Research

I build long-horizon agents that improve through expert iteration in context space, adapting a frozen model across many rounds by updating persistent state from verified reward signals rather than model weights.

Using this approach, my agents have discovered 150+ vulnerabilities in widely used OSS projects including Next.js, pnpm, MetaMask, and LiteLLM; improved the upper bound for an open math problem; and topped the Spider 2.0 dbt benchmark for data science tasks. Earlier, my master’s thesis work at MIT focused on model robustness under distribution shift.

Preprints

• Beyond the Model: How Harness Design Varies with LLM Performance on the 2026 International Mathematical Olympiad.
  Adib Hasan, Akashnil Dutta, Tarik Adnan Moon.
  Preprint, 2026.
  Code & Data

• AutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents.
  Adib Hasan, Daniel Schaffield, Akashnil Dutta, Tarik Adnan Moon.
  Preprint, 2026.
  Code

Publications

• VITA: Variational Pretraining of Transformers for Climate-Robust Crop Yield Forecasting.
  Adib Hasan, Mardavij Roozbehani, Munther Dahleh.
  AAAI Conference on Artificial Intelligence (AAAI), 2026.
  Oral (Top 5%)
  Arxiv | GitHub | Blogpost | Slides

• Pruning for Protection: Increasing Jailbreak Resistance in Aligned LLMs Without Fine-Tuning.
  Adib Hasan, Ileana Rugina, Alex Wang.
  BlackboxNLP Workshop at EMNLP, 2024.
  Arxiv | Dataset

• Graphettes: Constant-time determination of graphlet and orbit identity including (possibly disconnected) graphlets up to size 8.
  Adib Hasan, Po-Chien Chung, Wayne Hayes.
  PLoS ONE, 2017.
  Journal