Zhenwei (Joseph) Tang

My research develops scalable and trustworthy machine-learning methods for modeling human behavior. I use large-scale behavioral data to study how decisions vary across people and contexts, from chess to industrial ranking and LLM-powered recommendation, with an emphasis on reliable evaluation in real-world settings.

Aug 2026: Decomposing Staleness in Recommender Systems, our Google Discover work on filtering stale recommendations caused by supersession and relevance decay, was accepted to the CIKM 2026 Applied Research Track. The system was validated in production.
Aug 2026: SafeGEO, on how generative engine optimization can manipulate LLM-powered recommendation agents, was accepted at EMNLP 2026. My first last-author paper! See the website, code, and data.
Jul 2026: Heterogeneous Ranking in Industrial-Scale Recommender Systems, our Google Discover case study of heterogeneous feed ranking, was accepted to the RecSys 2026 Industry Track. The approach was validated through production A/B testing.
Jun 2026: I passed my thesis proposal!
Jun 2026: I started as a Student Researcher at Google.
May 2026: New preprint from my internship at Layer 6 AI: RankJudge: A Multi-Turn LLM-as-a-Judge Synthetic Benchmark Generator. See the leaderboard, code, and data.
May 2026: SWE-Bench Mobile, our benchmark asking whether LLM agents can develop industry-level mobile apps, was accepted at KDD 2026. See the project website and code.
May 2026: MINER: Mining Multimodal Internal Representation for Efficient Retrieval is now on arXiv.
May 2026: We released Generate, Filter, Control, Replay, a comprehensive survey of rollout strategies for LLM reinforcement learning.
Apr 2026: LLM Safety From Within, where we introduce SIREN, a lightweight guard model built on LLM internal representations, was accepted at ACL 2026.
Apr 2026: New preprint: ThinkTwice, on jointly optimizing LLMs for reasoning and self-refinement. Code and model are open source.
Mar 2026: We put Grounded Chess Reasoning in Language Models via Master Distillation on arXiv, introducing C1, a 4B model that reasons about chess with explainable solutions. Code and data are open source.
Mar 2026: New preprint: Level Up, where we define and exploit transitional problems for curriculum learning.
Feb 2026: I started a research internship at Layer 6 AI in Toronto.
Jan 2026: Chessformer: A Unified Architecture for Chess Modeling, the architecture behind Maia-3, our new state of the art in human move prediction, was accepted at ICLR 2026! Weights are on Hugging Face, and you can play against Maia-3 on Maia Chess.
Dec 2025: Learning to Imitate with Less, where we introduce Maia4All for efficient individual behavior modeling in chess, was accepted by TMLR.
Oct 2025: New preprint: ChessQA, a comprehensive evaluation of how well LLMs actually understand chess. Code is open source.
Aug 2025: SEAM, our benchmark testing whether vision-language models reason consistently across semantically equivalent inputs, was accepted at COLM 2025. See you in Montreal! Check out the website, code, and data.
Jul 2025: Maia Chess is now in open beta! Powered by Maia-2, you can play against the most human-like chess AI, analyze your games with human-aware AI, and train with Maia as your personal coach.
Jul 2025: I passed the Thesis Topic Approval checkpoint and am now officially a Ph.D. candidate!
Apr 2025: Two extended abstracts on human behavior modeling in chess were accepted at IC2S2 2025.
Dec 2024: I passed my Ph.D. qualifying exam.
Sep 2024: Maia-2: A Unified Model for Human-AI Alignment in Chess was accepted at NeurIPS 2024. See you in Vancouver, and in the meantime you can play against it on Lichess.
May 2024: SPIN, on sparsifying and integrating internal neurons in LLMs, was accepted to ACL 2024 Findings. Try the interactive demo and code.
Apr 2024: Two extended abstracts were accepted at IC2S2 2024. See you in Philly!
Mar 2024: Our paper Predicting Protein Functions Using Positive-Unlabeled Ranking with Ontology-based Priors (PU-GO) was published in Bioinformatics as part of the ISMB 2024 proceedings. Code and data are open source.
Nov 2023: LQAC received an Honorable Mention for the Best Paper Award at ISWC 2023!
Sep 2023: I received an ISWC 2023 Travel Award. See you in Athens!
Jun 2023: Customized Conversational Recommender Systems (CCRS) was accepted at ECML-PKDD 2023.
Jun 2023: Our paper Neural Multi-hop Logical Query Answering with Concept-level Answers (LQAC) was accepted at ISWC 2023.
May 2023: DiffuDetox, a mixed diffusion model for text detoxification, was accepted to ACL 2023 Findings.
Apr 2023: Two papers, LogicRec and Bayesian Knowledge-driven Critiquing with Indirect Evidence, were accepted at SIGIR 2023.
Nov 2022: I joined the Vector Institute as a Faculty Affiliated Researcher.
Nov 2022: Our paper Self-Supervised Learning for Conversational Recommendation was accepted by Information Processing & Management.
Sep 2022: I started my Ph.D. in Computer Science at the University of Toronto.
Aug 2022: I completed my M.S. in Computer Science at KAUST.
Jun 2022: I joined the L3S Research Center in Hannover as a visiting student.
Apr 2022: Our paper Positive-Unlabeled Learning with Adversarial Data Augmentation for Knowledge Graph Completion (PUDA) was accepted at IJCAI 2022.
Oct 2021: Personalized Transfer of User Preferences for Cross-domain Recommendation (PTUPCDR), work from my internship at the Institute of Computing Technology, CAS, was accepted at WSDM 2022.
Jan 2021: I started my M.S. in Computer Science at KAUST.
Jul 2020: I joined the Institute of Computing Technology, CAS as a research intern.
Jul 2020: I completed my B.S. in Telecommunication Engineering at BUPT.
May 2020: Intelligent Exploration for User Interface Modules of Mobile App with Collective Learning, work from my internship at Baidu Research, was accepted at KDD 2020. My first publication!
Oct 2019: I joined Baidu Research as a research intern.
Sep 2016: I started my B.S. in Telecommunication Engineering at BUPT.