Research

I’m interested in test-time scaling and reasoning, especially how to build AI agents that can solve hard problems and remain effective over long periods. My preferred approach is to build strong mathematical foundations and then test limitations of theory through experiments.

Two of my most recent projects are AutoFyn and Discovery–Execution.

AutoFyn is an autonomous agent system inspired by reinforcement learning principles. It discovered 150+ vulnerabilities in major open-source projects, including MetaMask, LiteLLM, and pnpm, and topped the Spider 2.0 DBT benchmark for data science tasks.

Discovery–Execution is a framework for predicting how much AI models benefit from longer reasoning, before running those longer attempts. It showed promising results across multiple model families.

Earlier, my master’s work at MIT focused on robustness under distribution shift. This work included VITA, which was selected for an oral presentation at AAAI 2026.

You can find my papers below and on my Google Scholar profile.

Test Time Scaling

Discovery–Execution framework: measure discovery and execution, then predict success across compute allocations

When Does Longer Reasoning Help? Predicting Mathematical Reasoning Through Discovery and Execution

Adib Hasan, Lay Jain, Thanic Nur Samin

MATH-AI Workshop at NeurIPS, 2026

paper / code / dataset / blog post

AutoFyn interface running a long-horizon security audit

AutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents

Adib Hasan, Daniel Schaffield, Akashnil Dutta, Tarik Adnan Moon

Preprint, 2026

paper / code / blog post

Robustness Against Adverse Distribution Shift

Overview of VITA variational pretraining and crop-yield fine-tuning

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 Presentation (Top 5%)

paper / code / blog post / slides

Refusal rates of pruned and unpruned language models

Pruning for Protection: Increasing Jailbreak Resistance in Aligned LLMs Without Fine-Tuning

Adib Hasan, Ileana Rugina, Alex Wang

BlackboxNLP Workshop at EMNLP, 2024

paper / dataset

Algorithms and Data Structures

Learned binary heap architecture and performance results

Towards Learned Binary Heaps

Adib Hasan, Angelos Pelecanos

MIT 6.890: Learning-Augmented Algorithms, 2019

paper / blog post

Graph and bit-vector representations used by Graphettes

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

paper