Designing effective reward functions remains a fundamental challenge in reinforcement learning (RL), as it often requires extensive human effort and domain expertise. While RL from human feedback has been successful in aligning agents with human intent, acquiring high-quality feedback is costly and labor-intensive, limiting its scalability. Recent advancements in foundation models present a promising alternative--leveraging AI-generated feedback to reduce reliance on human supervision in reward learning. Building on this paradigm, we introduce ERL-VLM, an enhanced rating-based RL method that effectively learns reward functions from AI feedback. Unlike prior methods that rely on pairwise comparisons, ERL-VLM queries large vision-language models (VLMs) for absolute ratings of individual trajectories, enabling more expressive feedback and improved sample efficiency. Additionally, we propose key enhancements to rating-based RL, addressing instability issues caused by data imbalance and noisy labels. Through extensive experiments across both low-level and high-level control tasks, we demonstrate that ERL-VLM significantly outperforms existing VLM-based reward generation methods. Our results demonstrate the potential of AI feedback for scaling RL with minimal human intervention, paving the way for more autonomous and efficient reward learning.

The Maximum Cut (MaxCut) problem is NP-Complete, and obtaining its optimal solution is NP-hard in the worst case. As a result, heuristic-based algorithms are commonly used, though their design often requires significant domain expertise. More recently, learning-based methods trained on large (un)labeled datasets have been proposed; however, these approaches often struggle with generalizability and scalability. A well-known approximation algorithm for MaxCut is the Goemans-Williamson (GW) algorithm, which relaxes the Quadratic Unconstrained Binary Optimization (QUBO) formulation into a semidefinite program (SDP). The GW algorithm then applies hyperplane rounding by uniformly sampling a random hyperplane to convert the SDP solution into binary node assignments. In this paper, we propose a training-data-free approach based on a non-episodic reinforcement learning formulation, in which an agent learns to select improved rounding hyperplanes that yield better cuts than those produced by the GW algorithm. By optimizing over a Markov Decision Process (MDP), our method consistently achieves better cuts across large-scale graphs with varying densities and degree distributions.

In the past decades, machine learning with Empirical Risk Minimization (ERM) has demonstrated great capability in learning and exploiting the statistical patterns from data, or even surpassing humans. Despite the success, ERM avoids the modeling of causality the way of understanding and handling changes, which is fundamental to human intelligence. When deploying models beyond the training environment, distribution shifts are everywhere. For example, an autopilot system often needs to deal with new weather conditions that have not been seen during training, An Al-aided drug discovery system needs to predict the biochemical properties of molecules with respect to new viruses such as COVID-19. It renders the problem of Out-of-Distribution (OOD) generalization challenging to conventional machine learning. In this thesis, we investigate how to incorporate and realize the causality for broader tasks in modern machine learning. In particular, we exploit the invariance implied by the principle of independent causal mechanisms (ICM), that is, the causal mechanisms generating the effects from causes do not inform or influence each other. Therefore, the conditional distribution between the target variable given its causes is invariant under distribution shifts. With the causal invariance principle, we first instantiate it to graphs -- a general data structure ubiquitous in many real-world industry and scientific applications, such as financial networks and molecules. Then, we shall see how learning the causality benefits many of the desirable properties of modern machine learning, in terms of (i) OOD generalization capability; (ii) interpretability; and (iii) robustness to adversarial attacks. Realizing the causality in machine learning, on the other hand, raises a dilemma for optimization in conventional machine learning, as it often contradicts the objective of ERM...

The growing number of pretrained models in Machine Learning (ML) presents significant challenges for practitioners. Given a new dataset, they need to determine the most suitable deep learning (DL) pipeline, consisting of the pretrained model and the hyperparameters for finetuning to it. Moreover, as models grow in scale, the increasing reliance on real-world data poses a bottleneck for training and requires leveraging data more effectively. Addressing the first challenge often involves manual model selection and hyperparameter tuning. At the same time, as models grow larger and more and more of the available human-generated data is being used for training, data augmentation and synthetic data become critical elements. Automated machine learning offers a path to address these challenges but is traditionally designed for tabular data and classical ML methods. This dissertation adopts meta-learning to extend automated machine learning to the deep learning domain. We propose empirical approaches to automate DL pipeline selection for Computer Vision tasks using prior task knowledge to learn surrogate models for pipeline ranking. Extending these methods to the language domain, we learn to finetune large language models. As a result, we show that our approach can outperform finetuning foundation models. Additionally, we meta-learn data augmentation and synthetic data to enhance performance in up-stream and down-stream tasks. We empirically show the underestimated importance of data augmentation when using Self-Supervised Learning and meta-learn advanced data augmentation strategies. Leveraging synthetic data, we also propose to meta-learn neural synthetic data generators as proxies for Reinforcement Learning (RL) environments. Additionally, we learn a multiple-environment world model in an in-context learning fashion by purely using synthetic, randomly sampled data.

Preference-based Reinforcement Learning (PbRL) entails a variety of approaches for aligning models with human intent to alleviate the burden of reward engineering. However, most previous PbRL work has not investigated the robustness to labeler errors, inevitable with labelers who are non-experts or operate under time constraints. Additionally, PbRL algorithms often target very specific settings (e.g. pairwise ranked preferences or purely offline learning). We introduce Similarity as Reward Alignment (SARA), a simple contrastive framework that is both resilient to noisy labels and adaptable to diverse feedback formats and training paradigms. SARA learns a latent representation of preferred samples and computes rewards as similarities to the learned latent. We demonstrate strong performance compared to baselines on continuous control offline RL benchmarks. We further demonstrate SARA's versatility in applications such as trajectory filtering for downstream tasks, cross-task preference transfer, and reward shaping in online learning.

We introduce Ego-R1, a novel framework for reasoning over ultra-long (i.e., in days and weeks) egocentric videos, which leverages a structured Chain-of-Tool-Thought (CoTT) process, orchestrated by an Ego-R1 Agent trained via reinforcement learning (RL). Inspired by human problem-solving strategies, CoTT decomposes complex reasoning into modular steps, with the RL agent invoking specific tools, one per step, to iteratively and collaboratively answer sub-questions tackling such tasks as temporal retrieval and multi-modal understanding. We design a two-stage training paradigm involving supervised finetuning (SFT) of a pretrained language model using CoTT data and RL to enable our agent to dynamically propose step-by-step tools for long-range reasoning. To facilitate training, we construct a dataset called Ego-R1 Data, which consists of Ego-CoTT-25K for SFT and Ego-QA-4.4K for RL. Furthermore, our Ego-R1 agent is evaluated on a newly curated week-long video QA benchmark, Ego-R1 Bench, which contains human-verified QA pairs from hybrid sources. Extensive results demonstrate that the dynamic, tool-augmented chain-of-thought reasoning by our Ego-R1 Agent can effectively tackle the unique challenges of understanding ultra-long egocentric videos, significantly extending the time coverage from few hours to a week.

Data-driven approaches struggle with precise manipulation; imitation learning requires many hard-to-obtain demonstrations, while reinforcement learning yields brittle, non-generalizable policies. We introduce VisuoTactile Local (ViTaL) policy learning, a framework that solves fine-grained manipulation tasks by decomposing them into two phases: a reaching phase, where a vision-language model (VLM) enables scene-level reasoning to localize the object of interest, and a local interaction phase, where a reusable, scene-agnostic ViTaL policy performs contact-rich manipulation using egocentric vision and tactile sensing. This approach is motivated by the observation that while scene context varies, the low-level interaction remains consistent across task instances. By training local policies once in a canonical setting, they can generalize via a localize-then-execute strategy. ViTaL achieves around 90% success on contact-rich tasks in unseen environments and is robust to distractors. ViTaL's effectiveness stems from three key insights: (1) foundation models for segmentation enable training robust visual encoders via behavior cloning; (2) these encoders improve the generalizability of policies learned using residual RL; and (3) tactile sensing significantly boosts performance in contact-rich tasks. Ablation studies validate each of these insights, and we demonstrate that ViTaL integrates well with high-level VLMs, enabling robust, reusable low-level skills. Results and videos are available at https://vitalprecise.github.io.

This paper introduces MARCO (Multi-Agent Reinforcement learning with Conformal Optimization), a novel hardware-aware framework for efficient neural architecture search (NAS) targeting resource-constrained edge devices. By significantly reducing search time and maintaining accuracy under strict hardware constraints, MARCO bridges the gap between automated DNN design and CAD for edge AI deployment. MARCO's core technical contribution lies in its unique combination of multi-agent reinforcement learning (MARL) with Conformal Prediction (CP) to accelerate the hardware/software co-design process for deploying deep neural networks. Unlike conventional once-for-all (OFA) supernet approaches that require extensive pretraining, MARCO decomposes the NAS task into a hardware configuration agent (HCA) and a Quantization Agent (QA). The HCA optimizes high-level design parameters, while the QA determines per-layer bit-widths under strict memory and latency budgets using a shared reward signal within a centralized-critic, decentralized-execution (CTDE) paradigm. A key innovation is the integration of a calibrated CP surrogate model that provides statistical guarantees (with a user-defined miscoverage rate) to prune unpromising candidate architectures before incurring the high costs of partial training or hardware simulation. This early filtering drastically reduces the search space while ensuring that high-quality designs are retained with a high probability. Extensive experiments on MNIST, CIFAR-10, and CIFAR-100 demonstrate that MARCO achieves a 3-4x reduction in total search time compared to an OFA baseline while maintaining near-baseline accuracy (within 0.3%). Furthermore, MARCO also reduces inference latency. Validation on a MAX78000 evaluation board confirms that simulator trends hold in practice, with simulator estimates deviating from measured values by less than 5%.

Preference-based reinforcement learning (PbRL) has emerged as a promising approach for learning behaviors from human feedback without predefined reward functions. However, current PbRL methods face a critical challenge in effectively exploring the preference space, often converging prematurely to suboptimal policies that satisfy only a narrow subset of human preferences. In this work, we identify and address this preference exploration problem through population-based methods. We demonstrate that maintaining a diverse population of agents enables more comprehensive exploration of the preference landscape compared to single-agent approaches. Crucially, this diversity improves reward model learning by generating preference queries with clearly distinguishable behaviors, a key factor in real-world scenarios where humans must easily differentiate between options to provide meaningful feedback. Our experiments reveal that current methods may fail by getting stuck in local optima, requiring excessive feedback, or degrading significantly when human evaluators make errors on similar trajectories, a realistic scenario often overlooked by methods relying on perfect oracle teachers. Our population-based approach demonstrates robust performance when teachers mislabel similar trajectory segments and shows significantly enhanced preference exploration capabilities,particularly in environments with complex reward landscapes.

One proposed mechanism to improve exploration in reinforcement learning is through the use of macro-actions. Paradoxically though, in many scenarios the naive addition of macro-actions does not lead to better exploration, but rather the opposite. It has been argued that this was caused by adding non-useful macros and multiple works have focused on mechanisms to discover effectively environment-specific useful macros. In this work, we take a slightly different perspective. We argue that the difficulty stems from the trade-offs between reducing the average number of decisions per episode versus increasing the size of the action space. Namely, one typically treats each potential macro-action as independent and atomic, hence strictly increasing the search space and making typical exploration strategies inefficient. To address this problem we propose a novel regularization term that exploits the relationship between actions and macro-actions to improve the credit assignment mechanism by reducing the effective dimension of the action space and, therefore, improving exploration. The term relies on a similarity matrix that is meta-learned jointly with learning the desired policy. We empirically validate our strategy looking at macro-actions in Atari games, and the StreetFighter II environment. Our results show significant improvements over the Rainbow-DQN baseline in all environments. Additionally, we show that the macro-action similarity is transferable to related environments. We believe this work is a small but important step towards understanding how the similarity-imposed geometry on the action space can be exploited to improve credit assignment and exploration, therefore making learning more effective.