Research
Papers and ongoing work.
KeyFlow: Agentic Policy Adaptation with Episodic Memory
CoRL 2026 Workshop Learn@Deploy
A coding agent adapts a manipulation policy by revising semantic 3D keypoints and retrieval programs from simulated play, without new demonstrations or weight updates. KeyFlow reaches 62.6% held-out success across 13 RoboCasa tasks and improves a base VLA by 37.2 percentage points across six IsaacSim tasks.
Generalized Planning Interfaces for Agentic Robotics
CoRL 2026 Workshop on Agentic Robotics
ENCORE: Few-Shot Agentic Discovery of Manipulation Strategies
NeurIPS 2026 Robot Learning Workshop (Oral, 6/43) · CoRL 2026 Workshop on Agentic Robotics
A coding agent reads three demonstrations, writes an executable policy program, and repairs it from its own execution traces.
Interpreting Diffusion Language Models Over Denoising Time
CoLM 2026 Workshop on Non-Autoregressive Language Models
Sparse autoencoders and cross-layer transcoders for diffusion LLMs, with timestep-aware tracing and causal interventions.
[Poster]
Doodle Agent: Exploring Freeform Visual Generation with Multimodal LLMs
ICCV 2025 Workshop on AI for Visual Arts (AI4VA)
A drawing agent that turns natural-language prompts into drawing actions, iteratively choosing brushes, colors, and coordinates.
[Website]
CNOT-Optimal Circuit Synthesis
Top 5 Paper, Research Science Institute (RSI), 2023
Synthesizing CNOT circuits under directed topological constraints, with hardness results for minimizing circuit size.
[Paper]