Seminar description

Large language models are increasingly embedded in agentic systems that address complex tasks in machine learning and science, including model development, experiment design, algorithm discovery, idea generation, literature analysis, biological design, and laboratory planning. In these systems, an LLM is not used merely as a stand-alone predictor. Instead, it operates within an agentic harness that provides tools, memory, prompts, roles, communication structures, evaluators, and iterative feedback.

Designing these harnesses is currently a substantial human engineering effort. Given an executable representation of a system, suitable benchmarks or evaluators, and sufficient computational resources, however, parts of this design process can themselves be automated. Prompts can be optimized through reflection or evolution; agent roles and communication topologies can be searched automatically; optimization and machine-learning algorithms can be generated as code; and entire scientific workflows can be organized by autonomous research agents.

This seminar studies the emerging intersection of AutoAI, AutoML for and with agents, automated agent-system design, algorithm discovery, and AutoResearch. It covers several levels of automation:

• optimization of prompts and agent instructions;
• automatic design of agent workflows and multi-agent topologies;
• evolution of programs, optimizers, and machine-learning algorithms;
• self-modifying and recursively improving agents;
• autonomous hypothesis generation, experimentation, analysis, and scientific writing;
• applications in mathematics, physics, machine learning, and biological design.

Particular attention will be paid to prominent success stories such as FunSearch, AlphaEvolve, automated AI research, autonomous quantum-physics research, and agentic biological design. At the same time, the seminar will critically examine the empirical foundations of the field. Recent studies show that multi-agent systems can suffer from specification errors, miscommunication, ineffective verification, high variance, benchmark overfitting, and unnecessary complexity. In some settings, carefully constructed single-agent systems, conventional AutoML, majority voting, or simple sampling baselines remain competitive with—or outperform—elaborate agentic systems.

The central questions of the seminar are therefore not only whether AI can automate parts of AI and scientific research, but also:

• What exactly is being automated: prompts, workflows, code, algorithms, hypotheses, experiments, or evaluation?
• Which parts of the system and search space remain designed by humans?
• What feedback signals and evaluators make autonomous improvement possible?
• Do discovered systems generalize beyond the benchmarks on which they were optimized?
• How should compute cost, inference cost, variance, and failed runs be reported?
• How can we distinguish genuine discovery from benchmark overfitting, data contamination, or rediscovery?
• When are multiple agents beneficial, and when do they merely add cost and failure modes?
• What forms of human oversight, reproducibility, scientific integrity, and safety are required?

Students will independently study recent research literature, present and critically discuss a selected topic, compare competing methodological perspectives, and prepare a written seminar paper. Where feasible, seminar papers may include a small reproduction study, an ablation, a comparison with a simple baseline, or a critical examination of an evaluation protocol.

Registration via email to Matthias Feurer and Rean Fernandes