Caiming Xiong

Overview

Caiming Xiong co-founded Recursive after leading AI research at Salesforce. He also previously worked at MetaMind.⁠[1]

His research spans machine learning, natural language processing, computer vision, multimedia, and data mining.⁠[2]

Contents16 sections

Key facts

2001–2014Education

He earned bachelor’s and master’s degrees in computer science at Huazhong University of Science and Technology, then pursued a PhD in computer science and engineering at SUNY Buffalo during 2008–2014.⁠[3]

2014–2015Research at UCLA

During 2014–2015, he was a postdoctoral researcher at UCLA under Song-Chun Zhu. Zhu’s academic record identifies his topic as robot learning from demonstrations, studying how robots acquire behavior from examples.⁠[4]

March 2016Dynamic Memory Networks

In March 2016, he, Stephen Merity, and Richard Socher published research improving the input and memory modules of Dynamic Memory Networks and applying the framework to visual and textual question answering.⁠[5]

November 2016Dynamic Coattention

In November 2016, he, Victor Zhong, and Richard Socher introduced Dynamic Coattention Networks. The model combines question and document representations and revisits candidate answer spans, allowing an initially incorrect answer to be revised.⁠[6]

August 2017Generating SQL from Natural Language

In August 2017, he co-authored Seq2SQL with Victor Zhong and Richard Socher. It translates natural-language questions into database queries, constrains generation using SQL structure, and learns from query-execution rewards.⁠[7]

March 2022Multi-Turn Program Synthesis

In March 2022, he co-authored CodeGen, which constructs programs through successive natural-language instructions. The team released model weights and training code to support research into this interactive approach to program synthesis.⁠[8]

July 2022Vision-Language Pre-training

In the BLIP paper published at ICML 2022, he, Junnan Li, Dongxu Li, and Steven Hoi combined vision-language understanding and generation, using generated captions and noise filtering to improve web image-text training data.⁠[9]

March 2024Time Series and Research Deployment

In March 2024, he co-authored an introduction to Moirai for forecasting across domains and frequencies. His accompanying biography described responsibility for language, multimodal, and action models, including moving research into production.⁠[10]

April 2024Evaluating Agents in Computer Environments

In April 2024, he co-authored OSWorld, which evaluates multimodal agents on tasks in real computer environments. Its web, desktop-application, and file-operation tasks use execution results to assess completion.⁠[11]

September 2024Action Models

In September 2024, he co-authored xLAM, a family of models for agent tool use and action tasks. The work combines different model sizes and architectures with unified, augmented, and synthetic training data.⁠[12]

March 2025A Public Lecture on Multimodal Agents

On March 17, 2025, he spoke for Salesforce AI Research in a public course organized by UC Berkeley professor Dawn Song and colleagues. His lecture concerned multimodal agents moving from perception to action.⁠[13]

Advisory Work

ACTAVA’s official team directory lists him as an advisor. This role accompanies his position as a Recursive co-founder; it does not identify him as an ACTAVA founder.⁠[14]

Research Views

In a public comment on automated research, he described reward hacking as a major challenge to address: optimizing a scoring objective can diverge from the result researchers actually intend.⁠[15]