| |
|
| |
Join Workato's AI Journal Club series-we're bringing together the best AI researchers to share papers & exchange perspectives on how AI research is shaping real-world systems.
Who Should Attend
AI Researchers & practitioners working at the intersection of AI research & real-world systems.
Schedule
5:30-6:00 PM: Check-in & registration
6:00-6:15 PM: Welcome to Workato
6:15-6:45 PM: Talk by Allen Chuang (Google)
6:45-7:00 PM: Q&A
7:00-7:30 PM: Talk by Ray Liu (NVIDIA)
7:30-7:45 PM: Q&A
7:45-8:30 PM: Networking
Please arrive by 6:00 PM. We politely ask that attendees arrive by this time out of respect for our speaker.
Sessions
Allen Chuang - Taming the Dynamics of LLM Reasoning: Test-Time Exploration, Data Scheduling, & On-Policy Distillation
Scaling reasoning capabilities in Large Language Models (LLMs) requires navigating complex optimization & search dynamics across both inference & post-training. However, modern reasoning pipelines suffer from critical inefficiencies throughout the model lifecycle: test-time decoding wastes compute on redundant search trajectories, reinforcement learning (RL) struggles with uniform data pacing, & on-policy distillation frequently undergoes catastrophic truncation collapse driven by length inflation. In this talk, I will present a unified perspective on understanding & taming reasoning dynamics. First, we examine inference-time search, demonstrating how Decoding Tree Sketching (DTS) enables structured exploration & early termination at decision tokens to eliminate redundant rollouts without retraining. Next, we turn to post-training optimization, showing how Adaptive Data Scheduling (ADS) leverages semantic clustering & policy-boundary selection to dynamically pace LLM RL curricula. Finally, we analyze the failure modes of On-Policy Distillation & introduce StableOPD to suppress trajectory explosion & restore distillation stability. Together, these methods provide practical mechanisms for building efficient, robust, & scalable LLM reasoning pipelines.
Ray Liu - Nondeterminism in LLM Inference & Training-Rollout Mismatch
LLM generation is not deterministic even when the temperature is set to zero. System-level configuration changes, such as variations in batch size & parallel strategy, which commonly occur in real-world serving due to continuous batching. This issue is more pronounced in RL, where the training & rollout engines naturally operate with different batch sizes, kernel selections, & parallelization strategies. This training-rollout mismatch problem leads to suboptimal performance & training collapse, specifically for the MoE model. In this talk, he will analyze why this happens & how to solve the problem at the system level by building deterministic GPU kernels.
Featured Speakers
Allen Chuang - Yu-Neng (Allen) Chuang is a Research Scientist at Google DeepMind. He works on building reliable & efficient LLM agentic systems through continued pre-training & post-training. He received his Ph.D. in Computer Science from Rice University. His research focuses on LLM reasoning, post-training, & agentic systems, with the goal of developing scalable & reliable AI systems for real-world applications. His work has been published at leading AI & machine learning venues, including ICML, ICLR, & NeurIPS, & has received several recognitions, including an ICML Spotlight, a CIKM Best Demo Paper Honorable Mention, & a NAACL Best Paper Award nomination.
Ray Liu - Zirui (Ray) Liu is an Assistant Professor of Computer Science at the University of Minnesota. His interests lie in the broad area of LLM & MLSys. He regularly published papers in top venues such as NeurIPS, ICML, ICLR, & MLSys. His work has been integrated into widely used NLP tools like Llama.cpp, SGLang [in progress], & Huggingface Transformers, & was highlighted at Google I/O sessions.
Host
About Workato
Workato is the Enterprise MCP company, providing the connective layer that gives AI agents secure, governed access to enterprise systems & data. Built on a decade of integration expertise spanning 14,000+ applications, Workato's platform enables organizations to move from simple automation to agentic AI that can reason, act, & orchestrate work across the entire business. You can explore Workato's end-to-end capabilities in our developer sandbox here.
|
|
|
|
|
|
|
|
