MechVQA / MechVL
An ICML 2026 benchmark with 3.3K mechanical drawings and 21K question-answer pairs, plus a domain-specialized multimodal model trained with SFT and self-play RL.
AI researcher & engineer · BAAI
Making foundation models work in specialized domains.
I build open data, post-training methods, and agentic/multimodal systems that connect research advances with demanding real-world knowledge and reasoning tasks.
让基础模型真正进入专业领域：开放数据、后训练方法与 Agent / 多模态系统。
ICML 2026
MechVQA / MechVL
AAAI 2025
CareBot
COLING 2025
MoSLD
3 BAAI collections
Led IndustryCorpus, IndustryCorpus2, and Industry Instruction
Signature research
Three connected research directions
My work spans the full path from data and post-training to systems that retrieve, reason, and operate in specialized domains.
SPAR & SciSage
Multi-agent systems for scholarly retrieval and scientific survey generation, paired with SPARBench and SurveyScope for systematic evaluation.
Open data & post-training
Multilingual industry corpora, instruction data, domain models, data-quality models, and methods for capability-preserving adaptation.
Open resources
Industry data and models
I led the development of these BAAI collections and work to release reusable datasets, models, evaluation assets, and training recipes.
IndustryCorpus2
Multilingual, multi-industry pre-training data with DataRater and classification models.
Industry Instruction
2.7M multilingual instruction samples and a family of domain-adapted models.
IndustryCorpus
Open pre-training corpora spanning finance, medicine, law, education, and technology.
Selected publications
Recent work
Complete and current bibliographic records are also available on Google Scholar, ORCID, and OpenReview.
2026
Wnuan: Staged Post-Training for Question Answering over Proprietary Enterprise KnowledgeFirst author / equal contribution
ICML 2026
MechVQA: Benchmarking and Enhancing Multimodal LLMs on Comprehensive Mechanical Drawing UnderstandingCo-first & corresponding author
2026
RAFT: Data Refinement and Adaptive Distillation for Domain Fine-Tuning with Alleviated ForgettingCo-first & corresponding author
2026
ChartWalker: Benchmarking the Cross-Chart RAG Task with Hierarchical Knowledge GraphsCorresponding author
ICML 2026 RLxF
Closing the Feedback Loop: From Experience Extraction to Insight Governance in Verbal Reinforcement LearningCo-author
2025
Rethinking Supervised Fine-Tuning: Emphasizing Key Answer Tokens for Improved LLM AccuracyFirst author · formerly SFTKey
2025
SPAR: Scholar Paper Retrieval with LLM-based Agents for Enhanced Academic SearchFirst author
2025
SciSage: A Multi-Agent Framework for High-Quality Scientific Survey GenerationFirst author
AAAI 2025
CareBot: A Pioneering Full-Process Open-Source Medical Language ModelOpen model, data, and training process
COLING 2025
MoSLD: An Extremely Parameter-Efficient Mixture-of-Shared LoRAs for Multi-Task LearningParameter-efficient multi-task adaptation
2024
CCI3.0-HQ: A Large-Scale Chinese Dataset of High Quality Designed for Pre-Training Large Language Models500GB high-quality Chinese pre-training corpus
2024
Aquila-Med LLM: Pioneering Full-Process Open-Source Medical Language ModelsBilingual medical LLM and training resources
About
Research with an engineering path to adoption
I am an AI researcher and engineer at the Beijing Academy of Artificial Intelligence. Before BAAI, I worked at ByteDance and Meituan. My work has evolved from computer vision and OCR to domain language models, post-training, AI agents, retrieval, and multimodal reasoning.
I aim to release the code, models, datasets, evaluation assets, and training recipes behind my research whenever possible.
Identity & contact
- Emailxfshi@baai.ac.cn
- GitHubxiaofengShi
- Hugging FaceXiaofengAlg
- ORCID0009-0009-6052-7713
- OpenReviewXiaofeng Shi