Postdoctoral Researcher

Learning and Evaluation of Safe Agentic Models

Bocconi · REALMilan, ItalyFull time · Two years

Call upcoming

Develop learning methods and experiments for capable agents working under explicit constraints.

Your role

You will investigate how capable agents can become reliable actors under explicit constraints. The research examines how learning, planning and verification affect useful task completion, compliance with the agreed constraints and responses to failed proof attempts. This is an opportunity to develop methods and evidence that shape how safe agentic capabilities are understood and evaluated.

The position offers ownership of an empirical research agenda, with scope to introduce new learning or planning methods and question how progress should be measured. The agenda spans planning that accounts for policies, agents that generate proofs, recovery from failed reasoning and reinforcement learning with verifiable rewards, offering distinct directions for a substantial research contribution.

Within PC2, you will bring these ideas into realistic task environments where agents must satisfy explicit policies while using tools and completing workflows. Working with researchers in formal methods and systems, you will connect those policies to contracts and reusable Airlock integrations. These settings will make it possible to study when verification helps, when it obstructs useful behaviour, and how agent methods can improve the balance.

You will collaborate with the ML research engineer on model development while leading the questions and comparisons behind your studies. Evaluation will distinguish task success, constraint satisfaction, intervention and computational cost, using controlled comparisons, held out cases and informative negative results. As the programme develops, you will help shape a public benchmark or evaluation resource and engage other researchers in reproducing or challenging its findings.

Your contribution to the field

Your contribution will shape how safe agentic capabilities are learned and evaluated. New planning or learning methods, together with reproducible studies of their performance under explicit policies, will reveal how verification changes useful autonomy and where current approaches fall short. Public evaluation resources and research results will give other teams a basis for comparing agents across tasks, constraints and computational budgets, and establish your own line of work at the intersection of empirical ML and verifiable action.

An initial research or development direction

An example starting point is to test whether an agent trained to anticipate contractual constraints completes more useful work than one that encounters them only after an action is rejected. A realistic task family, credible baselines and controlled comparisons will support the study and guide method development.

Essential background

A PhD in machine learning, artificial intelligence, computer science or a closely related field; the required stage of completion and formal eligibility will be specified in the final call.
Strong empirical research and implementation skills in machine learning, agent systems, reinforcement learning or a related area.
Ability to design controlled experiments, analyse failures and distinguish competing explanations for an observed improvement.
Interest in developing agent capabilities under explicit constraints and collaborating with researchers in formal methods and systems.

Additional relevant experience

Planning that accounts for policies, agents that generate proofs, reinforcement learning with verifiable rewards or learning from tool feedback.
Agent benchmarks, realistic task environments, reproducible evaluation or adversarial testing.
Formal methods, dependent type theory, proof assistants or testing based on specifications are welcome additional areas of experience.

Candidates with depth in a relevant area and an interest in connecting it to others are encouraged to apply. Evidence of your contribution can include research papers, formal developments, substantial software or a clear account of technical work that cannot be shared publicly.

Research environment and development

Your primary supervisor will be Alessandro Rudi. PC2 is a joint effort involving PRAXIS at AI4I, the REAL team at Bocconi and STAMP at Inria. The role requires close collaboration with the other participating groups, including opportunities for extended visiting periods at their laboratories, bringing your work into close contact with complementary expertise in machine learning, formal methods and systems.

The position is based in Milan, Italy. Remote working arrangements can be agreed when working from another laboratory in the PC2 collaboration. The timing and practical arrangements for research visits will be discussed with you and the participating hosts.

Computational resources. The joint programme is supported by AI4I's HPC infrastructure: approximately 150 NVIDIA GPUs (80 Blackwell B200 and 68 Hopper H200) and 5 PB of storage for agentic AI training, testing and validation. Access for collaborative projects will be arranged through the participating teams.

Application materials

Please prepare:

Your CV, including relevant research or technical experience.
A short statement of fit and interests, explaining what attracts you to the role, the experience you would bring and directions you would be interested in exploring. A full research proposal is not required.
One or two representative contributions, such as papers, formal developments or software. Briefly explain your own contribution. An account of work that cannot be shared publicly is welcome; a public repository is not required.

Submission instructions and the application deadline will be published here when the call is finalised.