Xiyang Wu (On the Job Market)

@wu_xiyang

Ph.D. Student at @gammaumd @eceumd @umiacs @UofMaryland. Intern @NECLabsAmerica @Dolby. Previous: @EmoryUniversity @GeorgiaTech @TJU1895.

College Park, MD

wuxiyang1996.github.io

Joined August 2021

1,872 Following

529 Followers

Replies Reposts Media

Pinned
Xiyang Wu (On the Job Market)
@wu_xiyang
Oct 17
Excited to share BeliefWeaver, our new work on failure recovery for long-horizon agents. Instead of simply retrying after failure, BeliefWeaver retains evidence across attempts and rollback, updates beliefs over possible repairs, and uses an implicit world model to decide what
00:00
3
Xiyang Wu (On the Job Market)
@wu_xiyang
Oct 19
Excited to share ASSEMBLE, our new work on long-video reasoning with traceable evidence. Our 9B model learns to find and connect the moments that support its answer. Across 3 benchmarks, it improves grounded accuracy by 6.7 points over Gemini-2.5-Pro. 📄 Paper:
00:00
1
Xiyang Wu (On the Job Market)
@wu_xiyang
Sep 27
I’ll be at #IROS2026 this week! I’m broadly interested in foundation models, reinforcement learning, and robotics, especially reliable embodied decision-making and VLA safety. I’ll also be presenting two projects on Monday: ⚡ FALCON — Mon 9/28 9:55 AM · Room 315/316 📄
Xiyang Wu (On the Job Market)
@wu_xiyang
Apr 8
(1/n) We’re excited to share SABER: a stealthy agentic black-box attack framework for Vision-Language-Action (VLA) models. It automatically generates small, plausible instruction edits that can reliably degrade robot behavior under bounded edit budgets. 🤖⚠️ Paper:
2
Xiyang Wu (On the Job Market)
@wu_xiyang
Jun 17
🚨 Excited to share that SABER has been accepted to IROS 2026 in Pittsburgh! 🤖⚠️ SABER is a stealthy agentic black-box attack framework for red-teaming VLA/robotic foundation models through small, plausible instruction edits. 📄 Paper: arxiv.org/abs/2603.24935 💻 Code:
Xiyang Wu (On the Job Market)
@wu_xiyang
Apr 8
(1/n) We’re excited to share SABER: a stealthy agentic black-box attack framework for Vision-Language-Action (VLA) models. It automatically generates small, plausible instruction edits that can reliably degrade robot behavior under bounded edit budgets. 🤖⚠️ Paper:
Xiyang Wu (On the Job Market)
@wu_xiyang
Apr 24
(1/n) Can an LLM agent learn reusable skills from its own rollouts — without hand-written skills or expert demonstrations? Introducing COS-PLAY 🎮: a co-evolving framework where a decision agent🧠 and a skill-bank agent 🧰 improve each other for long-horizon tasks. 📄 Paper:
1