United States
I'm a fourth-year Stanford Computer Science PhD student and NSERC PGS-D scholar, working with the Stanford AI Lab and Stanford NLP Group. I am co-advised by Michael C. Frank and Noah Goodman as part of the Language & Cognition (LangCog) and Computation & Cognition (CoCo) Labs. I am grateful to receive collaboration and support from Google DeepMind, NVIDIA, Amazon Science, Microsoft AFMR, Modal, Lambda Labs, and NSF ACCESS. My ultimate goal is to blend knowledge from multiple disciplines to advance AI research. My current research centers around aligning foundation model and human learning and capabilities, particularly in reasoning, generalization, and efficiency. I have explored ways to improve the controllability of language and visual generation models, and integrate structured and multimodal information to enhance their reasoning capabilities. I'm investigating psychologically and cognitively inspired methods for continual learning, self-improvement, and advanced reasoning in foundation models. I'm also exploring methods to bridge the data efficiency gap between human and model learning while shedding further light on human cognitive models and our efficient language and vision acquisition capabilities. Previously, I was a master's student at Carnegie Mellon University (CMU), where I worked with Eduard Hovy and Malihe Alikhani on language generation, data augmentation, and commonsense reasoning. Before that, I was an undergraduate student at the University of Waterloo, where I worked with Jesse Hoey on dialogue agents and text generation. My research contributions have been recognized with several publications at major conferences and a best paper award at INLG 2021. I am also a BCV Research Fellow, and Honorable Mention for the Jessie W.H. Zou Memorial Award and CRA Outstanding Undergraduate Researcher Award. I am the lead instructor for the Stanford CS25 Transformers course, and mentor and advise several students. I also led the organization of CtrlGen, a controllable generation workshop at NeurIPS 2021, and was involved in the GEM benchmark and workshop for NLG evaluation. In my free time, I enjoy gaming, playing the piano and guitar, martial arts, and table tennis. I am also the founder and president of the Stanford Piano Society. Personal website: https://styfeng.github.io/ Twitter: https://twitter.com/stevenyfeng Google Scholar: https://scholar.google.ca/citations?user=zwiszZIAAAAJ&hl=en CV: https://styfeng.github.io/data/steven_feng_CV.pdf
◦ Founded an independent AI research group bringing together collaborators from Stanford, CMU, and other institutions. ◦ Coordinate collaborations among PhD students, researchers, and engineers; oversee project scoping, experimental design, and writing. ◦ Main writer and applicant for several research and compute grants including NSF, Modal, NVIDIA, and Lambda Labs [all awarded]. ◦ Leading multiple research projects (submitted to ICML & COLM, and targeting NeurIPS & ICLR 2026), including hallucination in LLMs [https://arxiv.org/abs/2512.21577], scaling laws for RAG-considerate pretraining [https://arxiv.org/abs/2604.00715], diffusion language models, chain-of-thought reasoning, JEPA & world models, model curiosity & welfare, etc.
As a PhD student, I am working on projects related to machine learning (ML), natural language processing (NLP), computer vision (CV), psychology, and cognitive science (cogsci). Advisors: Dr. Michael C. Frank, Noah Goodman Other: Leonidas Guibas, Douwe Kiela ◦ Efficient Reasoning: Several projects supported by NSF, Modal, NVIDIA, and Lambda Labs to enhance and optimize the reasoning capabilities of models, in particular small and data-constrained LMs. ◦ BabyLM + VLM: Projects supported by Amazon and Microsoft AFMR studying why human children learn and reason with far less data than current models. We aim to close the performance gap while reducing training data and model size by orders of magnitude. Investigating optimal training data sources and cognitively inspired benchmarks. First-author paper at EMNLP 2024 and two submissions to COLM 2026; additional papers at CogSci and CCN 2025. ◦ Code Pretraining: Collaborative project with Google DeepMind. Investigated the features of code data that support better learning for small models. ◦ Diffusion Models & Text-to-Image Generation: Worked on projects to enhance the efficiency and controllability of text-to-image generation.
LLM research team in the Applied Deep Learning Research group. Investigated optimal data selection, mixing, ordering, and curriculum learning strategies for pretraining LLMs to enhance their performance, robustness, and generalizability. Paper: https://arxiv.org/abs/2412.15285
◦ Worked with the Alexa AI team on projects to enhance the reasoning capabilities of LLMs. ◦ Investigated ideas to improve and generalize chain-of-thought (CoT) reasoning for LLMs.