Introduced the need to study personalized safety in LLMs. Argued that alignment should not be purely global, but tailored to a user's background since risk profiles vary across users. Showed where current models fall short, and introduced an inference-time mitigation approach using LLM-guided Monte Carlo Tree Search.
Trading-R1 is an RL-trained reasoning LLM that transforms heterogeneous financial signals into structured, auditable investment theses and volatility-aware trade ratings by learning multi-perspective financial reasoning through reverse chain-of-thought distillation and an easy-to-hard RL curriculum, achieving superior risk-adjusted backtest performance over generic instruction-following and reasoning models.
Propose a scalable LLM-based pipeline for automatically generating researcher interest profiles from PubMed, addressing the prevalence of outdated, incomplete, and inaccurately scraped online profiles that hinder expert discovery and collaboration. We compare MeSH-term–based generation with abstract-based summarization, introduce a TF-IDF–KL divergence metric to analyze semantic differences from human-written profiles, and show that despite low lexical overlap, semantic similarity is moderate. Human evaluations favor MeSH-based profiles for readability, demonstrating the feasibility of automated researcher profiling at scale.
An agent that enables interactive product search by combining freehand sketches, diffusion-based visual refinement, and a conversational language agent with memory. The framework transforms rough user drawings into semantically meaningful visual queries and iteratively refines results through dialogue, then performs vector-based embedding search over a product index.
A multi-agent framework in which specialized agents generate code to create visuals and animations, extending LLM assisted learning beyond a simple chatbot experience. The system also incorporates STORM-style research and structured document generation to produce long-form study guides, practice tests, and problem sets.
A multi-agent LLM framework for financial trading, with agents crawling technical, news, sentiment, and more to make a trading decision. GitHub (27k stars)
Educational support that uses modified STORM-style research to conduct deep, topic-focused analysis and produce highly detailed long-form content. The output is organized into lesson plans, lessons and slides, practice problems, and practice tests, enabling full curriculum development and teaching preparation automation.
A multimodal LLM that aligns RNA sequence representations with large language models for direct sequence-to-language reasoning. It couples a pretrained RNA encoder with an instruction-tuned LLM via a lightweight projection module and is trained on RNA-QA, a literature-derived dataset of over 400,000 RNAs built with an LDA-based, topic-aware summarization pipeline to reduce topic drift and information loss.
Used diffusion models to perform weather-condition class balancing (especially in extreme, edge case weather conditions), thereby improving robustness to out-of-distribution (OOD) conditions in self-driving datasets and applications.
An agent that uses patient–doctor visit recordings not only to generate clinical notes for physicians, but also to support patients at home by reasoning over visit interactions and post-visit needs.
ProteinGPT is a multimodal LLM that accelerates protein research by enabling natural language interaction with protein sequences and structures. Trained on over 130,000 proteins, it integrates evolutionary-scale protein folding and inverse folding models (esm2) to encode sequence and structural information, aligns these modalities with language, and is instruction-tuned to deliver an interactive, high-performance protein-focused chat experience.
APR beamforming, combined with a marker-embedded assistant structure, enables flexible ultrasound arrays to reconstruct images without external shape sensing. By aligning strong marker-generated peak echoes across scanlines, the method estimates transmit and receive delays within a modified delay-and-sum framework. Simulation and phantom experiments show accurate target localization, supporting real-time pancreatic tumor tracking for more precise and safer radiotherapy.
A benchmark for evaluating logical reasoning in multimodal large language models within visual contexts. The dataset pairs images with carefully designed logic questions that require models to integrate visual perception and abstract reasoning, going beyond recognition or captioning.
An exploration of gamma index calculations for radiotherapy using GPU parallelization with CUDA. The gamma index is a standard quality assurance metric that compares planned and measured dose distributions by incorporating both dose and spatial agreement. GPU acceleration can significantly reduce computation time for large, high-resolution datasets.
An exploration into using YOLOv5 to quickly segment 3D point clouds by segmenting 2D image, masking depth, clustering point cloud and matching to fill in the other side of the pointcloud
Investigates how iodinated rectal hydrogel spacers affect the accuracy of proton dose calculations in prostate cancer therapy. Using validated material properties and Monte Carlo simulations, the authors show that CT-based interpretation of iodinated spacers can introduce clinically meaningful dose errors, particularly to nearby critical structures. A simple water material override is shown to restore dosimetric accuracy, offering a practical solution for safe clinical implementation of iodinated spacers in proton therapy.
Investigates a phantom-based validation of ultrasound imaging for real-time tracking of pancreatic tumor motion during radiation therapy. Using patient-derived respiratory motion and optical ground truth measurements, the authors demonstrate sub-millimeter tracking accuracy in the superior–inferior direction. The results support ultrasound as a non-ionizing, real-time solution for intra-fraction motion monitoring with potential to enable margin reduction and more precise dose delivery.
A novel beamforming technique for flexible ultrasound array transducers that relies on an auxiliary structure with highly reflective markers for calibration. The reflected signals provide reference peaks that are aligned to compensate for array deformation and achieve effective beam formation.