Tianren Zhang  |  章天任

I am interested in two intertwined areas of machine learning: representation learning and generalization. I believe that a good representation is all we need for human-level generalization and is key to achieving general intelligence.

My recent research primarily focuses on the theoretical foundations of learning world representations and out-of-distribution generalization, as well as translating these theoretical principles into efficient algorithms to solve real-world problems. I also work on reinforcement learning.

• When Do Neural Networks Learn World Models?
  Tianren Zhang, Guanyu Chen, Feng Chen
  ICML 2025  |  ICLR 2025 Workshop on World Models (Oral, Outstanding Paper Award)  |  paper
  TL;DR: We prove that in a multi-task setting, prediction models with a low-degree bias can provably identify latent data-generating variables (i.e., learning the world model) under mild assumptions.

• Exploring the Hidden Reasoning Process of Large Language Models by Misleading Them
  Guanyu Chen*, Peiyang Wang*, ..., Tianren Zhang^†, Feng Chen^†
  EMNLP 2025 (Oral)  |  paper
  TL;DR: We empirically demonstrate that LLMs can generalize unseen, false mathematical reasoning rules to real-world problems, implying the existence of an "abstract-then-reason" process in LLMs.

• OURO: A Self-Bootstrapped Framework for Enhancing Multimodal Scene Understanding
  Tianrun Xu*, Guanyu Chen*, ..., Tianren Zhang, Haichuan Gao^†, Feng Chen^†
  ICCV 2025
  

• Feature Contamination: Neural Networks Learn Uncorrelated Features and Fail to Generalize
  Tianren Zhang*, Chujie Zhao*, Yizhou Jiang, Feng Chen
  ICML 2024  |  paper  |  poster  |  code
  TL;DR: We identify that neural networks can learn task-irrelevant features due to an implicit bias of SGD, resulting in a failure to generalize under distribution shifts.

• Spatio-Temporal Approximation: A Training-Free SNN Conversion for Transformers
  Yizhou Jiang*, Kunlin Hu*, Tianren Zhang, Haichuan Gao, Yuqian Liu, Ying Fang^†, Feng Chen^†
  ICLR 2024  |  paper  |  code
  TL;DR: We propose the first training-free method for converting transformers to purely event-driven spiking neural networks.

• M³PL: Identifying and Exploiting View Bias of Prompt Learning
  Chujie Zhao*, Tianren Zhang*, Guanyu Chen, Yizhou Jiang, Feng Chen
  TMLR 2024  |  paper  |  code
  TL;DR: We identify a view bias in prompt learning of foundation models, i.e., it may extract only a partial subset of useful features while ignoring others, and we provide an effective fix.

• Fast Counterfactual Inference for History-Based Reinforcement Learning
  Haichuan Gao, Tianren Zhang, Zhile Yang, Yuqing Guo, Jinsheng Ren, Shangqi Guo^†, Feng Chen^†
  AAAI 2023  |  paper
  TL;DR: We propose a tree-based counterfactual inference method for learning history representations in reinforcement learning.

• Adjacency Constraint for Efficient Hierarchical Reinforcement Learning
  Tianren Zhang*, Shangqi Guo*^†, Tian Tan, Xiaolin Hu, Feng Chen^†
  TPAMI 2022  |  paper

• A Method of Supervised Learning from Conflicting Data with Hidden Contexts
  Tianren Zhang, Yizhou Jiang, Feng Chen
  arXiv preprint  |  paper
  TL;DR: A formulation and a theoretically grounded method for the problem of open-ended training on data with hidden contexts.

• Subjective Learning for Conflicting Data
  Tianren Zhang, Yizhou Jiang, Xin Su, Shangqi Guo, Feng Chen
  ICLR 2022 Workshop on Agent Learning in Open-Endedness  |  paper
  TL;DR: An initial attempt of formulating and addressing the problem of data conflicts in open-ended learning.

• CRIL: Continual Robot Imitation Learning via Generative and Prediction Model
  Chongkai Gao, Haichuan Gao, Shangqi Guo, Tianren Zhang, Feng Chen
  IROS 2021  |  paper  |  code
  TL;DR: A continual imitation learning method for robot learning based on generation and prediction.

• Generating Adjacency-Constrained Subgoals for Hierarchical Reinforcement Learning
  Tianren Zhang*, Shangqi Guo*, Tian Tan, Xiaolin Hu^†, Feng Chen^†
  NeurIPS 2020 (Spotlight)  |  paper  |  code
  TL;DR: We show that a state representation based on state adjacency can significantly improve the sample efficiency of hierarchical reinforcement learning.

• Deep Meta Metric Learning
  Guangyi Chen, Tianren Zhang, Jiwen Lu, Jie Zhou
  ICCV 2019  |  paper  |  code

• Conference Reviewer
  ICML (2025), NeurIPS (2024), ICLR (2025), AISTATS (2025), ACM MM (2024), CVPR (2024 - 2025), ICCV (2025),

• Journal Reviewer
  TPAMI, Artificial Intelligence Review