About me
I am a first-year Ph.D. student in Computer Science at the University of Maryland, College Park, advised by Prof. Yaodong Yu and Prof. Tianyi Zhou. Previously, I earned my master’s degree in Computer Science at Nanjing University and my bachelor’s degree in Computer Science at Xidian University. You can find my CV here: Yu Fangxu’s Curriculum Vitae.
My recent research topics include:
(1) LLM Reasoning and Post-Training: Flow of Reasoning, W2S-OPD
(2) Multimodal LLM Reasoning and Post-Training: TS-Reasoner, TSRBench, TSRouter, ArrowGEV, AudioRubrics
(3) Emotion & Social Reasoning: EACL, ECGN, PersuasiveToM
(4) AI Safety: COLD-Attack
(5) LLM Agent: ArcMemo, FlowBank
News
July 2026: One paper is accepted to TMLR! We propose TS-Reasoner, a new time series LLM for reasoning by aligning the Time Series Foundation Model with an LLM. In addition, we introduce a simple time series captioning approach for large-scale alignment data construction.
July 2026: One paper is accepted to COLM 2026! We propose TSRouter, a lightweight routing approach that selects the most appropriate modality to represent time series and model (e.g., LLMs, VLMs) to solve the reasoning problem.
April 2026: One paper is accepted to ICML 2026! We propose TSRBench, a large-scale, comprehensive benchmark designed to stress-test the time series understanding and reasoning capabilities of generalist models (LLMs, VLMs, and TSLLMs).
April 2026: One paper is accepted to Findings of ACL 2026! We quantitatively demonstrate the limitations of current VLMs in understanding temporal directionality, and propose an RL-based approach to simultaneously enhance both temporal video grounding and directional comprehension.
May 2025: One paper is accepted to Findings of ACL 2025! We propose an emotional contagion graph network (ECGN) inspired by the emotional contagion process in social interaction, which aims to improve emotion cause recognition by leveraging non-verbal information.
May 2025: One paper is accepted to ICML 2025! We adapt Generative Flow Networks (GFlowNets) to LLM multi-step reasoning, which aims to improve the capability of LLMs to find multiple solutions for each problem.
May 2024: One paper is accepted to ICML 2024! We adapt the Energy-based Constrained Decoding with Langevin Dynamics (COLD) to develop the COLD-Attack framework, which unifies and automates the search of adversarial LLM attacks under a variety of control requirements such as fluency, stealthiness, sentiment, and left-right coherence.
March 2024: One paper is accepted to Findings of NAACL 2024! We propose a novel contrastive learning framework to better distinguish similar emotions in Emotion Recognition in Conversation.
Sep 2021: I’m admitted to study for an M.Sc. degree in the School of Artificial Intelligence at Nanjing University without an entrance examination.