UCL LASP | Learning And Signal Processing

Agents for Optimization

Sun, 25 Jan 2026 00:00:00 +0000

Overview

Many industrial problems (routing, scheduling, circuit design) are NP-hard combinatorial optimization challenges. We investigate whether learning-based agents can “outsmart” or accelerate classical solvers.

Active Projects

1. Neural Combinatorial Optimization

Goal: Learning heuristics from data. Details: Instead of hand-crafting heuristics for every new problem, we train RL agents to learn construction and improvement heuristics automatically. We focus on graph-based problems where the agent learns to traverse the graph to build a valid solution.

2. Generalizable Solvers

Goal: Agents that generalize across problem sizes. Details: A major limitation of neural solvers is generalization. We are designing architectures (based on GNNs and attention) that allow an agent trained on small graphs (e.g., 20 nodes) to zero-shot generalize to large-scale instances (e.g., 1000 nodes) without retraining.

Works Done

AI for Sustainable Power Grids

Sun, 25 Jan 2026 00:00:00 +0000

Overview

The transition to renewable energy requires a smarter, more resilient grid. We apply graph-based learning to manage the combinatorial complexity of power networks and critical infrastructure.

Active Projects

1. Neural Unit Commitment

Goal: Optimize power dispatch in real-time. Details: The Unit Commitment (UC) problem—deciding which power plants to turn on—is a hard combinatorial problem. We are designing Graph Neural Networks that can approximate optimal solutions for UC faster than classical solvers, facilitating the integration of fluctuating renewable sources like wind and solar.

2. Resilient Infrastructure Monitoring

Goal: Detect failures before they become disasters. Details: Building on our work in water distribution networks, we develop graph-based anomaly detection systems. These models learn the topology of the infrastructure to localize leaks, faults, or attacks in complex sensor networks.

Works Done

Fundamentals of RL & Agents

Sun, 25 Jan 2026 00:00:00 +0000

Overview

Reinforcement Learning (RL) has achieved remarkable success, yet fundamental challenges remain in making agents sample-efficient, scalable, and capable of long-term reasoning. Our research delves into the theoretical underpinnings of RL to build more robust autonomous agents.

Active Projects

1. Scalable Environments with JAX

Goal: Accelerate RL research by orders of magnitude. Details: Building on our work Navix, we leverage JAX to create vectorised grid-world environments that compile directly to XLA. This allows for massive parallelisation, enabling us to train agents in seconds rather than hours and explore meta-learning frontiers previously out of reach.

2. Temporal Credit Assignment

Goal: Solve the “needle in a haystack” problem in long-horizon tasks. Details: When a reward is delayed, how does the agent know which past action caused it? We are developing new mechanisms for credit assignment that go beyond simple backpropagation through time, allowing agents to connect cause and effect over thousands of steps.

3. Sample Efficiency via Invariances

Goal: Learn faster by understanding symmetries. Details: We incorporate group theory into RL agents. By explicitly encoding known invariances (e.g., rotation, translation) into the network structure or the learning objective, we drastically reduce the number of samples needed to master a task.

Near-Optimal Sample Complexity in Reward-Free Kernel-Based Reinforcement Learning
A Kayal, S Vakili, L Toni, A Bernacchia. AISTATS 2025.

Reward-Free Kernel-Based Reinforcement Learning
A Kayal, S Vakili, L Toni, A Bernacchia. ICML 2024.

Navix: Scaling MiniGrid Environments with JAX
E Pignatelli, J Liesen, RT Lange, C Lu, PS Castro, L Toni. NeurIPS 2025 Dataset Track.

Assessing the zero-shot capabilities of LLMs for action evaluation in RL
E Pignatelli, J Ferret, T Rockäschel, E Grefenstette, D Paglieri, S Coward, et al. arXiv preprint 2024.

A survey of temporal credit assignment in deep reinforcement learning
E Pignatelli, J Ferret, M Geist, T Mesnard, H van Hasselt, O Pietquin, L Toni. arXiv preprint 2023.

Bayesian Optimization from Human Feedback: Near-Optimal Regret Bounds
A Kayal, S Vakili, L Toni, D Shiu, A Bernacchia. ICML 2025.

Geometric & Graph Generative AI

Sun, 25 Jan 2026 00:00:00 +0000

Overview

We investigate the fundamental limits of learning and information processing for geometric data. Our goal is to develop theoretically grounded generative models that can handle the complexity of 3D structures and molecular graphs.

Active Projects

1. Autoregressive Expansion for Latent Graph Diffusion

Goal: Extend Latent Graph Diffusion (LGDC) by introducing an autoregressive expansion mechanism. Details: Instead of expanding all nodes in a single step, this project generates fine-level structure iteratively, allowing local decisions to be conditioned on previously generated substructures.

2. Grounding Geometric Generative Models

Goal: Leverage discrete differential geometry to build better generative models. Details: We view graphs as samples from an underlying manifold. This project derives new families of diffusion and flow-based models grounded in curvature approximations and stochastic differential equations on manifolds.

LGDC: Latent Graph Diffusion via Spectrum-Preserving Coarsening
N Osman, K Jiang, D Buffelli, X Dong, L Toni. NeurIPS 2025 Workshop.

Effects of Random Edge-Dropping on Over-Squashing in Graph Neural Networks
J Singh, K Jiang, B Paige, L Toni. NeurIPS 2025.

Bures-Wasserstein Flow Matching for Graph Generation
K Jiang, J Cui, X Dong, L Toni. Submitted to ICLR, 2025.

From In Silico to In Vitro: Evaluating Molecule Generative Models
N Osman, V Lembo, G Bottegoni, L Toni. NeurIPS 2025 AI4Science Workshop.

Midi: Mixed graph and 3d denoising diffusion for molecule generation
C Vignac, N Osman, L Toni, P Frossard. ECML PKDD 2023.

Heterogeneous Graph Structure Learning through the Lens of Data-generating Processes
K Jiang, B Tang, X Dong, L Toni. AISTATS 2025.

Graph ML for Science

Sun, 25 Jan 2026 00:00:00 +0000

Overview

Biology and Chemistry are fundamentally relational—molecules are graphs of atoms, and cellular functions rely on complex interaction networks. We develop geometric deep learning methods to model, generate, and understand these structures.

Active Projects

1. Generative Biology & Drug Discovery

Goal: Move from “In Silico” generation to “In Vitro” validation. Details: We are building generative models (like MiDi and LGDC) that can design novel molecules with specific 3D geometries and chemical properties. A key focus is bridging the gap between computational metrics and actual wet-lab success rates.

2. Transcriptomics & Interaction Networks

Goal: Decode the language of the cell. Details: Using Graph Neural Networks (GNNs) and Graph Signal Processing, we model gene regulatory networks and protein-protein interactions. Our aim is to infer causal relationships in transcriptomic data to identify potential therapeutic targets.

From In Silico to In Vitro: Evaluating Molecule Generative Models
N Osman, V Lembo, G Bottegoni, L Toni. NeurIPS 2025 AI4Science Workshop.

LGDC: Latent Graph Diffusion via Spectrum-Preserving Coarsening
N Osman, K Jiang, D Buffelli, X Dong, L Toni. NeurIPS 2025 Workshop.

Heterogeneous Graph Structure Learning through the Lens of Data-generating Processes
K Jiang, B Tang, X Dong, L Toni. AISTATS 2025.

Life at LASP

Sun, 25 Jan 2026 00:00:00 +0000

Our Team in Action

We work hard, but we also enjoy our time together! Here are some highlights from recent conferences and social events.

LLM Alignment & Exploration

Sun, 25 Jan 2026 00:00:00 +0000

Overview

Large Language Models (LLMs) are powerful, but aligning them with human preferences and encouraging them to explore novel solutions remains difficult. We bring techniques from control theory and exploration research to LLMs.

Active Projects

1. Bayesian Optimization from Human Feedback

Goal: Optimize LLM outputs with minimal human labelling. Details: We treat alignment as a Bayesian Optimization problem. By efficiently querying human preferences, we aim to find optimal prompts or model weights with theoretical regret bounds, minimizing the cost of human annotation.

2. Post-Training Exploration

Goal: Encouraging LLMs to think “outside the box.” Details: Standard RLHF can lead to mode collapse (repetitive answers). We are investigating the impact of intrinsic rewards on LLMs, encouraging the model to explore diverse reasoning paths and discover creative solutions during the fine-tuning phase.

Bayesian Optimization from Human Feedback: Near-Optimal Regret Bounds
A Kayal, S Vakili, L Toni, D Shiu, A Bernacchia. ICML 2025.

Mon, 01 Jan 0001 00:00:00 +0000

Alan Guedes

Mon, 01 Jan 0001 00:00:00 +0000

Publications

Mon, 01 Jan 0001 00:00:00 +0000

Bures-Wasserstein Flow Matching for Graph Generation K Jiang, J Cui, X Dong, L Toni
2026. Submitted to ICLR arXiv:2506.14020

Reinforcement Learning Using known Invariances A Cioba, A Kayal, L Toni, S Vakili, A Bernacchia
2026. AISTATS [arXiv:2511.03473]

GT-MilliNoise: Graph transformer for point-wise denoising of indoor millimeter-wave point clouds P Gomes, W Brescia, S Mascolo, L Toni, L De Cicco
2025. Signal Processing: Image Communication

Effects of Random Edge-Dropping on Over-Squashing in Graph Neural Networks J Singh, K Jiang, B Paige, L Toni
2025. NeurIPS

Navix: Scaling minigrid environments with JAX E Pignatelli, J Liesen, RT Lange, C Lu, PS Castro, L Toni
2025. NeurIPS Dataset Track [preprint arXiv:2407.19396]

From In Silico to In Vitro: Evaluating Molecule Generative Models for Hit Generation N Osman, V Lembo, G Bottegoni, L Toni
2025. AI4Science Workshop @ NeurIPS 2025

LGDC: Latent Graph Diffusion via Spectrum-Preserving Coarsening N Osman, K Jiang, D Buffelli, X Dong, L Toni
2025. New Perspective in Graph Machine Learning Workshop @ NeurIPS 2025

MERINA+: Improving Generalization for Neural Video Adaptation via Information-Theoretic Meta-Reinforcement Learning N Kan, C Li, Y Jiang, W Dai, J Zou, H Xiong, L Toni
2025. IEEE Transactions on Circuits and Systems for Video Technology

The impact of intrinsic rewards on exploration in Reinforcement Learning A Kayal, E Pignatelli, L Toni
2025. Neural Computing and Applications

Bayesian Optimization from Human Feedback: Near-Optimal Regret Bounds A Kayal, S Vakili, L Toni, D Shiu, A Bernacchia
2025. ICML [arXiv:2505.23673]

Heterogeneous Graph Structure Learning through the Lens of Data-generating Processes K Jiang, B Tang, X Dong, L Toni
2025. AISTATS [arXiv:2503.08760]

Reward-Free Kernel-Based Reinforcement Learning A Kayal, S Vakili, L Toni, A Bernacchia
2024. ICML [arXiv:2502.07715]

Assessing the zero-shot capabilities of LLMs for action evaluation in RL E Pignatelli, J Ferret, T Rockäschel, E Grefenstette, D Paglieri, S Coward, et al.
2024. arXiv preprint arXiv:2409.12798

AGAR: Attention Graph-RNN for Adaptative Motion Prediction of Point Clouds of Deformable Objects PM Gomes, S Rossi, L Toni
2024. ACM Transactions on Multimedia Computing, Communications and Applications

Learning algorithm generalization error bounds via auxiliary distributions G Aminian, S Masiha, L Toni, MRD Rodrigues
2024. IEEE Journal on Selected Areas in Information Theory

Millinoise: a millimeter-wave radar sparse point cloud dataset in indoor scenarios W Brescia, P Gomes, L Toni, S Mascolo, L De Cicco
2024. Proceedings of the 15th ACM Multimedia Systems Conference

Conditional Meta-Reinforcement Learning with State Representation Y Sun, L Toni, Y Andreopoulos
2024. Automated Reinforcement Learning: Exploring Meta-Learning, AutoML, and LLMs

A survey of temporal credit assignment in deep reinforcement learning E Pignatelli, J Ferret, M Geist, T Mesnard, H van Hasselt, O Pietquin, L Toni
2023. arXiv preprint arXiv:2312.01072

Information-theoretic characterizations of generalization error for the Gibbs algorithm G Aminian, Y Bu, L Toni, MRD Rodrigues, GW Wornell
2023. IEEE Transactions on Information Theory

UCL LASP | Learning And Signal Processing

Agents for Optimization

Overview

Active Projects

1. Neural Combinatorial Optimization

2. Generalizable Solvers

Works Done

AI for Sustainable Power Grids

Overview

Active Projects

1. Neural Unit Commitment

2. Resilient Infrastructure Monitoring

Works Done

Fundamentals of RL & Agents

Overview

Active Projects

1. Scalable Environments with JAX

2. Temporal Credit Assignment

3. Sample Efficiency via Invariances

Related Publications

Geometric & Graph Generative AI

Overview

Active Projects

1. Autoregressive Expansion for Latent Graph Diffusion

2. Grounding Geometric Generative Models

Related Publications

Graph ML for Science

Overview

Active Projects

1. Generative Biology & Drug Discovery

2. Transcriptomics & Interaction Networks

Related Publications

Life at LASP

Our Team in Action

LLM Alignment & Exploration

Overview

Active Projects

1. Bayesian Optimization from Human Feedback

2. Post-Training Exploration

Related Publications

Alan Guedes

Publications