Haochen Shi ·史皓琛

Ph.D. Candidate · Université de Montréal & Mila, Montréal

I'm a graduating Ph.D. candidate at Mila and Université de Montréal, fortunate to be supervised by Bang Liu.

I study how machines reason: language agents that act in interactive worlds, and reasoning models that think in latent space.

I'm open to research roles in both industry and academia, as well as collaborations. If you would like to have a chat, please feel free to shoot me an email.

CV · PDF

Research

Latent Reasoning & Test-Time Scaling

Recent focus

Models that reason in latent space, and how to scale their compute at test time.

Input-Conditioned Search for Efficient Test-Time Scaling of Recursive Reasoners

Haochen Shi, Tianwen Fu, Qianggang Ding, Yinhan Lu, Yizhan Li, Huan Zhang, Yajie Zhao, Xingdi Yuan, Hengyang Lu, Bang Liu

Preprint · 2026

Searches over recursive reasoning functions to explore diverse reasoning traces, without any additional training. Highest mean accuracy on six reasoning tasks, with up to 100× fewer FLOPs than scaling baselines.

Paper Project page Code Weights

LLM Agents & Embodied Learning

Recent focus

Language-model agents that perceive, plan and act in embodied and open-ended worlds, and keep acquiring skills.

Evolving Programmatic Skill Networks

Haochen Shi, Xingdi Yuan, Bang Liu

COLM · 2026

A compositional network of executable skills that an agent keeps refining through experience, with LLM-driven fault localization and maturity-aware gating.

Paper Project page Code

Enhancing Agent Learning through World Dynamics Modeling

Zhiyuan Sun*, Haochen Shi*, Marc-Alexandre Côté, Glen Berseth, Xingdi Yuan, Bang Liu

Findings of EMNLP · 2024

Discover, Verify, and Evolve: the agent extracts world dynamics from a few demonstrations, verifies them, and evolves new ones for the current situation, reaching near-human reward in Crafter.

Paper Code

4 more in this topic

Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems. Bang Liu, Xinfeng Li, Jiayi Zhang, …, Haochen Shi, … (48 authors). Preprint 2025A survey that maps agent modules such as memory, world models and reward onto brain functions, then covers self-evolution, multi-agent collaboration and safety.
TradExpert: Revolutionizing Trading with Mixture of Expert LLMs. Qianggang Ding, Haochen Shi, Jiadong Guo, Bang Liu. ICLR Workshop 2025Specialized LLM experts for news, prices, alpha factors and fundamentals, combined for trading.
OPEx: A Component-Wise Analysis of LLM-Centric Agents in Embodied Instruction Following. Haochen Shi, Zhiyuan Sun, Xingdi Yuan, Marc-Alexandre Côté, Bang Liu. ACL 2024 · AAMAS 2024 extended abstractSplits an embodied agent into Observer, Planner and Executor to find which component actually limits instruction following.
Deciphering Digital Detectives: Understanding LLM Behaviors and Capabilities in Multi-Agent Mystery Games. Dekun Wu*, Haochen Shi*, Zhiyuan Sun, Bang Liu. Findings of ACL 2024 · codeA multi-agent framework and dataset for Jubensha detective games, with metrics for LLM reasoning.

Multimodal Understanding & Grounding

Aligning images and videos with language and other modalities, so that models describe, retrieve and reason about what is actually there.

Consensus Graph Representation Learning for Better Grounded Image Captioning

Wenqiao Zhang*, Haochen Shi*, Siliang Tang, Jun Xiao, Qiang Yu, Yueting Zhuang

AAAI · 2021

Learns a consensus between an image's scene graph and its caption's language graph, so the captioner names objects that are actually there. Object hallucination drops by 9%.

Paper

6 more in this topic

Robust Modality-Incomplete Anomaly Detection: A Modality-Instructive Framework with Benchmark. Bingchen Miao, Wenqiao Zhang, Juncheng Li, Wangyu Wu, Siliang Tang, Zhaocheng Li, Haochen Shi, et al. ACM MM 2025The first study of industrial anomaly detection when the RGB image or the 3D point cloud may be missing, with a benchmark and a modality-instructive two-stage framework.
MAGIC: Multimodal relAtional Graph adversarIal inferenCe for Diverse and Unpaired Text-based Image Captioning. Wenqiao Zhang*, Haochen Shi*, Jiannan Guo*, et al. AAAI 2022Diverse captions for images that contain text, learned without any image–caption pairs through adversarial inference over multimodal relational graphs.
Dilated Context Integrated Network with Cross-Modal Consensus for Temporal Emotion Localization in Videos. Juncheng Li, Junlin Xie, Linchao Zhu, Long Qian, Siliang Tang, Wenqiao Zhang, Haochen Shi, et al. ACM MM 2022Introduces temporal emotion localization: finding when each emotion occurs in an untrimmed video with subtitles, with a coarse-fine two-stream network over visual and textual cues.
BOSS: Bottom-up Cross-modal Semantic Composition with Hybrid Counterfactual Training for Robust Content-based Image Retrieval. Wenqiao Zhang, Jiannan Guo, Mengze Li, Haochen Shi, et al. Preprint 2022Finds a target image from a reference image plus text describing the change, composing visual and textual semantics bottom-up and training with hybrid counterfactual examples.
Adaptive Hierarchical Graph Reasoning with Semantic Coherence for Video-and-Language Inference. Juncheng Li, Siliang Tang, Linchao Zhu, Haochen Shi, et al. ICCV 2021Decides whether a statement is entailed by a video and its subtitles, reasoning over a hierarchical graph with semantic coherence.
Relational Graph Learning for Grounded Video Description Generation. Wenqiao Zhang, Xin Eric Wang, Siliang Tang, Haizhou Shi, Haochen Shi, et al. ACM MM 2020A language-refined scene graph supplies fine-grained objects and relations, giving more accurate grounded video descriptions.

Robust Learning & Knowledge Extraction

Learning from scarce, noisy or shifted labels, in vision and in knowledge extraction from text.

Semi-supervised Active Learning for Semi-supervised Models: Exploit Adversarial Examples with Graph-based Virtual Labels

Jiannan Guo*, Haochen Shi*, Yangyang Kang, Kun Kuang, Siliang Tang, Zhuoren Jiang, Changlong Sun, Fei Wu, Yueting Zhuang

ICCV · 2021

Label propagation over a graph gives unlabeled samples virtual labels; adversarial examples then reveal which samples sit near cluster boundaries and are worth annotating. Semi-supervised and active learning improve each other in a closed loop.

Paper

3 more in this topic

Reasoning Makes Good Annotators: An Automatic Task-specific Rules Distilling Framework for Low-resource Relation Extraction. Yilin Lu, Juncheng Li, Xiaoqiang Wang, Haochen Shi, Tao Chen, Siliang Tang. Findings of EMNLP 2023A pre-trained language model reasons out task-specific labeling rules and composes them into robust compound rules, replacing hand-written rules for low-resource relation extraction.
Empower Distantly Supervised Relation Extraction with Collaborative Adversarial Training. Tao Chen*, Haochen Shi*, Liyuan Liu, et al. AAAI 2021Recycles the training instances that multi-instance learning discards, by coordinating virtual adversarial training and adversarial training.
Alleviate Dataset Shift Problem in Fine-grained Entity Typing with Virtual Adversarial Training. Haochen Shi, Siliang Tang, Xiaotao Gu, et al. IJCAI 2020Distantly supervised training data differs systematically from the human-annotated test set; a virtual adversarial training regularizer makes fine-grained entity typing robust to this shift.

* Equal contribution

Education

2021 – PresentPh.D. Computer ScienceUniversité de Montréal / Mila
2021M.S. Computer ScienceZhejiang University
2018B.S. Software EngineeringHuazhong University of Science and Technology

Teaching & Awards

TA, Advanced Projects in Machine Learning (IFT6759)Winter 2024
TA, NLP with Deep Learning (IFT6289)Winter 2023
Technical Assistant, IVADO NLP Workshop: Knowledge GraphsWinter 2021
TA, NLP and AI courses, Zhejiang University2019 – 2021
Tianzhou-Chen Scholarship (top 1%), Zhejiang University2020
Outstanding Graduate, Zhejiang University2020
CETC Scholarship, China Electronics Technology Group Corporation2020
4th place, NIST TAC KBP 2019 Entity Discovery and Linking2019