Direct Alignment Algorithms (DAAs), such as Direct Preference Optimisation (DPO) and Identity Preference Optimisation (IPO), have emerged as alternatives to online Reinforcement Learning from Human Feedback (RLHF) algorithms such as Proximal Policy Optimisation (PPO) for aligning language models to human preferences, without the need for explicit reward modelling. These methods generally aim to increase the likelihood of generating better (preferred) completions while discouraging worse (non-preferred) ones, while staying close to the original model's behaviour. In this work, we explore the relationship between completion likelihood and model performance in state-of-the-art DAAs, and identify a critical issue of likelihood over-optimisation. Contrary to expectations, we find that higher likelihood of better completions and larger margins between better and worse completion likelihoods do not necessarily lead to better performance, and may even degrade it. Our analysis reveals that while higher likelihood correlates with better memorisation of factual knowledge patterns, a slightly lower completion likelihood tends to improve output diversity, thus leading to better generalisation to unseen scenarios. Moreover, we identify two key indicators that signal when over-optimised output diversity begins to harm performance: Decreasing Entropy over Top-k Tokens and Diminishing Top-k Probability Mass. Our experimental results validate that these indicators are reliable signs of declining performance under different regularisation schemes, helping prevent overoptimisation and improve alignment with human preferences. Recent advancements in Large Language Models (LLMs) (Touvron et al., 2023; Achiam et al., 2023; Roziere et al., 2023; Dubey et al., 2024; Land & Bartolo, 2024) have significantly expanded their capabilities, enabling applications such as code generation, tool use, and interactive communication. As LLMs become increasingly powerful, the challenge of aligning them with human preferences has grown in importance. Direct Alignment Algorithms (DAAs), such as Direct Preference Optimisation (DPO) (Rafailov et al., 2023) and Identity Preference Optimisation (IPO) (Azar et al., 2024), have emerged as alternatives to Reinforcement Learning from Human Feedback (RLHF) (Ziegler et al., 2019; Bai et al., 2022) for training LMs on human preference data. These methods aim to bypass the traditional RLHF pipeline by directly optimising the policy without explicit reward modelling. DAAs are designed to increase the likelihood of better completions while reducing the likelihood of worse ones, all while staying close to the original model's behaviour.

Designing control policies whose performance level is guaranteed to remain above a given threshold in a span of environments is a critical feature for the adoption of reinforcement learning (RL) in real-world applications. The search for such robust policies is a notoriously difficult problem, related to the so-called dynamic model of transition function uncertainty, where the environment dynamics are allowed to change at each time step. But in practical cases, one is rather interested in robustness to a span of static transition models throughout interaction episodes. The static model is known to be harder to solve than the dynamic one, and seminal algorithms, such as robust value iteration, as well as most recent works on deep robust RL, build upon the dynamic model. In this work, we propose to revisit the static model. We suggest an analysis of why solving the static model under some mild hypotheses is a reasonable endeavor, based on an equivalence with the dynamic model, and formalize the general intuition that robust MDPs can be solved by tackling a series of static problems. We introduce a generic meta-algorithm called IWOCS, which incrementally identifies worst-case transition models so as to guide the search for a robust policy. Discussion on IWOCS sheds light on new ways to decouple policy optimization and adversarial transition functions and opens new perspectives for analysis. We derive a deep RL version of IWOCS and demonstrate it is competitive with state-of-the-art algorithms on classical benchmarks.

Reinforcement Learning (RL) has been used to finetune Large Language Models (LLMs) using a reward model trained from preference data, to better align with human judgment. The recently introduced direct alignment methods, which are often simpler, more stable, and computationally lighter, can more directly achieve this. However, these approaches cannot optimize arbitrary rewards, and the preference-based ones are not the only rewards of interest for LLMs (eg., unit tests for code generation or textual entailment for summarization, among others). RL-finetuning is usually done with a variation of policy gradient, which calls for on-policy or near-on-policy samples, requiring costly generations. We introduce Contrastive Policy Gradient, or CoPG, a simple and mathematically principled new RL algorithm that can estimate the optimal policy even from off-policy data. It can be seen as an off-policy policy gradient approach that does not rely on important sampling techniques and highlights the importance of using (the right) state baseline. We show this approach to generalize the direct alignment method IPO (identity preference optimization) and classic policy gradient. We experiment with the proposed CoPG on a toy bandit problem to illustrate its properties, as well as for finetuning LLMs on a summarization task, using a learned reward function considered as ground truth for the purpose of the experiments.

Robust reinforcement learning is essential for deploying reinforcement learning algorithms in real-world scenarios where environmental uncertainty predominates. Traditional robust reinforcement learning often depends on rectangularity assumptions, where adverse probability measures of outcome states are assumed to be independent across different states and actions. This assumption, rarely fulfilled in practice, leads to overly conservative policies. To address this problem, we introduce a new time-constrained robust MDP (TC-RMDP) formulation that considers multifactorial, correlated, and time-dependent disturbances, thus more accurately reflecting real-world dynamics. This formulation goes beyond the conventional rectangularity paradigm, offering new perspectives and expanding the analytical framework for robust RL. We propose three distinct algorithms, each using varying levels of environmental information, and evaluate them extensively on continuous control benchmarks. Our results demonstrate that these algorithms yield an efficient tradeoff between performance and robustness, outperforming traditional deep robust RL methods in time-constrained environments while preserving robustness in classical benchmarks. This study revisits the prevailing assumptions in robust RL and opens new avenues for developing more practical and realistic RL applications.

Robust reinforcement learning is the problem of learning control policies that provide optimal worst-case performance against a span of adversarial environments. It is a crucial ingredient for deploying algorithms in real-world scenarios with prevalent environmental uncertainties and has been a long-standing object of attention in the community, without a standardized set of benchmarks. This contribution endeavors to fill this gap. We introduce the Robust Reinforcement Learning Suite (RRLS), a benchmark suite based on Mujoco environments. RRLS provides six continuous control tasks with two types of uncertainty sets for training and evaluation. Our benchmark aims to standardize robust reinforcement learning tasks, facilitating reproducible and comparable experiments, in particular those from recent state-of-the-art contributions, for which we demonstrate the use of RRLS. It is also designed to be easily expandable to new environments. The source code is available at \href{https://github.com/SuReLI/RRLS}{https://github.com/SuReLI/RRLS}.

Reinforcement Learning from Human Feedback (RLHF) (Christiano et al., 2017) has rapidly become a standard method to align Large Language Models (LLMs). One of the main practical issues that all the prominent existing RLHF methods (offline or online) (Ouyang et al., 2022; Rafailov et al., 2023; Azar et al., 2023; Zhao et al., 2023b; Ahmadian et al., 2024) encounter is that their optimal solution heavily depends on the training task in terms of the distribution used to generate the preference data (behavior policy) (Munos et al., 2023; Azar et al., 2023). This makes the existing RLHF methods prone to out-of-distribution (OOD) tasks (Li et al., 2024; Kirk et al., 2024) where the evaluation distribution is significantly different from that of the behavior policy. Also, whenever the base/SFT models significantly differ from the behavior policy, the dependency of the RLHF solutions on the behavior policy makes the preference dataset and reward model less useful (Gao et al., 2022) as RLHF may undo the SFT/pretraining. To address this challenge, we introduce an alternative approach for aligning LLMs from human preferences based on more principled and robust foundations. Our goal is to find a solution that is robust to the changes in the preference dataset, meaning that changes in the distribution from which the completions are sampled do not affect the final outcome of learning significantly. To achieve this goal, we exploit the concept of self-improving (Huang et al., 2022; Bai et al., 2022) language models. By self-improving LLM we refer to a model capable of enhancing its outputs recursively with each inference iteration.

Many classic Reinforcement Learning (RL) algorithms rely on a Bellman operator, which involves an expectation over the next states, leading to the concept of bootstrapping. To introduce a form of pessimism, we propose to replace this expectation with an expectile. In practice, this can be very simply done by replacing the $L_2$ loss with a more general expectile loss for the critic. Introducing pessimism in RL is desirable for various reasons, such as tackling the overestimation problem (for which classic solutions are double Q-learning or the twin-critic approach of TD3) or robust RL (where transitions are adversarial). We study empirically these two cases. For the overestimation problem, we show that the proposed approach, ExpectRL, provides better results than a classic twin-critic. On robust RL benchmarks, involving changes of the environment, we show that our approach is more robust than classic RL algorithms. We also introduce a variation of ExpectRL combined with domain randomization which is competitive with state-of-the-art robust RL agents. Eventually, we also extend \ExpectRL with a mechanism for choosing automatically the expectile value, that is the degree of pessimism

The ability to autonomously assemble structures is crucial for the development of future space infrastructure. However, the unpredictable conditions of space pose significant challenges for robotic systems, necessitating the development of advanced learning techniques to enable autonomous assembly. In this study, we present a novel approach for learning autonomous peg-in-hole assembly in the context of space robotics. Our focus is on enhancing the generalization and adaptability of autonomous systems through deep reinforcement learning. By integrating procedural generation and domain randomization, we train agents in a highly parallelized simulation environment across a spectrum of diverse scenarios with the aim of acquiring a robust policy. The proposed approach is evaluated using three distinct reinforcement learning algorithms to investigate the trade-offs among various paradigms. We demonstrate the adaptability of our agents to novel scenarios and assembly sequences while emphasizing the potential of leveraging advanced simulation techniques for robot learning in space. Our findings set the stage for future advancements in intelligent robotic systems capable of supporting ambitious space missions and infrastructure development beyond Earth.

Mean Field Games (MFGs) have the ability to handle large-scale multi-agent systems, but learning Nash equilibria in MFGs remains a challenging task. In this paper, we propose a deep reinforcement learning (DRL) algorithm that achieves population-dependent Nash equilibrium without the need for averaging or sampling from history, inspired by Munchausen RL and Online Mirror Descent. Through the design of an additional inner-loop replay buffer, the agents can effectively learn to achieve Nash equilibrium from any distribution, mitigating catastrophic forgetting. The resulting policy can be applied to various initial distributions. Numerical experiments on four canonical examples demonstrate our algorithm has better convergence properties than SOTA algorithms, in particular a DRL version of Fictitious Play for population-dependent policies.

Non-cooperative and cooperative games with a very large number of players have many applications but remain generally intractable when the number of players increases. Introduced by Lasry and Lions, and Huang, Caines and Malham\'e, Mean Field Games (MFGs) rely on a mean-field approximation to allow the number of players to grow to infinity. Traditional methods for solving these games generally rely on solving partial or stochastic differential equations with a full knowledge of the model. Recently, Reinforcement Learning (RL) has appeared promising to solve complex problems at scale. The combination of RL and MFGs is promising to solve games at a very large scale both in terms of population size and environment complexity. In this survey, we review the quickly growing recent literature on RL methods to learn equilibria and social optima in MFGs. We first identify the most common settings (static, stationary, and evolutive) of MFGs. We then present a general framework for classical iterative methods (based on best-response computation or policy evaluation) to solve MFGs in an exact way. Building on these algorithms and the connection with Markov Decision Processes, we explain how RL can be used to learn MFG solutions in a model-free way. Last, we present numerical illustrations on a benchmark problem, and conclude with some perspectives.