AI4OPT is a hands-on workshop and friendly competition exploring agentic AI systems that write scientific papers in optimization. We kick off the challenge during the AI Autumn School ASCOMP (competition start) and culminate at ICOMP with a showcase of the best agent-authored papers.
Domain experts (“oracles”) will supply several optimization research problems. Your goal: build or adapt an agentic LLM pipeline that reads the problem, generates hypotheses, runs experiments, analyzes results, and drafts a short-paper in standard scientific style.
What you’ll build
We will provide a simple, extensible baseline system (released before the start) that:
- reads the oracle’s problem statement and optional hints,
- proposes approaches and experiment plans based on a literature review
- executes experiments, collects metrics, and performs analyses,
- produces a first draft of a short paper (intro/method/experiments/discussion/related work).
Participants can modify any part of the pipeline—prompts, code, tools, heuristics—and inject their own intuition and domain knowledge.
What we provide
- Baseline agent code and starter prompts (simple to run, easy to extend). To situate the space, think of systems like Sakana AI Scientist v2, Zochi, or Agent Laboratory.
- A small set of optimization problem statements from oracles (released closer to launch).
- Compute resources for training/running experiments (fair-use quota; details TBA).
- Prizes & recognition for top submissions (announced at ICOMP).
Who should participate
All participants of ASCOMP and ICOMP are very welcome to participate in challenge. Also we encourage researchers, students, and practitioners in:
- optimization, OR, ML/AI, data science, systems;
- anyone curious about autonomous research agents and scientific writing.
No prior agent-framework experience is required—just Python skills and curiosity.
Key dates
- Competition start: October, 14-th
- Submission deadline: October, 18-th
- Workshop & awards @ ICOMP: October, 19-th