Framework for synthesizing potential functions for PBRS using VLM feedback.

An entropy-aware KV cache compression framework for improving LLM long-context reasoning.

Continual pretraining framework for visually grounded commonsense reasoning in CLIP.

A comprehensive suite of SAEs for studying sparse representations in Qwen3 models.

Reinforcement learning-based test-time adaptation for zero-shot text-to-speech.

Framework for composing simple behaviors for generalizable complex robot manipulation.

Lifelong Mixture of Dynamic Experts for robot manipulation.

A structured-data pipeline for building specialized conversational systems in low-resource languages.

A framework for weakness-driven coevolution of data and models for optimization modeling.

A verifiable multi-task gym environment for evaluating quantitative-finance agents.

A single-file executable for running LLMs locally, discussed for potential Android porting.

Coordinated, confidential vulnerability remediation for critical open source software.

An open-source secret management platform for developers.

A Linux distribution utilizing a functional configuration language for reproducible builds.

Project to defend critical open-source software from AI-enabled exploits and bugs.

Self-hosted job search cockpit for tracking and tailoring applications with AI integration.

A high-performance browser engine written in Rust, focusing on parallelism and security.

A 3D graphics rendering engine toolbox implemented in Scheme.

Pure Python as a systems language with AOT compilation.

Feedforward model for real-time photorealistic volumetric rendering of CT scans.