About Me
I’m Wenyue Hua, a Senior Researcher at Microsoft Research, AI Frontiers.
My work runs across three lines. Strategic and social decision-making in LLM agents, from simulation to training: multi-agent simulation of strategic conflict (WarAgent), game-theoretic workflows that make LLM negotiators more rational (Game-theoretic LLM), and reinforcement learning that lets a 4B model match or exceed the GPT-5 family in-domain across six negotiation domains (SocialRL). Trustworthiness: TrustAgent, the Agentic Risk Standard, and EmojiPrompt. Efficiency: Interactive Speculative Planning, Dynamic Speculative Agent Planning, and AgentOpt. Earlier, I worked on generative recommendation (How to Index Item IDs) and reasoning evaluation (NPHardEval, InductionBench, OpenAGI).
Before Microsoft, I was a postdoctoral researcher at UC Santa Barbara with Prof. William Yang Wang (2024 to 2025). I received my Ph.D. in Computer Science from Rutgers University (2020 to 2024), advised by Prof. Yongfeng Zhang. I also hold an MA in Linguistics from Rutgers (advised by Prof. Adam Jardine) and a BS in Mathematics and a BA in Linguistics and Philosophy from UCLA (advised by Prof. Edward Keenan).
I founded NICE (2023), an AI/NLP research community that hosts talks on new papers and projects. If you have work you would like to present, contact nice.ai.academy@gmail.com.
- Strategic and social reasoning in LLM agents
- RL post-training for agents
- Trustworthy agents: safety, risk, privacy
- Efficient agent systems
-
Postdoctoral Research in Computer Science, 2024-2025
Computer Science Department, University of California, Santa Barbara
-
Ph.D. in Computer Science, 2020-2024
Computer Science Department, Rutgers University, New Brunswick
-
Master of Arts in Linguistics (Ph.D. track transfer out), 2018-2020
Department of Linguistics, Rutgers University, New Brunswick
-
B.S. in Mathematics, General & B.A. in Linguistics&Philosophy with Specialization in Computing, 2014-2018
UCLA
Selected Works
Papers I led as first, co-first, or senior author. Full list on the Publications page.
Strategic and social decision-making
SocialRL: From Passive Delegates to Strategic Negotiators (2026, arXiv 2608.13787). A general RL recipe that trains social reasoning directly. A 4B model matches or exceeds the GPT-5 family in-domain across six negotiation domains. PDF
Game-theoretic LLM: Agent Workflow for Negotiation Games (2024, arXiv 2411.05990). Shows where LLMs depart from rational play in complete- and incomplete-information games, and designs game-theoretic workflows that steer them toward equilibrium and better negotiation outcomes. PDF
WarAgent (2023, arXiv 2311.17227). LLM multi-agent simulation of WWI, WWII, and the Warring States period, used to study the triggers and conditions that lead to war. PDF
Trustworthy agents
TrustAgent (2024, Findings of EMNLP). An agent-constitution framework with pre-planning, in-planning, and post-planning safety strategies. PDF
Quantifying Trust: the Agentic Risk Standard (2026, arXiv 2604.03976). A settlement-layer standard that integrates risk assessment, underwriting, and compensation for AI-mediated transactions. Spotlighted in Fortune. PDF
EmojiPrompt (2025, NAACL). Generative prompt obfuscation so cloud LLMs can complete tasks without seeing raw private content. Co-first author. PDF
Efficient agent systems
Interactive Speculative Planning (2025, ICLR). Speculative execution for agent planning, co-designed with a user interface that treats human interruption as a first-class part of the system rather than an exception. PDF
Dynamic Speculative Agent Planning (2026, ICLR). An asynchronous online RL framework for lossless acceleration of agent planning, exposing a single parameter that trades latency against dollar cost and cutting total cost by 30%. Senior author. PDF
AgentOpt (2026, technical report, arXiv 2604.06296). Client-side model selection for agent pipelines. The cost gap between the best and worst model combinations reaches 13–32x, and Arm Elimination cuts evaluation budget by 24–67% at near-optimal accuracy. Open-source package. PDF · Code
Foundations: generative recommendation and reasoning evaluation
How to Index Item IDs for Recommendation Foundation Models (2023, SIGIR-AP). Systematic study of item ID construction for LLM-based generative recommendation, with sequential, collaborative, semantic, and hybrid indexing. PDF
NPHardEval (2024, ACL). A dynamic reasoning benchmark organized by computational complexity class and refreshed regularly to resist overfitting. Co-first author. PDF
InductionBench (2025, ACL). A benchmark for inductive reasoning, inferring the underlying rule from observations rather than applying a given one. Frontier models fail even the simplest complexity classes of the subregular hierarchy. PDF · Code
News
Honors, Press, and Invited Talks
Honors
- KAUST AI Rising Star, 2025 (news)
- National Science Foundation SBIR research funding ($50,000), 2021
- Phi Beta Kappa, UCLA College and Departmental Honors, 2018
Press
- “Quantifying Trust” spotlighted in Fortune, April 2026, with commentary on systemic financial risk from AI agents (news)
Invited talks
- From Philosophy of Language to AI Agents: Fermilab (March 2026), MBZUAI and UC Santa Barbara (June 2026) (news)
- Agentic Risk Standard: Columbia University (May 2026), Carnegie Mellon University (May 2026), UC Santa Barbara (April 2026)
- Magentic Marketplace: Columbia Agent Workshop (October 2025), RecSys 2025 EARL Workshop (September 2025)
Service
- Co-organizer, ICLR 2026 Workshop on Memory for LLM-Based Agentic Systems (MemAgents) and the 2nd COLM 2026 Workshop on Lifelong Agents
- Area Chair, ACL, EMNLP, NAACL, EACL, COLM (2025 to 2026)
Experience
- Advisor: Prof. William Yang Wang
- Dissertation: Trustworthy Large Language Model
- Advisor: Prof. Yongfeng Zhang
- Thesis: Learning Underlying Representations and Input-Strictly-Local Functions
- Advisor: Prof. Adam Jardine
- Thesis: Boolean-Algebraic Representation of Possible Worlds
- Advisor: Prof. Edward Keenan