HUmLrn — Unified Learning from Diverse Human Feedback

About

HUmLrn investigates how artificial intelligence agents can learn from different forms of human feedback within a unified framework.

RESEARCH QUESTION

How can artificial intelligence agents make the best use of diverse human feedback to enhance their abilities?

Human expertise is communicated in several ways. People may demonstrate how to perform a task, express preferences between alternatives, correct an action, or provide an evaluative signal. These observations reveal information about the objective behind human behaviour even when they cannot express an explicit reward function.

Duration36 months

January 2026 — December 2028

FundingFIS 2 Starting Grant

Italian Science Fund · MUR

CoordinatorPolitecnico di Milano

Department of Electronics, Information and Bioengineering (DEIB)

Current statusActive
26%

Objectives

01

Methodological

Formalize the learning problem and develop a framework capable of incorporating diverse human feedback through a unified representation.

02

Theoretical

Study optimality and statistical complexity, characterize fundamental limitations, and derive conditions under which effective learning is possible.

03

Algorithmic

Design and analyse principled learning algorithms for heterogeneous feedback, alongside practical approaches for more complex environments.

04

Experimental

Develop software interfaces and evaluate learning methods in simulated sequential decision-making tasks, including an autonomous driving scenario.

People

Diego Alovisetti

Ph.D. Student

Politecnico di Milano · DEIB

LinkedIn

Riccardo Venturelli

Research Assistant

Politecnico di Milano · DEIB

GitHubLinkedIn

Maria Cristina Centorame

Research Assistant

Politecnico di Milano · DEIB

LinkedIn

Costanza Arluno

Administrative Staff

Politecnico di Milano · DEIB

Publications

20 publications

2026
Conference

Worst-Case Regret Bounds for Combinatorial Bandits with Ranking Feedback

Cristiano Migali, Gianmarco Genalti, Alberto Maria Metelli, and Marco Mussi

Advances in Neural Information Processing Systems (NeurIPS)
To AppearAcceptance rate: 7900/30709 (25.7%)ICORE 2026: A*GGS 2021: A++
Paper ↗ Poster ↗ BibTeX ↓
2026
Conference

Why Does Action Chunking Improve Behavioral Cloning Performance in Robotic Control?

Filippo Lazzati, Kyle Stachowicz, William Chen, Alberto Maria Metelli, Andrew Wagenmaker, and Sergey Levine

Conference on Robot Learning (CoRL)
To AppearAcceptance rate: 687/2094 (32.8%)
Paper ↗ Slides ↗ BibTeX ↓
2026
Workshop

Combinatorial Bandits with Plackett-Luce Feedback: A Worst-Case Analysis

Cristiano Migali, Gianmarco Genalti, Alberto Maria Metelli, and Marco Mussi

European Workshop on Reinforcement Learning (EWRL)
Paper ↗ Poster ↗ BibTeX ↓
2026
Workshop

Variance-Aware Optimal Ranking in Log-Concave Random Utility Models

Diego Alovisetti, Marco Mussi, and Alberto Maria Metelli

European Workshop on Reinforcement Learning (EWRL)
Paper ↗ Poster ↗ BibTeX ↓
2026
Conference

Fast Mixing Steady-State Control in Markov Decision Processes

Federico Corso, Marco Mussi, and Alberto Maria Metelli

International Conference on Machine Learning (ICML)
Acceptance rate: 6552/24661 (26.6%)ICORE 2026: A*GGS 2021: A++
Paper ↗ Poster ↗ BibTeX ↓
2026
Conference

Learning to Rank from Incomplete Rankings

Cristiano Migali, Gianmarco Genalti, Alberto Maria Metelli, and Marco Mussi

International Conference on Machine Learning (ICML)
Acceptance rate: 6552/24661 (26.6%)ICORE 2026: A*GGS 2021: A++
Paper ↗ Poster ↗ BibTeX ↓
2026
Workshop

Modeling Incomparability: A New Rationality Paradigm for Preference-Based Reinforcement Learning

Simone Drago, Marco Mussi, Leonardo Bianconi, and Alberto Maria Metelli

Workshop on Epistemic Intelligence in Machine Learning at ICML
Paper ↗ Poster ↗ BibTeX ↓
2026
Conference

Online Compatible Reward Identification from Preference Feedback

Simone Drago, Marco Mussi, and Alberto Maria Metelli

International Conference on Machine Learning (ICML)
Acceptance rate: 6552/24661 (26.6%)ICORE 2026: A*GGS 2021: A++
Paper ↗ Poster ↗ BibTeX ↓
2026
Conference

Reusing Trajectories in Policy Gradients Enables Fast Convergence

Alessandro Montenegro, Federico Mansutti, Marco Mussi, Matteo Papini, and Alberto Maria Metelli

International Conference on Machine Learning (ICML)
Acceptance rate: 6552/24661 (26.6%)ICORE 2026: A*GGS 2021: A++
Paper ↗ Poster ↗ BibTeX ↓
2026
Workshop

Robust Learning to Rank from Incomplete Rankings under Positional Censoring

Cristiano Migali, Gianmarco Genalti, Alberto Maria Metelli, and Marco Mussi

Workshop on Epistemic Intelligence in Machine Learning at ICML
Spotlight
Paper ↗ Slides ↗ Poster ↗ BibTeX ↓
2026
Workshop

Addressing Stochastic Rising Bandits with Thompson Sampling

Marco Fiandri, Alberto Maria Metelli, and Francesco Trovò

OPTIMAL: Optimisation and Post-Bayesian Inference in Machine Learning @ AISTATS 2026
Paper ↗ Poster ↗ BibTeX ↓
2026
Conference

Catoni-Style Change Point Detection for Regret Minimization in Piecewise-Stationary Heavy-Tailed Bandits

Gianmarco Genalti, Sujay Bhatt, Nicola Gatti, and Alberto Maria Metelli

International Conference on Artificial Intelligence and Statistics (AISTATS)
Acceptance rate: 589/2102 (28.0%)ICORE 2026: AGGS 2021: A+
Paper ↗ Poster ↗ BibTeX ↓
2026
Workshop

Closed-Form Reward Centroids for Inverse Reinforcement Learning

Filippo Lazzati and Alberto Maria Metelli

OPTIMAL: Optimisation and Post-Bayesian Inference in Machine Learning @ AISTATS 2026
Paper ↗ Poster ↗ BibTeX ↓
2026
Conference

Generalizing Behavior via Inverse Reinforcement Learning with Closed-Form Reward Centroids

Filippo Lazzati and Alberto Maria Metelli

International Conference on Artificial Intelligence and Statistics (AISTATS)
Acceptance rate: 589/2102 (28.0%)ICORE 2026: AGGS 2021: A+
Paper ↗ Poster ↗ BibTeX ↓
2026
Conference

Thompson Sampling-like Algorithms for Stochastic Rising Bandits

Marco Fiandri, Alberto Maria Metelli, and Francesco Trovò

International Conference on Artificial Intelligence and Statistics (AISTATS)
Acceptance rate: 589/2102 (28.0%)ICORE 2026: AGGS 2021: A+
Paper ↗ Poster ↗ BibTeX ↓
2026
Conference

Imitation Learning as Return Distribution Matching

Filippo Lazzati and Alberto Maria Metelli

International Conference on Learning Representations (ICLR)
Acceptance rate: 5355/19525 (27.4%)ICORE 2026: A*GGS 2021: A++
Paper ↗ Poster ↗ BibTeX ↓
2026
Workshop

Reformulating Imitation Learning as Return Distribution Matching

Filippo Lazzati and Alberto Maria Metelli

The 2nd Workshop on Advances in Financial AI Workshop: Towards Agentic and Responsible Systems at ICLR
Runner up
Paper ↗ Slides ↗ Poster ↗ BibTeX ↓
2026
Conference

Robustness in the Face of Partial Identifiability in Reward Learning

Filippo Lazzati and Alberto Maria Metelli

International Conference on Learning Representations (ICLR)
Acceptance rate: 5355/19525 (27.4%)ICORE 2026: A*GGS 2021: A++
Paper ↗ Poster ↗ BibTeX ↓
2026
Workshop

Robustness in the Face of Partial Identifiability in Reward Learning Problems

Filippo Lazzati and Alberto Maria Metelli

ICLR 2026 the 2nd Workshop on World Models: Understanding, Modelling and Scaling
Paper ↗ Poster ↗ BibTeX ↓
2026
Journal

Rising Rested Bandits: Lower Bounds and Efficient Algorithms

Marco Fiandri, Alberto Maria Metelli, and Francesco Trovò

Artificial Intelligence (AIJ)
In pressCORE 2020: A*SJR 2025: Q1
Paper ↗ BibTeX ↓

Software and datasets

News & Events

Event

Workshop on Reinforcement Learning and Game Theory

The HUmLrn team organizes a workshop at AIML Systems 2026 in Lake Como.

↗
Event

Invited talk at AIML Systems 2026

Alberto Maria Metelli gives an invited talk on Reinforcement Learning Beyond Scalar Rewards. Slides ↗

↗
News

ICML 2026

4 conference papers and 1 workshop paper presented at ICML 2026 in Seoul.

↗
Event

Research talk at TU Wien

Alberto Maria Metelli gives a research talk on Learning with Diverse Human Feedback. Slides ↗

↗
News

AISTATS 2026

3 conference papers and 2 workshop papers presented at AISTATS 2026 in Tangier.

↗
News

ICLR 2026

2 conference papers and 2 workshop papers presented at ICLR 2026 in Rio de Janeiro.

↗

COLLABORATE WITH US

Interested in our research?

Get in touch to discuss collaborations and research opportunities.
Contact the team