Siwei Wu

I am a second-year PhD student in the Manchester NLP group at the University of Manchester, supervised by Prof. Chenghua Lin. Since October 2025, I have been interning with the foundation model team at iQuest, where I work on Agentic Post-training. Before that, I interned at the Institute for Intelligent Computing (通义千问) of Alibaba (2024.2-2024.8). I obtained my bachelor's degree from Jiangnan University and my master's degree from Nanjing University of Science and Technology.

My research interests mainly focus on Code Agentic, Agentic Everything, and natural language processing (NLP).

Email  /  Google Scholar  /  Twitter  /  GitHub

• 2026-03-17 – Our paper (IQuest-Coder-V1 Technical Report) was released on arXiv!

• 2026 – Our paper (MMRA) was accepted to Findings of EACL 2026!

• 2026-03 – Our paper (COIG-P) was accepted to Findings of EACL 2026!

• 2026-02-03 – Our paper (TerminalTraj) was released on arXiv!

• 2024-10-18 – Our paper (A Comparative Study on Reasoning Patterns of OpenAI's o1 Model) was released on arXiv!

• 2024-09-13 – I started my PhD in the NLP group at the University of Manchester!

• 2024-05-17 – Our paper (SciMMIR) was accepted by ACL'24!

• 2024-02-18 – I intern at the Institute for Intelligent Computing (通义千问) of Alibaba!

• 2023-07-25 – Our paper (Dense-ATOMIC) achieved the ACL'23 Outstanding Paper Award!

• 2023-05-25 – Our paper (Dense-ATOMIC) was nominated for ACL'23 Best Paper!

• 2023-05-04 – 2 papers (Dense-ATOMIC and CPNC) were accepted by ACL'23!

• 2023-4-05 – Created this homepage.

TerminalTraj presents a large-scale pipeline for constructing dockerized environments, generating terminal agent trajectories, and verifying them with executable checks.

We introduce COIG-P, a large-scale Chinese preference dataset designed for alignment with human values, covering diverse domains including chat, coding, and mathematics.

In this technical report, I was responsible for the construction of TerminalBench-related data.

We introduce DocMMIR, a unified framework for document-level multi-modal information retrieval across Wikipedia articles, scientific papers, and presentation slides.

In our work, to investigate the reasoning patterns of o1, we compare o1 with existing Test-time Compute methods (BoN, Step-wise BoN, Agent Workflow, and Self-Refine) by using OpenAI's GPT-4o as a backbone on general reasoning benchmarks in three domains (i.e., math, code and commonsense reasoning).

we first design a pipeline by optimizing existing tools and annotating numerous lyric-note pairs of songs. Then, based on the annotated data, we train a unified SongTrans model that can directly transcribe lyrics and notes while aligning them simultaneously, without requiring pre-processing songs.

We propose the multi-image relation association task and a meticulously curated Multi-granularity Multi-image Relational Association (MMRA) benchmark, comprising 1,024 samples. In order to systematically and comprehensively evaluate current LVLMs, we establish an associational relation system among images that contain 11 subtasks (e.g, UsageSimilarity, SubEvent, etc.) at two granularity levels (i.e., “image” and “entity”) according to the relations in ConceptNet.

To bridge this information retrieval gap in the scientific domain, this work develops a specialised scientific MMIR (SciMMIR) benchmark by leveraging open-access paper collections to extract data relevant to the scientific domain.

This work mainly solves the bipartite graph properties of commonsense knowledge graph ATOMIC, mines potential multiple paths in ATOMIC, and builds a more complete knowledge graph Dense-ATOMIC.

This work alleviates the problems of edge sparsity and nodes redundancy in the commonsense knowledge graph, and proposes a new commonsense knowledge graph completion framework.

Last modified in Apr. 2026. Design and source code from Jon Barron.