Ali Shehper

I am an AI Researcher at NVIDIA, working as a Senior Research Scientist on the Nemotron team.

Previously, I worked on language model pre-training at Essential AI. Prior to that, I was a Research Scientist at Caltech, focusing on Reinforcement Learning for Math, and earlier, a Physicist studying supersymmetric and topological quantum field theories.

Pre-training Science

• A Training-Time Diagnostic for Generalization via the Log-Alignment Ratio. A. Shehper, A. Vaswani.

RL For Math

• The Two-Hump Problem: Bridging the Difficulty Gap in Mathematical Reinforcement Learning, (Accepted at ICML 2026)
  L. Fagan*, M. Tarquini*, A. Shehper*, M. Manko, A. Gruen, C. Huang, G. Butbaia, D. Passaro, S. Gukov. (* equal contribution)

• Hierarchical Reinforcement Learning for Sparse-Reward Search in Commutative Algebra, (Accepted at ICML 2026)
  G. Butbaia, P. Orland, C. Huang, D. Passaro, L. Fagan, M. Tarquini, H. Dao, D. Eisenbud, A. Shehper, S. Gukov.

• What makes math problems hard for reinforcement learning: a case study, (Accepted at NeurIPS 2025)
  A. Shehper, A. M. Medina-Mardones, L. Fagan, B. Lewandowski, A. Gruen, Y. Qiu, P. Kucharski, Z. Wang, S. Gukov.

Physics and Math

• PhD Thesis: Aspects of Supersymmetric and Topological Quantum Field Theories.

• Nonabelian Twists of the D4 Theory J. Distler, B. Ergun, A. Shehper (equal contribution with B. Ergun, alphabetical order).

• Symmetries of 2d TQFTs and Equivariant Verlinde Formulae for General Groups S. Gukov, D. Pei, C. Reid, A. Shehper (lead contributor, alphabetical order).

• Distinguishing 4d N=2 SCFTs J. Distler, B. Ergun, A. Shehper (equal contribution with B. Ergun, alphabetical order).

• Deformations of surface defect moduli spaces A. Neitzke, A. Shehper (lead contributor, alphabetical order).

Research Blog

• A Features’ Perspective on Neural Scaling Laws

Code

• AC-Solver A Python library for tackling long-horizon, ultra-sparse-reward RL environments, designed to accompany our case study.

• Sparse-Dictionary-Learning An open-source implementation of Anthropic’s Towards Monosemanticity: Decomposing Language Models with Dictionary Learning.

• Neural Scaling Laws An implementation of Scaling Laws for Neural Language Models, along with results from An Empirical Model of Large-Batch Training.

• Language Model Feature Browser A visualizer for features learned by a 1-layer Language Model, with the GitHub Repository.

Media

• How an Unsolved Math Problem Could Train AI to Predict Crises Years in Advance, Scientific American. August, 2025.
• AI Program Plays the Long Game to Solve Decades-Old Math Problems, Caltech Media. February, 2025.

Google Scholar / GitHub / Twitter / LinkedIn