Research
10 papers and preprints. * denotes equal contribution.
arXiv preprint (under review)
Dynamic Predictive Planning treats manipulation of moving targets as counterfactual planning: the world-action model's own rollout estimates when an interaction will happen and where the target will be, and a counterfactual observation lets the policy invoke a skill it already has instead of improvising a recovery. Runs in real time on a single consumer GPU with no training on dynamic data.
arXiv preprint (under review)
A foundation-to-lightweight VLA framework: it adapts a flow map for single-state teacher supervision and pairs it with a self-consistency objective, distilling a large VLA into a compact student that provably recovers the teacher's distribution.
arXiv preprint (under review)
Teacher-free, rollout-free sequence-level distillation: the sequence-level reverse KL splits into a chunk-level term and a future-potential term that carries the long-horizon effect of the current action, and a one-step drifting objective optimizes both.
arXiv preprint (under review)
A training-free way to stretch a chunked policy's execution horizon by reusing the actions it would otherwise discard, extending it only while action velocity stays smooth. Cuts policy calls by 1.2-1.7x with no loss in success rate across several VLAs and a world-action model.
arXiv preprint (under review)
Casts critic-guided flow generation as optimal control and regresses a lightweight guidance network onto the resulting costate, so a frozen VLA gains trajectory-aware critic guidance at one extra forward pass per step - 3.6x faster than QGF with 7x fewer parameters.
NeurIPS 2026 (Poster)
A positional encoding that models curved rays so a single video diffusion model can be controlled by pinhole, fisheye and panoramic cameras alike. An earlier version appeared at the ECCV 2026 Workshop on 3D in the Era of World Models.
arXiv preprint (under review)
Combines distribution matching distillation with an unpaired neural Schrödinger bridge to translate ultra-low-field brain MRI toward high-field quality.
MICCAI 2026 ULF-EnC Challenge Workshop
A residual diffusion pipeline with an artifact-suppression network for the MICCAI ultra-low-field enhancement challenge.
Korea Computer Congress (KCC) 2025
Measures how Kubernetes traffic-scheduling algorithms behave across homogeneous and heterogeneous clusters.
KDD Cup 2025 Workshop on CRAG-MM
Prompt-level adaptation and knowledge-driven fine-tuning for multimodal RAG question answering in the Meta CRAG-MM challenge.