arXiv subjects

Alfredo Garcia

Publications and source records attributed to Alfredo Garcia.

At least 19 recordsLinked to original sources

arXiv2026

Reinforcing Multi-Turn Reasoning in LLM Agents via Fine-Grained Reward Structure and Credit Assignment

Reinforcement Learning (RL) approaches have been wildly used to enhance the reasoning capabilities of Large Language Model (LLM) agents in long-horizon, multi-turn scenarios. Such interactions can be formalized as turn-level Markov decision processes (MDPs), where intermediate rewards are often available. However, most prior work relies on sparse trajectory-level rewards, resulting in poor credit assignment, while dense turn-level rewards remain underexplored. In this paper, we investigate how to effectively leverage dense turn-level reward structures in RL algorithms, specifically Group Relative Policy Optimization (GRPO) and Proximal Policy Optimization (PPO), to enable fine-grained credit assignment. We categorize reward structures into three types based on their granularity: (1) terminal reward; (2) delayed reward; (3) per-turn reward, each corresponding to a distinct turn-level MDP formulation, and derive GRPO and PPO algorithms tailored to each case, respectively. Experiments on multi-turn search and game agent tasks show that, for both GRPO and PPO, dense per-turn reward structures consistently outperform sparse terminal and delayed reward structures in terms of training dynamics and numerical results. Furthermore, on search tasks, PPO with dense per-turn rewards achieves greater training stability and faster convergence, and delivers the highest answer correctness among all state-of-the-art methods across diverse question-answering datasets.

cs.LG↗

arXiv2026

HiPER: Hierarchical Reinforcement Learning with Explicit Credit Assignment for Large Language Model Agents

Training LLMs as interactive agents for multi-turn decision-making remains challenging, particularly in long-horizon tasks with sparse and delayed rewards, where agents must execute extended sequences of actions before receiving meaningful feedback. Most existing reinforcement learning (RL) approaches model LLM agents as flat policies operating at a single time scale, selecting one action at each turn. In sparse-reward settings, such flat policies must propagate credit across the entire trajectory without explicit temporal abstraction, which often leads to unstable optimization and inefficient credit assignment. We propose HiPER, a novel Hierarchical Plan-Execute RL framework that explicitly separates high-level planning from low-level execution. HiPER factorizes the policy into a high-level planner that proposes subgoals and a low-level executor that carries them out over multiple action steps. To align optimization with this structure, we introduce a key technique called hierarchical advantage estimation (HAE), which carefully assigns credit at both the planning and execution levels. By aggregating returns over the execution of each subgoal and coordinating updates across the two levels, HAE provides an unbiased gradient estimator and provably reduces variance compared to flat generalized advantage estimation. Empirically, HiPER achieves state-of-the-art performance on challenging interactive benchmarks, reaching 97.4\% success on ALFWorld and 83.3\% on WebShop with Qwen2.5-7B-Instruct (+6.6\% and +8.3\% over the best prior method), with especially large gains on long-horizon tasks requiring multiple dependent subtasks. These results highlight the importance of explicit hierarchical decomposition for scalable RL training of multi-turn LLM agents.

cs.LG↗

arXiv2026

Stabilizing Off-Policy Training for Long-Horizon LLM Agent via Turn-Level Importance Sampling and Clipping-Triggered Normalization

Reinforcement learning (RL) algorithms such as PPO and GRPO are widely used to train large language models (LLMs) for multi-turn agentic tasks. However, in off-policy training pipelines, these methods often exhibit unstable optimization dynamics and are prone to performance collapse. Through empirical analysis, we identify two fundamental sources of instability in this setting: (1)~a granularity mismatch between token-level policy optimization and turn-structured interactions, and (2) high-variance and unreliable gradient updates induced by off-policy importance sampling and inaccurate advantage estimation. To address these challenges, we propose SORL, \underline{S}tabilizing \underline{O}ff-Policy \underline{R}einforcement \underline{L}earning for Long-Horizon Agent Training. SORL introduces principled mechanisms that align policy optimization with the structure of multi-turn interactions and adaptively suppress unreliable off-policy updates, yielding more conservative and robust learning dynamics. Within this framework, we instantiate two stabilized algorithms: SO-PPO and SO-GRPO. Both algorithms are designed to mitigate gradient variance and prevent optimization collapse without requiring careful early stopping or heuristic tuning. We evaluate SO-PPO and SO-GRPO on a range of multi-turn search benchmarks, including general question answering, multi-hop question answering, and medical multiple-choice QA tasks. Experimental results show that both methods consistently prevent training instabilities and performance collapses observed in standard PPO and GRPO, maintain lower clipping ratios and more stable optimization trajectories, and achieve superior or comparable task performance. These results demonstrate that the proposed algorithm provides a practical, scalable, and general framework for stabilizing reinforcement learning in multi-turn LLM agent training.

cs.LG↗

arXiv2025

Do LLMs Recognize Your Latent Preferences? A Benchmark for Latent Information Discovery in Personalized Interaction

Large Language Models (LLMs) excel at producing broadly relevant text, but this generality becomes a limitation when user-specific preferences are required, such as recommending restaurants or planning travel. In these scenarios, users rarely articulate every preference explicitly; instead, much of what they care about remains latent, waiting to be inferred. This raises a fundamental question: Can LLMs uncover and reason about such latent information through conversation? We address this problem by introducing a unified benchmark for evaluating latent information discovery - the ability of LLMs to reveal and utilize hidden user attributes through multi-turn interaction. The benchmark spans three progressively realistic settings: the classic 20 Questions game, Personalized Question Answering, and Personalized Text Summarization. All tasks share a tri-agent framework (User, Assistant, Judge) enabling turn-level evaluation of elicitation and adaptation. Our results reveal that while LLMs can indeed surface latent information through dialogue, their success varies dramatically with context: from 32% to 98%, depending on task complexity, topic, and number of hidden attributes. This benchmark provides the first systematic framework for studying latent information discovery in personalized interaction, highlighting that effective preference inference remains an open frontier for building truly adaptive AI systems.

cs.LG↗

arXiv2025

ADARL: Adaptive Low-Rank Structures for Robust Policy Learning under Uncertainty

Robust reinforcement learning (Robust RL) seeks to handle epistemic uncertainty in environment dynamics, but existing approaches often rely on nested min--max optimization, which is computationally expensive and yields overly conservative policies. We propose \textbf{Adaptive Rank Representation (AdaRL)}, a bi-level optimization framework that improves robustness by aligning policy complexity with the intrinsic dimension of the task. At the lower level, AdaRL performs policy optimization under fixed-rank constraints with dynamics sampled from a Wasserstein ball around a centroid model. At the upper level, it adaptively adjusts the rank to balance the bias--variance trade-off, projecting policy parameters onto a low-rank manifold. This design avoids solving adversarial worst-case dynamics while ensuring robustness without over-parameterization. Empirical results on MuJoCo continuous control benchmarks demonstrate that AdaRL not only consistently outperforms fixed-rank baselines (e.g., SAC) and state-of-the-art robust RL methods (e.g., RNAC, Parseval), but also converges toward the intrinsic rank of the underlying tasks. These results highlight that adaptive low-rank policy representations provide an efficient and principled alternative for robust RL under model uncertainty.

cs.LG↗

arXiv2025

Aligning Frozen LLMs by Reinforcement Learning: An Iterative Reweight-then-Optimize Approach

Aligning large language models (LLMs) with human preferences usually requires fine-tuning methods such as RLHF and DPO. These methods directly optimize the model parameters, so they cannot be used in test-time to improve model performance, nor are they applicable when the model weights are not accessible. In contrast, test-time methods sidestep weight updates by leveraging reward functions to guide and improve output quality. However, they incur high inference costs, and their one-shot guidance is often based on imperfect reward or value functions, leading to suboptimal outputs. In this work, we present a method named Iterative Reweight-then-Optimize (IRO), a reinforcement learning (RL) framework that performs RL-style alignment of the (frozen) base model without touching its parameters. During training, each iteration (i) samples candidates from the base model, (ii) resamples using current value functions, and (iii) trains a new lightweight value function that guides the next decoding pass. At test time, the value functions are used to guide the base model generation via a search-based optimization process. Notably, users can apply IRO to align a model on their own dataset, similar to OpenAI's reinforcement fine-tuning (RFT), but without requiring access to the model weights.

cs.LG↗

arXiv2025

Learning An Active Inference Model of Driver Perception and Control: Application to Vehicle Car-Following

In this paper we introduce a general estimation methodology for learning a model of human perception and control in a sensorimotor control task based upon a finite set of demonstrations. The model's structure consists of i the agent's internal representation of how the environment and associated observations evolve as a result of control actions and ii the agent's preferences over observable outcomes. We consider a model's structure specification consistent with active inference, a theory of human perception and behavior from cognitive science. According to active inference, the agent acts upon the world so as to minimize surprise defined as a measure of the extent to which an agent's current sensory observations differ from its preferred sensory observations. We propose a bi-level optimization approach to estimation which relies on a structural assumption on prior distributions that parameterize the statistical accuracy of the human agent's model of the environment. To illustrate the proposed methodology, we present the estimation of a model for car-following behavior based upon a naturalistic dataset. Overall, the results indicate that learning active inference models of human perception and control from data is a promising alternative to black-box models of driving.

cs.LG↗

arXiv2025

Understanding Inverse Reinforcement Learning under Overparameterization: Non-Asymptotic Analysis and Global Optimality

The goal of the Inverse reinforcement learning (IRL) task is to identify the underlying reward function and the corresponding optimal policy from a set of expert demonstrations. While most IRL algorithms' theoretical guarantees rely on a linear reward structure, we aim to extend the theoretical understanding of IRL to scenarios where the reward function is parameterized by neural networks. Meanwhile, conventional IRL algorithms usually adopt a nested structure, leading to computational inefficiency, especially in high-dimensional settings. To address this problem, we propose the first two-timescale single-loop IRL algorithm under neural network parameterized reward and provide a non-asymptotic convergence analysis under overparameterization. Although prior optimality results for linear rewards do not apply, we show that our algorithm can identify the globally optimal reward and policy under certain neural network structures. This is the first IRL algorithm with a non-asymptotic convergence guarantee that provably achieves global optimality in neural network settings.

stat.ML↗

arXiv2025

Regularized Q-Learning with Linear Function Approximation

Regularized Markov Decision Processes serve as models of sequential decision making under uncertainty wherein the decision maker has limited information processing capacity and/or aversion to model ambiguity. With functional approximation, the convergence properties of learning algorithms for regularized MDPs (e.g. soft Q-learning) are not well understood because the composition of the regularized Bellman operator and a projection onto the span of basis vectors is not a contraction with respect to any norm. In this paper, we consider a bi-level optimization formulation of regularized Q-learning with linear functional approximation. The {\em lower} level optimization problem aims to identify a value function approximation that satisfies Bellman's recursive optimality condition and the {\em upper} level aims to find the projection onto the span of basis vectors. This formulation motivates a single-loop algorithm with finite time convergence guarantees. The algorithm operates on two time-scales: updates to the projection of state-action values are `slow' in that they are implemented with a step size that is smaller than the one used for `faster' updates of approximate solutions to Bellman's recursive optimality equation. We show that, under certain assumptions, the proposed algorithm converges to a stationary point in the presence of Markovian noise. In addition, we provide a performance guarantee for the policies derived from the proposed algorithm.

cs.AI↗

arXiv2024

Retraction-Free Decentralized Non-convex Optimization with Orthogonal Constraints

In this paper, we investigate decentralized non-convex optimization with orthogonal constraints. Conventional algorithms for this setting require either manifold retractions or other types of projection to ensure feasibility, both of which involve costly linear algebra operations (e.g., SVD or matrix inversion). On the other hand, infeasible methods are able to provide similar performance with higher computational efficiency. Inspired by this, we propose the first decentralized version of the retraction-free landing algorithm, called \textbf{D}ecentralized \textbf{R}etraction-\textbf{F}ree \textbf{G}radient \textbf{T}racking (DRFGT). We theoretically prove that DRFGT enjoys the ergodic convergence rate of $\mathcal{O}(1/K)$, matching the convergence rate of centralized, retraction-based methods. We further establish that under a local Riemannian PŁ condition, DRFGT achieves a much faster linear convergence rate. Numerical experiments demonstrate that DRFGT performs on par with the state-of-the-art retraction-based methods with substantially reduced computational overhead.

cs.LG↗

arXiv2024

Local Linear Convergence of Infeasible Optimization with Orthogonal Constraints

Many classical and modern machine learning algorithms require solving optimization tasks under orthogonality constraints. Solving these tasks with feasible methods requires a gradient descent update followed by a retraction operation on the Stiefel manifold, which can be computationally expensive. Recently, an infeasible retraction-free approach, termed the landing algorithm, was proposed as an efficient alternative. Motivated by the common occurrence of orthogonality constraints in tasks such as principle component analysis and training of deep neural networks, this paper studies the landing algorithm and establishes a novel linear convergence rate for smooth non-convex functions using only a local Riemannian PŁ condition. Numerical experiments demonstrate that the landing algorithm performs on par with the state-of-the-art retraction-based methods with substantially reduced computational overhead.

math.OC↗

arXiv2024

Learning Reward and Policy Jointly from Demonstration and Preference Improves Alignment

Aligning human preference and value is an important requirement for building contemporary foundation models and embodied AI. However, popular approaches such as reinforcement learning with human feedback (RLHF) break down the task into successive stages, such as supervised fine-tuning (SFT), reward modeling (RM), and reinforcement learning (RL), each performing one specific learning task. Such a sequential approach results in serious issues such as significant under-utilization of data and distribution mismatch between the learned reward model and generated policy, which eventually lead to poor alignment performance. We develop a single stage approach named Alignment with Integrated Human Feedback (AIHF), capable of integrating both human preference and demonstration to train reward models and the policy. The proposed approach admits a suite of efficient algorithms, which can easily reduce to, and leverage, popular alignment algorithms such as RLHF and Directly Policy Optimization (DPO), and only requires minor changes to the existing alignment pipelines. We demonstrate the efficiency of the proposed solutions with extensive experiments involving alignment problems in LLMs and robotic control problems in MuJoCo. We observe that the proposed solutions outperform the existing alignment algorithms such as RLHF and DPO by large margins, especially when the amount of high-quality preference data is relatively limited.

cs.AI↗

arXiv2024

Getting More Juice Out of the SFT Data: Reward Learning from Human Demonstration Improves SFT for LLM Alignment

Aligning human preference and value is an important requirement for contemporary foundation models. State-of-the-art techniques such as Reinforcement Learning from Human Feedback (RLHF) often consist of two stages: 1) supervised fine-tuning (SFT), where the model is fine-tuned by learning from human demonstration data; 2) Preference learning, where preference data is used to learn a reward model, which is in turn used by a reinforcement learning (RL) step to fine-tune the model. Such reward model serves as a proxy to human preference, and it is critical to guide the RL step towards improving the model quality. In this work, we argue that the SFT stage significantly benefits from learning a reward model as well. Instead of using the human demonstration data directly via supervised learning, we propose to leverage an Inverse Reinforcement Learning (IRL) technique to simultaneously build an reward model and a policy model. This approach leads to new SFT algorithms that are not only efficient to implement, but are robust to the presence of low-quality supervised learning data. Moreover, we discover a connection between the proposed IRL based approach, and a recent line of works called Self-Play Fine-tune (SPIN). Theoretically, we show that the proposed algorithms converge to the stationary solutions of the IRL problem. Empirically, we align 1B and 7B models using proposed methods and evaluate them on a reward benchmark model and the HuggingFace Open LLM Leaderboard. The proposed methods show significant performance improvement over existing SFT approaches. Our results indicate that it is beneficial to leverage reward learning throughout the entire alignment process.

cs.AI↗

arXiv2024

Distributed Networked Multi-task Learning

We consider a distributed multi-task learning scheme that accounts for multiple linear model estimation tasks with heterogeneous and/or correlated data streams. We assume that nodes can be partitioned into groups corresponding to different learning tasks and communicate according to a directed network topology. Each node estimates a linear model asynchronously and is subject to local (within-group) regularization and global (across groups) regularization terms targeting noise reduction and generalization performance improvement respectively. We provide a finite-time characterization of convergence of the estimators and task relation and illustrate the scheme's general applicability in two examples: random field temperature estimation and modeling student performance from different academic districts.

cs.MA↗

arXiv2024

FedGlu: A personalized federated learning-based glucose forecasting algorithm for improved performance in glycemic excursion regions

Continuous glucose monitoring (CGM) devices provide real-time glucose monitoring and timely alerts for glycemic excursions, improving glycemic control among patients with diabetes. However, identifying rare events like hypoglycemia and hyperglycemia remain challenging due to their infrequency. Moreover, limited access to sensitive patient data hampers the development of robust machine learning models. Our objective is to accurately predict glycemic excursions while addressing data privacy concerns. To tackle excursion prediction, we propose a novel Hypo-Hyper (HH) loss function, which significantly improves performance in the glycemic excursion regions. The HH loss function demonstrates a 46% improvement over mean-squared error (MSE) loss across 125 patients. To address privacy concerns, we propose FedGlu, a machine learning model trained in a federated learning (FL) framework. FL allows collaborative learning without sharing sensitive data by training models locally and sharing only model parameters across other patients. FedGlu achieves a 35% superior glycemic excursion detection rate compared to local models. This improvement translates to enhanced performance in predicting both, hypoglycemia and hyperglycemia, for 105 out of 125 patients. These results underscore the effectiveness of the proposed HH loss function in augmenting the predictive capabilities of glucose predictions. Moreover, implementing models within a federated learning framework not only ensures better predictive capabilities but also safeguards sensitive data concurrently.

cs.LG↗

arXiv2024

A Bayesian Approach to Robust Inverse Reinforcement Learning

We consider a Bayesian approach to offline model-based inverse reinforcement learning (IRL). The proposed framework differs from existing offline model-based IRL approaches by performing simultaneous estimation of the expert's reward function and subjective model of environment dynamics. We make use of a class of prior distributions which parameterizes how accurate the expert's model of the environment is to develop efficient algorithms to estimate the expert's reward and subjective dynamics in high-dimensional settings. Our analysis reveals a novel insight that the estimated policy exhibits robust performance when the expert is believed (a priori) to have a highly accurate model of the environment. We verify this observation in the MuJoCo environments and show that our algorithms outperform state-of-the-art offline IRL algorithms.

cs.LG↗

arXiv2024

A Unified View on Solving Objective Mismatch in Model-Based Reinforcement Learning

Model-based Reinforcement Learning (MBRL) aims to make agents more sample-efficient, adaptive, and explainable by learning an explicit model of the environment. While the capabilities of MBRL agents have significantly improved in recent years, how to best learn the model is still an unresolved question. The majority of MBRL algorithms aim at training the model to make accurate predictions about the environment and subsequently using the model to determine the most rewarding actions. However, recent research has shown that model predictive accuracy is often not correlated with action quality, tracing the root cause to the objective mismatch between accurate dynamics model learning and policy optimization of rewards. A number of interrelated solution categories to the objective mismatch problem have emerged as MBRL continues to mature as a research area. In this work, we provide an in-depth survey of these solution categories and propose a taxonomy to foster future research.

cs.LG↗

arXiv2024

Structural Estimation of Markov Decision Processes in High-Dimensional State Space with Finite-Time Guarantees

We consider the task of estimating a structural model of dynamic decisions by a human agent based upon the observable history of implemented actions and visited states. This problem has an inherent nested structure: in the inner problem, an optimal policy for a given reward function is identified while in the outer problem, a measure of fit is maximized. Several approaches have been proposed to alleviate the computational burden of this nested-loop structure, but these methods still suffer from high complexity when the state space is either discrete with large cardinality or continuous in high dimensions. Other approaches in the inverse reinforcement learning (IRL) literature emphasize policy estimation at the expense of reduced reward estimation accuracy. In this paper we propose a single-loop estimation algorithm with finite time guarantees that is equipped to deal with high-dimensional state spaces without compromising reward estimation accuracy. In the proposed algorithm, each policy improvement step is followed by a stochastic gradient step for likelihood maximization. We show that the proposed algorithm converges to a stationary solution with a finite-time guarantee. Further, if the reward is parameterized linearly, we show that the algorithm approximates the maximum likelihood estimator sublinearly. Finally, by using robotics control problems in MuJoCo and their transfer settings, we show that the proposed algorithm achieves superior performance compared with other IRL and imitation learning benchmarks.

cs.LG↗