence into the frozen RGB stream, preserving the pretrained RGB-language interface while enabling multispectral reasoning. The authors also release a pixel-aligned RGB-thermal instruction-tuning dataset and Thermo-VL-Bench, a curated VQA benchmark targeting low-light and cross-spectrum scenarios. The approach demonstrates strong performance on thermal-only and RGB+thermal tasks, validating prompt-conditioned multispectral fusion as an effective paradigm for degraded-illumination perception.