Yeonju Ro

@j777ro

UT Austin PhD Student @UTCompSci @utnslab @VITAGroupUT

Austin, Texas sites.google.com/view/hey-yeonju Joined July 2022

184 Following 76 Followers 16 Posts

Latest posts 4

Opus 4.8 was released today, and it features dynamic workflows where the planner constructs a workflow for a given task. “Some problems are too big for one pass by a single agent,” Claude says. For dynamic workflows, how can we identify the most important nodes and verify them to prevent error propagation? Take a look at our recent work Sherlock 🕵🏻‍♀️🔍 arxiv.org/abs/2511.00330

ClaudeDevs Verified account @ClaudeDevs May 28

New in Claude Code (research preview): dynamic workflows.

Claude writes an orchestration script on the fly, then spins up a large fleet of coordinated subagents in parallel to take on your most complex tasks.

Use the word "workflow" in a prompt to get started.

View quoted post on X

Yeonju Ro @j777ro May 27

A deployed agent is not just a model. It is a stateful system: its behavior depends not only on model weights, but also on accumulated memory, retrieved context, and operational events like updates or maintenance.

Our recent work, ⏳AgingBench⏳, moves agent evaluation beyond snapshot accuracy toward reliability over time. It makes these hidden state changes measurable, which is exactly the kind of test we need if agents are going to be operated, debugged, and repaired as production systems.

Take a look and try it on your system. I’m around at #CAIS this week and happy to chat in person!

Jianing Zhu Verified account @Jianing9810 May 27

AI agents are increasingly deployed as persistent operational systems, but do they remain reliable over time?

Unfortunately no, our new work shows agents can quietly fail after deployment, despite passing day-1 evaluation. We call this "agent aging", akin to human aging.

View quoted post on X

Yeonju Ro @j777ro May 14, 2025

@divytweet will present our take on OS Guardrails today! 🚧
ML has made its way into everything—even the OS.
How do we keep it in check?
Come hear our proposal (and Divyanshu’s well-crafted jokes)! 👀🔥 #hotos

Divyanshu Saxena Verified account @divytweet May 14, 2025

I will be (remotely) presenting our proposal tomorrow at HotOS'25 on gaining the adaptability of learned policies for the OS, while maintaining control over them!

We introduce OS Guardrails - an abstraction for monitoring properties at run-time and taking corrective actions.

View quoted post on X

Yeonju Ro @j777ro Dec 5, 2024

(1/n) Do you think token batching in MoE is inefficient? Are you looking for ways to transform pre-trained LLMs into MoEs? Then you should check out Read-ME at NeurIPS'24! 📖

arxiv.org/abs/2410.19123

arxiv.org Read-ME: Refactorizing LLMs as Router-Decoupled Mixture of Experts... The proliferation of large language models (LLMs) has led to the adoption of Mixture-of-Experts (MoE) architectures that dynamically leverage specialized subnetworks for improved efficiency and...