Clinical Agent Safety: Benchmark and Runtime Guard (VMAG)
July 2026 – present
Built a synthetic-FHIR benchmark that scores each agent step as act, gather missing information, abstain, or escalate, with a policy guard outside the model.
In a 20-case guard-off versus guard-on evaluation, a 7B open model's unsafe-action rate fell from 0.35 to 0.0. A frontier model remained at 0.0 unsafe-action rate while its over-refusal rate fell from 0.375 to 0.0. These are small synthetic-benchmark results, not evidence of clinical safety or generalization.
Evaluations can be reproduced offline using a committed response cache without credentials. This work instantiates and measures prior enforcement approaches; it does not claim a new architecture. Project details · Earlier 14-case browser demo
MOPA: Multimodal Ordinal Preference Alignment
February 2026 – present · Co-first author
Micro-video popularity prediction at upload time, before any interaction. Recast the task as ordinal exceedance estimation: the model emits monotone threshold probabilities rather than a scalar, with monotonicity guaranteed by construction, over frozen VideoPrism features fused with upload-time metadata by a gated multimodal unit.
A second, critic-free pairwise preference stage — DPO-style, from the RLHF/DPO/GRPO lineage but applied outside language — contrasts top-percentile videos against average ones to correct the conservative bias supervised training leaves on the tail.
Best of 10 prior methods on all four metrics, on a 3,989-video sample of SMPD-Video (SMP Challenge) with a 399-video evaluation split. Against the strongest baseline per metric: nMSE 0.6691 to 0.4989 and Spearman 0.5849 to 0.7171. On the top-5% tail, against the best baseline, Recall@5% 0.220 to 0.380. Substituting random preference pairs for targeted ones degrades every metric, which is what shows the gain comes from what is contrasted rather than from adding a stage.
Results are single-run (seed 42) and model-selected on the same validation split the metrics are reported on, with no separate held-out test split. A complete 9-page manuscript exists; no venue is claimed here. Project details
Interactive Three-Body reading platform
Built a participatory narrative system with four perspectives, cross-character memory, rewindable decisions, and a live Newtonian simulation.
RL reward-signal validation
Analyzed 220 WebArena-lite and Terminal Bench trajectories, audited counterexamples, and developed a prototype reward feature that improved combined AUC from 0.726 to 0.767.
Paper-to-MP3 research tool
Built a Python CLI that fetches papers, extracts body text, summarizes paragraphs in Chinese with a local model, and exports Markdown plus MP3 narration.