/ˈtsɔŋ.feŋ ˈtsaʊ/

Research interests

LLMs, RAG, agents, and their integration with graph-based methods.

Research experience

Institute for Logic, Language and Computation, UvA

Promoter: Khalil Sima'an, Supervisor: Jelke Bloem

October 2022 -- Present

Research on LLMs, RAG, agents, and their integration with graph-based methods.

University of Chinese Academy of Sciences (UCAS)

Supervisor: Junyong Fang

October 2014 -- July 2017

Analyzing multi-band remote sensing images using deep learning and computer vision techniques.

Medical College, Peking University

Part-time Researcher

2018 -- 2019

Information retrieval and question answering for Chinese medicine literature.

Industry experience

Kunlun Digital Technology Co., Ltd. (CNPC Richfit)

Senior Machine Learning Researcher

January 2021 -- October 2022

Team Leader in Conversation Service Group. Responsible for building an automated customer question-answering system, leading the team, and managing projects. [Video] [Description]

Tsinghua Tongfang Co., Ltd.

Machine Learning Researcher

July 2017 -- January 2021

Team Leader in Machine Learning Group. Responsible for research on natural language processing, knowledge graphs, and text mining algorithms; applying these algorithms in projects; and leading the team while managing projects.

Selected Publications

Fusion Training for Mathematical Generalization in Large Language Models (ACL 2026 SRW)

Congfeng Cao, Pengyu Zhang, Jelke Bloem.

Contribution: Systematically studies Thinking Mode Fusion in mathematical reasoning, showing that thinking and non-thinking modes interfere with each other and that training schedule and data ratio control this trade-off.

Technical: Thinking Mode Fusion; Qwen3-4B; LoRA-based SFT; GSM8K for non-thinking data; OpenMathReasoning for thinking data; thinking/non-thinking data-ratio control; T-NT, NT-T, and Mix training schedules; exact-match evaluation on mathematical reasoning.

Graph-TempCZ: A Graph Representation of Software Mentions for Predicting Software Usage in Scientific Publications (LREC 2026)

Congfeng Cao, Pengyu Zhang, Jelke Bloem.

Contribution: Introduces Graph-TempCZ, a large-scale temporal publication--software mention graph, and formulates software usage prediction as a link prediction task.

Technical: Large-scale publication--software mention graph; Europe PMC metadata augmentation; temporal graph construction from 2010--2020; link prediction; XGBoost with node/topological features; Sentence-BERT embeddings; GraphSAGE for GNN-based link prediction and temporal transfer experiments.

How Aligned Are Unimodal Language and Graph Encodings of Chemical Molecules? (IJCNLP-AACL Main 2025)

Congfeng Cao, Zhi Zhang, Jelke Bloem, Khalil Sima'an.

Contribution: Studies whether molecular graph and language encoders are inherently aligned before force-alignment training, and introduces the U2A benchmark for evaluating graph--language alignment.

Technical: Molecular graph--language alignment benchmark; ChEBI-20 molecule-description pairs; RDKit and MAYGEN for isomers/tautomers; MolCLR, MolFormer, SciBERT, SentenceBERT, BERT, ModernBERT, GPT; Debiased CKA for alignment evaluation; structural-level alignment analysis.

Under Review

SoftwareCiter: Agentic Generation of Verifiable Software Citations with a Metadata-Aware Benchmark

Congfeng Cao, Yixian Shen, Kayle Verhiel, Jelke Bloem.

Contribution: Proposes a multi-agent framework for generating complete and verifiable software citations from incomplete software mentions in scientific papers, supported by the metadata-aware SoftwareCiteBench benchmark.

Technical: Multi-agent LLM framework; publication parsing; software metadata extraction; missing-field planning; external metadata search from web/GitHub/Zenodo/official pages; FORCE11-style citation construction; citation verification; evaluated with GPT-5-mini and LLM-as-a-Judge on SoftwareCiteBench.

Molecule Generation through Reasoning with Large Language Models

Congfeng Cao, Di Wu, Jiahong Huang, Jelke Bloem, Khalil Sima'an.

Contribution: Frames molecule generation as a step-by-step reasoning task and introduces MolCoT-4K and MolGeneration to improve SMILES generation from molecular descriptions.

Technical: Reasoning-based molecule generation; curated CoT dataset MolCoT-4K; Qwen2.5-7B fine-tuning; SFT plus GRPO-based reinforcement learning; molecule-specific rewards including exact match, SMILES validity via RDKit, and format reward; evaluation using Validity, RDK-FTS, EM, and Pass@N.

Rethinking Feature Augmentation in Graph Contrastive Learning

Congfeng Cao, Pengyu Zhang, Zeyu Zhang, Jelke Bloem.

Contribution: Introduces Random Feature Masking, showing that a simple feature augmentation strategy improves graph contrastive learning performance and robustness at high masking rates.

Technical: Graph Contrastive Learning; Random Feature Masking (RFM); comparison with Stochastic Masking Node Feature (MF); integrated into GRACE, GraphCL, and CCA-SSG; node classification on Cora, Citeseer, Pubmed, Coauthor, and Amazon datasets; uniformity analysis using KDE and variance.

A Graph-Based Analysis of Software Usage in Scientific Publications

Congfeng Cao, Na Li, Kristy James, Jelke Bloem.

Contribution: Shows that biomedical research software usage forms structured, domain-specific ecosystems, moving beyond frequency-based analysis of software mentions.

Technical: Large-scale scientometric graph analysis; CZ Software Mentions + Europe PMC metadata; Graph-CZ publication--software bipartite graph; Software-CZ software co-occurrence graph; Publication-CZ publication co-usage graph; temporal analysis, co-occurrence analysis, Louvain community detection, and cross-community interaction analysis.

Grants

NWO Compute Projects

2024, 2025, 2026

UCAS Academic Scholarship

2015, 2016, 2017

Education

University of Amsterdam

Amsterdam, Netherlands

PhD in Computer Science

October 2022 -- Present

University of Chinese Academy of Sciences

Beijing, China

MEng in Signal and Information Processing

October 2014 -- July 2017

Presentations

ACL 2026, LREC 2026, IJCNLP-AACL 2025, EMNLP 2025, CLIN 2025, CLIN 2024, ELLIS 2024.

Course & Training

Knowledge Graphs (2026), IBM RAG and Agentic AI Professional Certificate (2025), Agentic AI (2025), Fine-tuning & RL for LLMs (2025), Generative AI with Large Language Models (2025), Machine Learning and Optimisation (2024), LxMLS (2024), LOGML (2025), Human-Centred AI (2023)

Skills

Programming
Proficient in: Python, PyTorch, TensorFlow. Familiar with: C/C++, Java.
Database: SQL (MySQL, Oracle, PostgreSQL), NoSQL (Neo4J, MongoDB).

Languages
Chinese (native), English (work language), Dutch (beginner).

Visitor Statistics

Visitor count: