Self-Supervised Learning, Cellular Representation, Neuronal Segmentation, Fate Prediction

DINOv3-based cell embeddings with unsupervised subclusters; PSPA-based neuron and dendrite segmentation in brightfield confocal microscopy.

4D Reconstruction, Surgical Scene Understanding, Scene Graphs, Representation Learning

4D surgical video reconstruction with Gaussian Splatting and foundation model embeddings; enables spatial reasoning for autonomous surgery.

Graph Neural Networks, Synthesis Prediction, Materials Science

RetroSynth: graph-of-phases synthesizability prediction that models phase competition with context-aware message passing for higher PR-AUC and better-calibrated lab-ready candidates.

Tool-augmented LLMs still struggle with reliable tool choice, argument filling, and recovery from flaky calls, especially when interfacing via the emerging Model Context Protocol (MCP) across diverse environments. We will tackle this with reinforcement learning, optimizing sequential decisions for tool selection, parameterization, and error handling using rewards from task success, latency, and safety checks, benchmarking against baselines.

We study sycophancy, the tendency of language models to echo a user’s stated beliefs or flatter them at the expense of truth. Building on evidence that RLHF-trained assistants exhibit sycophancy and that it can surface in real deployments, we aim to develop open datasets and a lean benchmark suite that stress-test agreement-against-truth, belief-invariance, and calibrated refusal across factual, social, and mathematical settings. The goal is a reproducible pipeline to detect, track, and potentially mitigate sycophancy without eroding helpfulness in deployed assistants.

SemCity-LoC is a new approach for localizing UAV cameras in cities without relying on GNSS or dense, textured 3D meshes. Instead, it uses lightweight, semantic 3D city models (with simple geometric primitives and labels) and aligns them with image-inferred semantics and depth to estimate the camera pose end-to-end in a differentiable way. The project aims to enable scalable, GNSS-free aerial localizationfor applications such as search-and-rescue, autonomous drone navigation, and urban mapping.

Working on long-form video understanding, reasoning over hours of surgical video.

Retrosynthesis, Inorganic Chemistry, Reaction Ranking

Retro-Rank-In: ranking framework for inorganic retrosynthesis with state-of-the-art generalization.

Graph Neural Networks, Synthesis Prediction, Materials Science

Reaction Graph Network (RGN) for predicting solid-state synthesis conditions, accelerating materials discovery.

Numerical Reasoning, Language Models, Token-Level Regression

Number Token Loss (NTL): token-level regression for improved numerical reasoning in language models with zero runtime overhead.