Adapting Large Language Models (LLMs) to downstream tasks using Reinforcement Learning (RL) has proven to be an effective approach. However, LLMs do not inherently define the structure of an agent for RL training, particularly in terms of defining the action space. This paper studies learning a compact latent action space to enhance the controllability and exploration of RL for LLMs. We propose Controlling Large Language Models with Latent Actions (CoLA), a framework that integrates a latent action space into pre-trained LLMs. We apply CoLA to the Llama-3.1-8B model. Our experiments demonstrate that, compared to RL with token-level actions, CoLA's latent action enables greater semantic diversity in text generation. For enhancing downstream tasks, we show that CoLA with RL achieves a score of 42.4 on the math500 benchmark, surpassing the baseline score of 38.2, and reaches 68.2 when augmented with a Monte Carlo Tree Search variant. Furthermore, CoLA with RL consistently improves performance on agent-based tasks without degrading the pre-trained LLM's capabilities, unlike the baseline. Finally, CoLA reduces computation time by half in tasks involving enhanced thinking prompts for LLMs by RL. These results highlight CoLA's potential to advance RL-based adaptation of LLMs for downstream applications.

Published as a workshop paper at "Tackling Climate Change with Machine Learning", ICLR 2025 Mesoscale ocean dynamics play a critical role in climate systems, governing heat transport, hurricane genesis, and drought patterns. However, simulating these processes at high resolution remains computationally prohibitive due to their nonlinear, multiscale nature and vast spatiotemporal domains. Implicit neural representations (INRs) reduce the computational costs as resolution-independent surrogates but fail in many-query scenarios (inverse modeling) requiring rapid evaluations across diverse parameters. We present PINROD, a novel framework combining dynamics-aware implicit neural representations with parametrized neural ordinary differential equations to address these limitations. Experiments on ocean mesoscale activity data show superior accuracy over existing baselines and improved computational efficiency compared to standard numerical simulations.

The Internet of Things requires intelligent decision making in many scenarios. To this end, resources available at the individual nodes for sensing or computing, or both, can be leveraged. This results in approaches known as participatory sensing and federated learning, respectively. We investigate the simultaneous implementation of both, through a distributed approach based on empowering local nodes with game theoretic decision making. A global objective of energy minimization is combined with the individual node's optimization of local expenditure for sensing and transmitting data over multiple learning rounds. We present extensive evaluations of this technique, based on both a theoretical framework and experiments in a simulated network scenario with real data. Such a distributed approach can reach a desired level of accuracy for federated learning without a centralized supervision of the data collector. However, depending on the weight attributed to the local costs of the single node, it may also result in a significantly high Price of Anarchy (from 1.28 onwards). Thus, we argue for the need of incentive mechanisms, possibly based on Age of Information of the single nodes.

Large Language Models (LLMs) are increasingly relied upon for coding tasks, yet in most scenarios it is assumed that all relevant information can be either accessed in context or matches their training data. We posit that LLMs can benefit from the ability to interactively explore a codebase to gather the information relevant to their task. To achieve this, we present a textual environment, namely debug-gym, for developing LLM-based agents in an interactive coding setting. Our environment is lightweight and provides a preset of useful tools, such as a Python debugger (pdb), designed to facilitate an LLM-based agent's interactive debugging. Beyond coding and debugging tasks, this approach can be generalized to other tasks that would benefit from information-seeking behavior by an LLM agent.

Accurately predicting adsorption properties in nanoporous materials using Deep Learning models remains a challenging task. This challenge becomes even more pronounced when attempting to generalize to structures that were not part of the training data.. In this work, we introduce SymGNN, a graph neural network architecture that leverages material symmetries to improve adsorption property prediction. By incorporating symmetry operations into the message-passing mechanism, our model enhances parameter sharing across different zeolite topologies, leading to improved generalization. We evaluate SymGNN on both interpolation and generalization tasks, demonstrating that it successfully captures key adsorption trends, including the influence of both the framework and aluminium distribution on CO$_2$ adsorption. Furthermore, we apply our model to the characterization of experimental adsorption isotherms, using a genetic algorithm to infer likely aluminium distributions. Our results highlight the effectiveness of machine learning models trained on simulations for studying real materials and suggest promising directions for fine-tuning with experimental data and generative approaches for the inverse design of multifunctional nanomaterials.

The increasing complexity of power grid management, driven by the emergence of prosumers and the demand for cleaner energy solutions, has needed innovative approaches to ensure stability and efficiency. This paper presents a novel approach within the model-free framework of reinforcement learning, aimed at optimizing power network operations without prior expert knowledge. We introduce a masked topological action space, enabling agents to explore diverse strategies for cost reduction while maintaining reliable service using the state logic as a guide for choosing proper actions. Through extensive experimentation across 20 different scenarios in a simulated 5-substation environment, we demonstrate that our approach achieves a consistent reduction in power losses, while ensuring grid stability against potential blackouts. The results underscore the effectiveness of combining dynamic observation formalization with opponent-based training, showing a viable way for autonomous management solutions in modern energy systems or even for building a foundational model for this field.

Federated Learning (FL) enables distributed ML model training on private user data at the global scale. Despite the potential of FL demonstrated in many domains, an in-depth view of its impact on model accuracy remains unclear. In this paper, we investigate, systematically, how this learning paradigm can affect the accuracy of state-of-the-art ML models for a variety of ML tasks. We present an empirical study that involves various data types: text, image, audio, and video, and FL configuration knobs: data distribution, FL scale, client sampling, and local and global computations. Our experiments are conducted in a unified FL framework to achieve high fidelity, with substantial human efforts and resource investments. Based on the results, we perform a quantitative analysis of the impact of FL, and highlight challenging scenarios where applying FL degrades the accuracy of the model drastically and identify cases where the impact is negligible. The detailed and extensive findings can benefit practical deployments and future development of FL.

Federated learning (FL) for time series forecasting (TSF) enables clients with privacy-sensitive time series (TS) data to collaboratively learn accurate forecasting models, for example, in energy load prediction. Unfortunately, privacy risks in FL persist, as servers can potentially reconstruct clients' training data through gradient inversion attacks (GIA). Although GIA is demonstrated for image classification tasks, little is known about time series regression tasks. In this paper, we first conduct an extensive empirical study on inverting TS data across 4 TSF models and 4 datasets, identifying the unique challenges of reconstructing both observations and targets of TS data. We then propose TS-Inverse, a novel GIA that improves the inversion of TS data by (i) learning a gradient inversion model that outputs quantile predictions, (ii) a unique loss function that incorporates periodicity and trend regularization, and (iii) regularization according to the quantile predictions. Our evaluations demonstrate a remarkable performance of TS-Inverse, achieving at least a 2x-10x improvement in terms of the sMAPE metric over existing GIA methods on TS data. Code repository: https://github.com/Capsar/ts-inverse

In modern resource-sharing systems, multiple agents access limited resources with unknown stochastic conditions to perform tasks. When multiple agents access the same resource (arm) simultaneously, they compete for successful usage, leading to contention and reduced rewards. This motivates our study of competitive multi-armed bandit (CMAB) games. In this paper, we study a new N-player K-arm competitive MAB game, where non-myopic players (agents) compete with each other to form diverse private estimations of unknown arms over time. Their possible collisions on same arms and time-varying nature of arm rewards make the policy analysis more involved than existing studies for myopic players. We explicitly analyze the threshold-based structures of social optimum and existing selfish policy, showing that the latter causes prolonged convergence time $\Omega(\frac{K}{\eta^2}\ln({\frac{KN}{\delta}}))$, while socially optimal policy with coordinated communication reduces it to $\mathcal{O}(\frac{K}{N\eta^2}\ln{(\frac{K}{\delta})})$. Based on the comparison, we prove that the competition among selfish players for the best arm can result in an infinite price of anarchy (PoA), indicating an arbitrarily large efficiency loss compared to social optimum. We further prove that no informational (non-monetary) mechanism (including Bayesian persuasion) can reduce the infinite PoA, as the strategic misreporting by non-myopic players undermines such approaches. To address this, we propose a Combined Informational and Side-Payment (CISP) mechanism, which provides socially optimal arm recommendations with proper informational and monetary incentives to players according to their time-varying private beliefs. Our CISP mechanism keeps ex-post budget balanced for social planner and ensures truthful reporting from players, achieving the minimum PoA=1 and same convergence time as social optimum.

Innovation in nanophotonics currently relies on human experts who synergize specialized knowledge in photonics and coding with simulation and optimization algorithms, entailing design cycles that are time-consuming, computationally demanding, and frequently suboptimal. We introduce MetaChat, a multi-agentic design framework that can translate semantically described photonic design goals into high-performance, freeform device layouts in an automated, nearly real-time manner. Multi-step reasoning is enabled by our Agentic Iterative Monologue (AIM) paradigm, which coherently interfaces agents with code-based tools, other specialized agents, and human designers. Design acceleration is facilitated by Feature-wise Linear Modulation-conditioned Maxwell surrogate solvers that support the generalized evaluation of metasurface structures. We use freeform dielectric metasurfaces as a model system and demonstrate with MetaChat the design of multi-objective, multi-wavelength metasurfaces orders of magnitude faster than conventional methods. These concepts present a scientific computing blueprint for utilizing specialist design agents, surrogate solvers, and human interactions to drive multi-physics innovation and discovery.