Jiale Liu
I am a Ph.D. student in Informatics at Penn State University, advised by Qingyun Wu, and a co-creator and maintainer of AutoGen. In 2026, I was a Student Researcher at Google DeepMind in New York, working on continual learning and memory mechanisms for LLM agents that learn from experience. In Summer 2025, I interned at Adobe Research, where I developed multi-agent systems for open-ended collaborative and creative workflows. Before Penn State, I obtained my B.Eng. in Computer Science from Xidian University, ranking first in my class of 134.
My research develops scalable frameworks and algorithms that enable LLM agents to reason, collaborate, and self-improve in complex, open-ended environments. In particular, I study how to debug and optimize agentic behavior and how to form effective multi-agent teams.
I am on the job market and looking for full-time roles starting in 2027. If you are hiring, I would love to hear from you. jiale.liu@psu.edu
Experience
-
May – Oct 2026
Google DeepMind Student Researcher, New York, NY
Continual learning and memory for LLM agents that learn from experience.
Mentors: Zun Li, John Schultz
-
May – Aug 2025
Adobe Research Research Scientist Intern, San Jose, CA
Multi-agent intelligence for creative support.
Mentors: Victor S. Bursztyn, Saayan Mitra
Publications
* equal contribution-
arXiv 2026
Learning In-Context Adaptivity for LLMs through Continual Gameplay
TL;DR We post-train LLMs with trajectory-level RL to excel at learning from memory in repeated poker—our 9B policy boosts experience gain up to 5x over SFT on unseen opponents and matches or surpasses frontier models across 35 opponents.
-
NeurIPS E&D 2026
Who&When Pro: Can LLMs Really Attribute Failures in AI Agents?
TL;DR We build a 12,326-trajectory multimodal benchmark across 26 datasets with controlled failure injection, showing frontier LLMs frequently confuse failure symptoms with root causes.
-
ACL 2026
TeamFusion: Supporting Open-ended Teamwork with Multi-Agent Systems
TL;DR We propose a multi-agent framework that pairs each team member with a preference-conditioned proxy agent, surfacing disagreements through structured discussion to synthesize higher-consensus creative deliverables.
-
ICML 2026
Digital Agents Require Unified Agent-Native Environments
TL;DR We propose agent-native computers and build AgentVM, unifying GUI and programmatic text actions over shared OS state for general-purpose digital agents.
-
EMNLP Findings 2025
Divide, Optimize, Merge: Fine-Grained LLM Agent Optimization at Scale
TL;DR We propose a divide-optimize-merge framework for scalable LLM agent optimization, improving benchmark accuracy while cutting prompt token consumption by 56%.
-
EMNLP 2025
SimpleDoc: Multi-Modal Document Understanding with Dual-Cue Page Retrieval and Iterative Refinement
TL;DR We build a lightweight multimodal document QA framework combining dual-cue page retrieval with iterative VLM working-memory refinement, boosting accuracy while reading fewer pages.
-
ICML 2025 Spotlight
TL;DR We are the first to formalize automated failure attribution for multi-agent systems with an annotated benchmark spanning 127 systems, exposing how poorly frontier reasoning models localize decisive error steps.
-
COLM Workshop 2024
Adaptive In-conversation Team Building for Language Model Agents
TL;DR We build Captain Agent, which dynamically recruits and reflects on nested multi-agent teams per subtask, improving accuracy by 21.9% without task-specific prompt engineering.
-
COLM 2024 ICLR Workshop Best Paper Award
Autogen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation Framework
TL;DR We build AutoGen, an open-source multi-agent conversation framework enabling customizable, conversable LLM agents across coding, research, and tool-use workflows.
-
ICML 2024 Spotlight
Refined Coreset Selection: Towards Minimal Coreset Size under Model Performance Constraints
TL;DR We are the first to formalize minimal-size coreset selection under explicit model performance constraints, providing an optimization algorithm with theoretical convergence guarantees.
-
ICML 2024
Offline Training of Language Model Agents with Functions as Learnable Weights
TL;DR We propose treating external agent tools as learnable parameters via AgentOptimizer, iteratively forging and refining functions from execution traces without updating LLM weights.
-
ICLR 2024
IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models
TL;DR We propose framing selective annotation for in-context learning as graph influence maximization, selecting high-coverage demonstrations with lower annotation cost and submodular guarantees.
-
NeurIPS Workshop 2024
Embodied LLM Agents Learn to Cooperate in Organized Teams
TL;DR We propose prompt-based organizational structures and a Criticize-Reflect evolution loop for embodied multi-agent teams, cutting communication overhead and boosting efficiency by up to 30%.
-
Nexus 2024
TL;DR We build a systematic trustworthiness benchmark exposing reliability and robustness challenges of deep learning models for continental-scale water quality prediction.
-
J-BHI 2024
RLAD: A Reliable Hippo-guided Multi-task Model for Alzheimer's Disease Diagnosis
TL;DR We propose a hippocampus-guided multi-task deep learning framework for reliable Alzheimer's disease diagnosis from neuroimaging.
-
ICLR 2023
Moderate Coreset: A Universal Method of Data Selection for Real-world Data Analysis
TL;DR We propose selecting training examples near the median distance to class centers, yielding a simple and distribution-robust coreset baseline across diverse real-world settings.
Open Source
-
Co-creator, Maintainer
AutoGen microsoft/autogen
An open-source framework enabling next-gen LLM applications via multi-agent conversation.
-
Co-creator, Maintainer
AG2 ag2ai/ag2
The evolution of AutoGen with integrated adaptive team building for LLM agents.
-
Maintainer
FLAML microsoft/FLAML
A fast library for automated machine learning and tuning.