Hi, my name is Humzah Merchant.

humzahm@uchicago.edu

I'm an incoming PhD student at Brown Computer Science co-advised by Randall Balestriero and Bradford Levy. Currently working on financial world modeling and LLM based forecasting. See https://frontiertopit.com

Conference Papers

• Towards Financial World Modeling - Humzah Merchant, Alec Guthrie, Simon Mahns, Randall Balestriero, Bradford Levy
  NeurIPS 2026 (Paper)
  We introduce a trillion-observation financial markets dataset, establish a rigorous evaluation protocol, and systematically study self-supervised encoder training as a foundation for DINO-WM-style financial world models.
• Divergence Decoding: Targeted Unlearning via Auxiliary Models - Humzah Merchant, Bradford Levy
  ICML 2026 (Poster | Paper)
  Motivated by earlier work on addressing look-ahead bias—which is a special case of unlearning—applied a method of debiasing logits more broadly to unlearning benchmarks and found that it is state of the art across most of them.

Workshop Papers

• A Fast and Effective Solution to the Problem of Look-ahead Bias in LLMs - Humzah Merchant, Bradford Levy
  Best Poster Award at NeurIPS 2025 Workshop: Generative AI in Finance (Poster | Paper)
  Drew on product of experts and importance sampling to develop a method for reducing and/or eliminating look-ahead bias in LLMs. Demonstrated its efficacy across a variety of finance-specific tasks.
• Forecasting With LLMs: Improved Generalization Through Feature Steering - Humzah Merchant, Bradford Levy
  ICML 2026 Workshop: Forecasting as a New Frontier of Intelligence (Poster | Paper)
  Amplifying features associated with time-awareness in LLMs can reduce reliance on information from after a cutoff specified at inference-time.
• Optimizing Pre-Training of Tabular Foundation Models by Shaping Geometry - Humzah Merchant, Sriniketh Vangaru, Randall Balestriero
  ICML 2026 Workshop: Foundation Models for Structured Data (Poster)
  TFMs motivate well-distributed intermediate embeddings; although supervised pretraining does not naturally produce them, explicitly shaping early-layer geometry improves classification performance.