Research: Generative Discovery Beyond the Data

My research centers on exploration for out-of-distribution discovery: from the theoretical foundations of exploration and discovery, to principled flow- and diffusion-based discovery methods, and applications in biochemistry. More broadly, I aim to understand and establish the principles of discovery systems at the intersection of generative modeling, optimization, and sequential decision-making, toward a science of generative discovery.

Research Overview

Mathematical Foundations of Exploration and Discovery

Mathematical Foundations animation

I develop mathematical foundations for exploration and discovery, from discrete dynamical systems to flow- and diffusion-based generative models. My work showed that maximum-entropy exploration can require non-Markovian policies [M1], developed frameworks for optimizing complex exploration and experimental-design objectives [M2, M3], and studied how geometric priors can improve the statistical complexity of active exploration [M4]. To understand generative discovery processes, I recently extended these principles to answer questions such as:

  • How can exploration be formalized on spaces implicitly represented by flow models? [M5, M6]
  • When can local flow expansion yield global coverage of new valid design regions? [M7]
  • What guarantees are possible for distributional flow adaptation beyond average reward? [M8, M9]

Selected Papers

  1. ICML
    Provable Maximum Entropy Manifold Exploration via Diffusion Models
    Riccardo De Santi*, Marin Vlastelica*, Ya-Ping Hsieh, and 3 more authors
    International Conference on Machine Learning (ICML), 2025
    Bibtex PDF Code
  2. ICMLOutstanding Paper
    The Importance of Non-Markovianity in Maximum State Entropy Exploration
    Mirco Mutti*, Riccardo De Santi*, and Marcello Restelli
    International Conference on Machine Learning (ICML), 2022
    Outstanding Paper Award at ICML 2022
    Bibtex Talk PDF

See all publications →

Discovery Algorithms via Flow and Diffusion Models

Generative Discovery Algorithms animation

I develop scalable algorithms that turn flow and diffusion models into practical engines for out-of-distribution discovery. My methods go beyond standard reward-guided fine-tuning: they adapt pre-trained generative models to amplify low-probability modes hidden within the prior — effectively debiasing it from its pre-training data [A1, A2]; to expand into new valid regions through verifier-constrained entropy expansion, yielding higher novelty and diversity [A3]; and to target rare, high-value outcomes via distributional fine-tuning [A4, A5], or access intermediate states via reward-guided merging [A6]. Concretely, my algorithms contributed to answering questions such as:

  • How can fine-tuning allow to access low-probability modes hidden in a pre-trained flow? [A1, A2]
  • How can verifier feedback drive flow expansion into new valid design regions? [A3]
  • How can flow fine-tuning target rare outcomes in the tails of the reward distribution? [A4, A5]

Selected Papers

  1. ICMLOral Presentation
    A Unified Density Operator View of Flow Control and Merging
    Riccardo De Santi, Malte Franke, Ya-Ping Hsieh, and 1 more author
    International Conference on Machine Learning (ICML), 2026
    Oral at Real-World Constrained and Preference-Aligned Flow and Diffusion-Based Models Workshop at ICLR 2026
    Bibtex PDF
  2. ICLR
    Verifier-Constrained Flow Expansion for Discovery Beyond the Data
    Riccardo De Santi*, Kimon Protopapas*, Ya-Ping Hsieh, and 1 more author
    International Conference on Learning Representations (ICLR), 2026
    Bibtex PDF Code
  3. NeurIPSSpotlight and Oral Presentation
    Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning
    Riccardo De Santi, Marin Vlastelica, Ya-Ping Hsieh, and 3 more authors
    Advances in Neural Information Processing Systems (NeurIPS), 2025
    Spotlight at NeurIPS 2025 and Oral at Workshop on Generative AI and Biology at ICML 2025
    Bibtex PDF Code

See all publications →

Biochemistry Applications

Real-World Science Applications animation

I bring discovery algorithms to the design of drug-like molecules, therapeutic peptides, and proteins, partnering with chemistry and biology academic labs [e.g., LAB1, LAB2] and industry [e.g., LAB3] to close the loop between generative exploration and out-of-distribution discovery on real-world data. This line focuses on translating principled methods into measurable impact for sustainable chemistry and biotechnology. Concretely, I work on questions such as:

  • How should we measure diversity, novelty, and coverage across biochemical design spaces? [B1]
  • How can we generate novel molecules that satisfy both functional and synthetic constraints? [B2]
  • How can we sacrifice the average molecular quality to improve the top candidates? [B3, B4]

Selected Papers

  1. ICMLOral Presentation
    Constrained Molecular Generation via Sequential Flow Model Fine-Tuning
    Sven Gutjahr*, Riccardo De Santi*, Luca Schaufelberger*, and 2 more authors
    International Conference on Machine Learning (ICML), 2026
    Oral at Frontiers in Probabilistic Inference Workshop at NeurIPS 2025
    Bibtex
  2. ICML
    Efficient Tail-Aware Generative Optimization via Flow Model Fine-Tuning
    Zifan Wang, Riccardo De Santi, Xiaoyu Mo, and 3 more authors
    International Conference on Machine Learning (ICML), 2026
    Bibtex PDF
  3. ICLR
    Verifier-Constrained Flow Expansion for Discovery Beyond the Data
    Riccardo De Santi*, Kimon Protopapas*, Ya-Ping Hsieh, and 1 more author
    International Conference on Learning Representations (ICLR), 2026
    Bibtex PDF Code

See all publications →