Projects

We work on how machines see, remember, retrieve, and explain themselves, and on whether the results hold up when you look closely. Please click each project to learn more.

Same Ranking, Different Winner

The scoring target a memory benchmark picks can flip which LLM wins, even when the overall ranking looks stable.

#ConversationalMemory#Evaluation

Demo arXiv GitHub

SEAM

Measuring how much typed user speech gets absorbed into edited artifacts at unmarked within-turn seams.

#HumanCenteredAI#LLMEditing

Demo GitHub

4D Gaussian Streaming

Which Gaussians you pick as anchors barely matters: cheap samplers match the default strategy at deployment budgets, so the standard 8,192-anchor choice over-spends compute.

#ComputerVision#GaussianSplatting

arXiv

Outcome Monitors

Recovery affordances that let an agent notice and repair tool calls that failed silently.

#AIAgents#TrustworthyAI

arXiv

Fixed RAG Compression

A fixed compression layer in a RAG pipeline collapses the reader-scaling trend that the uncompressed setup shows.

#RAG#Evaluation

arXiv GitHub

LAPSE

Memory consolidation flattens the temporal shape of user facts: an ongoing activity is stored as a standing one.

#ConversationalMemory#Evaluation

arXiv GitHub

Thin-Object Segmentation

A cross-domain survey of power-line, crack, and retinal-vessel segmentation, from classical pipelines to promptable foundation models.

#ComputerVision#FoundationModels

Paper

UAV Smoke Detection

A lightweight CNN screens every frame on the drone and hands only uncertain ones to a stronger model, cutting processing time by up to 56.9%.

#ComputerVision#EdgeAI

Paper Poster