We are the Deep Learning Fundamentals group at Texas A&M, led by Prof. Tomer Galanti. We connect rigorous theory with practical algorithms to understand and improve modern deep learning and large language models.

Theory of LLM reasoning: Which architectures, objectives, and training procedures reliably improve multi-step reasoning in LLMs?

• LLM-PV [1]
• Auto-regressive decision trees [1]
• DisCO: constrained optimization for reasoning [1]
• Rare-token depreciation bias [1]

Deep representation learning: What representations do neural networks actually learn, and which ones support strong generalization?

• Understanding pre-training: Supervised [1, 2, 3], and self-supervised [1, 2, 3, 4]
• Representation properties: Low-rank bias [1], CRH [1], and intermediate neural collapse [1, 2]

See the full list of publications.

Contact: galanti@tamu.edu  ·  GitHub: DLFundamentals  ·  Google Scholar: Profile
Texas A&M University · Department of Computer Science & Engineering