Project: Terminal-Bench Science, a benchmark for AI agents working through computational tasks in the natural sciences.

Release: v0.1.0, released in August 2026.

Contribution: I was a main contributor to Terminal-Bench Science v0.1. Four tasks I contributed to were accepted into v0.1.

Overview Link to heading

Terminal-Bench Science studies whether AI systems can work through realistic scientific-computing workflows in a terminal environment. The benchmark is about more than producing code that looks plausible: an agent must understand the scientific setting, carry out the workflow, and produce a result that can be checked.

Four v0.1 tasks Link to heading

The four tasks I contributed to span mathematical optimization and mechanical-engineering workflows. Each title links directly to its task folder in the tagged v0.1 release:

I worked with collaborators on the scientific framing, task implementation, and evaluation setup. The descriptions here stay intentionally high-level; the public repository contains the release artifacts and task details.

Why it matters Link to heading

AI for Science needs evaluations that reflect how scientific work is actually done. A useful benchmark should test the full path from problem understanding to a meaningful, checkable outcome, while remaining clear enough for researchers to reproduce and improve.

The v0.1 release is an initial step in that direction. I am happy to have contributed four tasks at the intersection of scientific computing, optimization, and AI-agent evaluation.