Edward Sun
Portrait
Email / Scholar / GitHub

I am a third-year computer science student at UCLA. Most recently, I was a Safety Research Fellow at A\, where I worked on robotics alignment and safety.

At UCLA, I work closely with Yijia Xiao and am advised by Professor Wei Wang and Professor Yuchen Cui. I also collaborate with Yuchen Wu at the University of Washington and Professor Jindong Wang at William & Mary.

I am interested in the safety, robustness, and interpretability of foundation models. I am also interested in robotics and making embodied systems more capable, reliable, and safe in real-world deployment.

Previously, I was a SPAR research fellow working on interpretable language model updates with Professor Aaron Mueller. Before that, I worked on multimodal LLMs, reasoning, and agentic systems with applications in biology and finance, and co-founded the open-source research community Tauric Research with Yijia Xiao.

I was named a Goldwater Scholar in 2025.

Highlights

When Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMs

Edward Sun*, Yuchen Wu*, Zixian Ma, Eric Hanchen Jiang, Yijia Xiao, Xiaoyuan Yi, Ranjay Krishna, Wei Wang, Jindong Wang, Aylin Caliskan

COLM 2026

Studies personalized safety in vision-language models, where a generally reasonable response may be unsafe for a specific user whose medical, emotional, or situational context is hidden. Introduces MPS-Bench, a benchmark of 5,181 scenarios from 584 real-world images across 12 high-risk domains, and finds frontier VLMs almost always answer directly (86–99%) instead of seeking missing context. Identifies visual dominance—visual affect enters the text stream in early layers and suppresses textual risk signals during fusion—and proposes PRISM, a lightweight input monitor using bidirectional cross-modal modulation to predict when a query requires deferral, achieving 0.978 AUC and dominating the safety-utility Pareto frontier.

Meta-Ctrl: Guaranteed Plan Generation by Decoupling Syntactic and Semantic Constraints

Gwen Yidou-Weng*, Edward Sun*, Tianyi Ma, Metin Alp Dogan, Benjie Wang, Allen Peng, Guy Van den Broeck, Yuchen Cui

CoRL 2026

A constrained-decoding framework that guarantees LLM-generated robot plans satisfy syntactic and semantic constraints while preserving the base LM's plan quality. Meta-Ctrl introduces meta-tokens—a compact vocabulary of grounded actions—enforcing syntax at the token level and semantics (preconditions, goals, ordering) at the action level, an exact factorization that cuts constrained-decoding memory from over 107 TB to under 2 GB. Lifts small open-weight LMs above frontier models like GPT-4o and o1-preview on embodied planning benchmarks, and is demonstrated on a real tabletop robot. Project page

Visible Touch: Rendering Contact for Visuomotor Policies

Metin Alp Dogan*, Edward Sun*, Feng Xu*, Daniel Wu, Allen Peng, Dennis Hong, Yuchen Cui

CoRL 2026

Integrates contact information into visuomotor policies without specialized tactile encoders or architectural changes. The key insight is that the bottleneck is not the contact signal itself but how it is delivered: rendering contact into the same spatial frame the policy already attends to makes it directly usable by any image-conditioned policy, including pretrained VLAs. Paired with an open-source, low-cost magnetic contact sensor fabricated from off-the-shelf parts via a parametric CAD-to-mold pipeline. On LIBERO, Visible Touch improves BC-Transformer success by 15.7 points and miniVLA by 25 points on average; fine-tuning π0.5 on four real-world contact-rich tasks gains 30 points. Project page

Personalized Safety in LLMs: A Benchmark and A Planning-Based Agent Approach

Yuchen Wu, Edward Sun, Kaijie Zhu, Jianxun Lian, Jose Hernandez-Orallo, Aylin Caliskan, Jindong Wang

NeurIPS 2025

Introduced the need to study personalized safety in LLMs. Argued that alignment should not be purely global, but tailored to a user's background since risk profiles vary across users. Showed where current models fall short, and introduced an inference-time mitigation approach using LLM-guided Monte Carlo Tree Search.

Created: 2026-09-16 Wed 05:39

Emacs 29.3 (Org mode 9.6.15)