Projects

Research on human–AI co-creation, robot world models, biomedical reasoning, and single-cell dynamics.

• Nothing Happens for a Reason: Null Interventions for Hit/Miss Ranking in Robot World Models — 2026. PROJECT. Geometry-screened counterfactual null interventions for action-conditioned robot world models. Conservative planned-geometry checks generate execution-free no-effect labels, paired with same-state hit-over-miss ranking so predicted misses stay static without collapsing to an all-static predictor. On DINO-WM near-miss ranking, HardSucc rises from 0.73 to 0.95 in simulation and from 0.48 to 0.83 on a real SO-101 arm.
• scWorldModel: An Imagination-Trained Recurrent State-Space World Model for Time-Course Single-Cell RNA-seq — 2026. PROJECT. An imagination-trained recurrent state-space model for unpaired single-cell populations. scWM learns a shared recurrent population state with posterior and free-running sliced-Wasserstein training, per-cell stochastic decoding, and modality-specific action encoders, reducing temporal extrapolation error by up to 16× versus latent-dynamics baselines.
• Who Gets to Decide? Conflict-Visible Multi-Agent Planning for Human-Governed Text-to-Image Co-Creation — 2026. PROJECT. A planning workflow where specialist agents propose alternatives, disagreements stay visible, and high-impact conflicts can be adjudicated by the user rather than resolved automatically. A 50-person study finds that conflict-visible adjudication preserves outcome authorship and creativity support better than hidden multi-agent resolution.
• MorphoJEPA: Cross-Modality-Robust In-Domain Self-Supervision for Low-Label Live-Cell Division Forecasting — 2026. PROJECT. An in-domain joint-embedding predictive model for live-cell division forecasting from unlabeled microscopy. MorphoJEPA holds up when imaging modality shifts — especially on phase-contrast — where off-the-shelf image, video, and cell encoders degrade.
• BioSkillGate: Validation-Gated Skill Adaptation for Biomedical Reasoning — 2026. RESEARCH / AUTHOR. A framework that adapts natural-language skill documents for a frozen language model, keeping only edits that pass an inner held-out gate. On Qwen3-8B it lifts average accuracy across biomedical multiple-choice benchmarks while leaving model weights and decoding unchanged.
• CRoCal: Trajectory-Level Conformal Calibration for Clinical World Models — 2026. PROJECT. A backbone-independent calibrator that turns clinical world-model forecasts into joint 24-hour prediction boxes, conditioned on ICU regime. Across PhysioNet, MIMIC-IV Sepsis, and eICU-CRD it is the only evaluated method that tracks the 90% joint-coverage target under heterogeneous ICU types.