SETA: Scaling Environments for Terminal Agents
A framework for synthesizing and evolving grounded, executable terminal environments with reliable verification and adaptive difficulty.
Research directions
Research
A concise archive grouped by the scientific direction each project contributes to.
01 / Current direction
Agentic AI
Post-training, agent environments, tool use, and infrastructure for capable LLM agents.
02 / Previous direction
Generative 3D & Vision
Structured generation and spatial reasoning from large, uncurated image collections.
03 / Previous direction
Medical Imaging & Neuroscience
Acquisition, reconstruction, representation learning, and image analysis across MRI, connectomics, and pathology.
Ultra-high temporal resolution dynamic angiography
Subspace reconstruction for arterial spin labelling angiography at 14.7 ms temporal resolution.
Efficient 3D cone trajectory design
A two-stage k-space trajectory for combined high-resolution angiographic and perfusion imaging.
Motion correction with subspace self-navigation
Joint navigator and image reconstruction to reduce motion artefacts in combined angiography and perfusion imaging.
Multi-modal representation learning with UK Biobank
Data-driven MRI features for predicting neurocognitive and health phenotypes from large-scale multimodal data.
Large-scale 3D mitochondria segmentation
3D instance segmentation using flux, distance transforms, and consistency regularization on MitoEM.
Hierarchical pathology screening for cervical abnormality
A three-stage framework for robust whole-slide cervical abnormality screening and suspicious-cell localization.
Real-time endoscopic cancer detection
A fast and interpretable detection system for carcinoma pharyngeal screening in endoscopy video.