Ensure that artificial general intelligence benefits all of humanity.
— OpenAI Mission · the north star I work toward
A verification-centric deep-research agent: sub-agents specialise in retrieval and verification, route their reports through a shared evidence pool, and a global verifier reasons over the assembled evidence graph before any claim is delivered.
Research agents trained with structured planning and tool interaction, then hardened by local and global verification that refines intermediate steps and validates final answers.
IRIS is the first benchmark to score fairness on the understanding and the generation side of a unified MLLM at once — 60 metrics over three dimensions — exposing generation gaps and mind–action inconsistencies in leading models.
A single LLM plays both planner and worker roles, trained by RL with granular process rewards and a theoretically grounded credit assignment across the two rollouts.
A benchmark of 819 high-risk scenarios that measures behavioral jailbreaks of LLM agents on a real OS, using semantic–physical dual verification and OS-level state rollback.
Streams every reasoning step to downstream agents instead of waiting for the whole chain — cutting latency and, because early steps are the reliable ones, improving accuracy.
Long reasoning chains make multimodal models lean on accumulated text and forget the image; process-level RL rewards visual keyword coverage and keeps attention on the relevant regions.
Structured, degradation-aware reasoning chains — supervised fine-tuning plus reward-driven alignment with adaptive reasoning depth — for visual understanding under real-world degradations.
Predicts where to act from information entropy and optimizes a distance-based location reward with GRPO, giving GUI agents markedly more accurate interaction.
Brings video frame interpolation down to a single diffusion step, replacing multi-step sampling with a one-step model for efficiency.
Aligns diffusion models with reward signals at the group level — fully offline, independent of the sampler, and less prone to reward hacking.
Represents motion as dense point trajectories projected into latent space, steering image-to-video models without touching their architecture.
More selected work in Publications and Technical Reports — or see Research Summary below for how these threads fit together.
.png){kind=logo}
Scholar
LinkedIn
GitHub
