About

I am a Research Scientist at ByteDance Seed, working on post-training and personalization for large language models. I received my Ph.D. in EECS from UC Berkeley (graduated Dec 2024), advised by Prof. Anca Dragan in the InterACT Lab. Before Berkeley I was at Stanford (class of 2018), where I worked with Amir Zamir, Silvio Savarese, and Dorsa Sadigh at SVL and ILIAD.

I'm interested in making digital and physical agents more powerful, reliable, and safely aligned. Currently I'm focused on agentic systems, reasoning models, efficient large language models, and RL with verifiable rewards (RLVR). At ByteDance Seed I've contributed to Seed-OSS and the Seed-Thinking series (v1.5, v1.6, v2.0, and beyond).

During my Ph.D., I worked on the safety of human–robot systems from three angles: (1) causal confusion — how learned reward models latch onto spurious correlates of human preference, and how to diagnose it; (2) adversarial perturbations — quantifying robustness of assistive policies along a natural–adversarial frontier; and (3) steerability — making model behavior controllable and personalizable at inference time without retraining.

News

Aug 2025 Seed-OSS-36B released open-source under Apache-2.0.
Apr 2025 Seed1.5-Thinking tech report on arXiv.
Jan 2025 Context Steering accepted at ICLR 2025.
Aug 2023 Quantifying Assistive Robustness accepted at CoRL 2023.
Jan 2023 Causal Confusion in Preference-Based Reward Learning accepted at ICLR 2023.

Tech Reports

ByteDance Seed model releases I've contributed to.

2026 Seed 2.0: Next-Generation Foundation Model ByteDance Seed Team [overview]
2025 Seed-OSS: Open-Source Foundation Models ByteDance Seed Team [code] [models]
2025 Seed1.6: Adaptive Thinking and Multimodal Reasoning ByteDance Seed Team [overview] [tech blog]
2025 Seed1.5-Thinking: Advancing Superb Reasoning Models with Reinforcement Learning ByteDance Seed Team [arXiv] [code]

Selected Publications

Full list on Google Scholar · * indicates equal contribution

2025 Context Steering: Controllable Personalization at Inference Time Jerry Zhi-Yang He, Sashrika Pandey, Mariah L. Schrum, Anca Dragan ICLR 2025
2023 Quantifying Assistive Robustness Via the Natural-Adversarial Frontier Jerry Zhi-Yang He, Daniel S. Brown, Zackory Erickson, Anca Dragan CoRL 2023
2023 Causal Confusion and Reward Misidentification in Preference-Based Reward Learning Jeremy Tien, Jerry Zhi-Yang He, Zackory Erickson, Anca Dragan, Daniel S. Brown ICLR 2023
2022 Learning Representations that Enable Generalization in Assistive Tasks Jerry Zhi-Yang He, Zackory Erickson, Daniel S. Brown, Aditi Raghunathan, Anca Dragan CoRL 2022
2021 Assisted Robust Reward Design Jerry Zhi-Yang He, Anca Dragan CoRL 2021
2018 Gibson Env: Real-World Perception for Embodied Agents Fei Xia*, Amir Zamir*, Zhi-Yang He*, Alexander Sax, Silvio Savarese, Jitendra Malik CVPR 2018 (Spotlight) [website] [code] [paper]

Selected Projects

Gibson Environment Co-led the development of a simulation platform for real-world active perception. Recipient of the 2018 NVIDIA Pioneering Research Award.
ThingPedia Open-source platform for a personalized Internet of Things, with Prof. Monica Lam.