Rémy Portelas

Grgur Kovač, Masataka Sawayama, Rémy Portelas et al.

Large Language Models (LLMs) are often misleadingly recognized as having a personality or a set of values. We argue that an LLM can be seen as a superposition of perspectives with different values and personality traits. LLMs exhibit context-dependent values and personality traits that change based on the induced perspective (as opposed to humans, who tend to have more coherent values and personality traits across contexts). We introduce the concept of perspective controllability, which refers to a model's affordance to adopt various perspectives with differing values and personality traits. In our experiments, we use questionnaires from psychology (PVQ, VSM, IPIP) to study how exhibited values and personality traits change based on different perspectives. Through qualitative experiments, we show that LLMs express different values when those are (implicitly or explicitly) implied in the prompt, and that LLMs express different values even when those are not obviously implied (demonstrating their context-dependent nature). We then conduct quantitative experiments to study the controllability of different models (GPT-4, GPT-3.5, OpenAssistant, StableVicuna, StableLM), the effectiveness of various methods for inducing perspectives, and the smoothness of the models' drivability. We conclude by examining the broader implications of our work and outline a variety of associated scientific questions. The project website is available at https://sites.google.com/view/llm-superpositions .

Grgur Kovač, Rémy Portelas, Peter Ford Dominey et al.

Developmental psychologists have long-established the importance of socio-cognitive abilities in human intelligence. These abilities enable us to enter, participate and benefit from human culture. AI research on social interactive agents mostly concerns the emergence of culture in a multi-agent setting (often without a strong grounding in developmental psychology). We argue that AI research should be informed by psychology and study socio-cognitive abilities enabling to enter a culture too. We discuss the theories of Michael Tomasello and Jerome Bruner to introduce some of their concepts to AI and outline key concepts and socio-cognitive abilities. We present The SocialAI school - a tool including a customizable parameterized uite of procedurally generated environments, which simplifies conducting experiments regarding those concepts. We show examples of such experiments with RL agents and Large Language Models. The main motivation of this work is to engage the AI community around the problem of social intelligence informed by developmental psychology, and to provide a tool to simplify first steps in this direction. Refer to the project website for code and additional information: https://sites.google.com/view/socialai-school.

Clément Romac, Rémy Portelas, Katja Hofmann et al.

Training autonomous agents able to generalize to multiple tasks is a key target of Deep Reinforcement Learning (DRL) research. In parallel to improving DRL algorithms themselves, Automatic Curriculum Learning (ACL) study how teacher algorithms can train DRL agents more efficiently by adapting task selection to their evolving abilities. While multiple standard benchmarks exist to compare DRL agents, there is currently no such thing for ACL algorithms. Thus, comparing existing approaches is difficult, as too many experimental parameters differ from paper to paper. In this work, we identify several key challenges faced by ACL algorithms. Based on these, we present TeachMyAgent (TA), a benchmark of current ACL algorithms leveraging procedural task generation. It includes 1) challenge-specific unit-tests using variants of a procedural Box2D bipedal walker environment, and 2) a new procedural Parkour environment combining most ACL challenges, making it ideal for global performance assessment. We then use TeachMyAgent to conduct a comparative study of representative existing approaches, showcasing the competitiveness of some ACL algorithms that do not use expert knowledge. We also show that the Parkour environment remains an open problem. We open-source our environments, all studied ACL algorithms (collected from open-source code or re-implemented), and DRL students in a Python package available at https://github.com/flowersteam/TeachMyAgent.

Rémy Portelas, Cédric Colas, Katja Hofmann et al.

We consider the problem of how a teacher algorithm can enable an unknown Deep Reinforcement Learning (DRL) student to become good at a skill over a wide range of diverse environments. To do so, we study how a teacher algorithm can learn to generate a learning curriculum, whereby it sequentially samples parameters controlling a stochastic procedural generation of environments. Because it does not initially know the capacities of its student, a key challenge for the teacher is to discover which environments are easy, difficult or unlearnable, and in what order to propose them to maximize the efficiency of learning over the learnable ones. To achieve this, this problem is transformed into a surrogate continuous bandit problem where the teacher samples environments in order to maximize absolute learning progress of its student. We present a new algorithm modeling absolute learning progress with Gaussian mixture models (ALP-GMM). We also adapt existing algorithms and provide a complete study in the context of DRL. Using parameterized variants of the BipedalWalker environment, we study their efficiency to personalize a learning curriculum for different learners (embodiments), their robustness to the ratio of learnable/unlearnable environments, and their scalability to non-linear and high-dimensional parameter spaces. Videos and code are available at https://github.com/flowersteam/teachDeepRL.

Mathieu Petitbois, Rémy Portelas, Sylvain Lamprier

We study offline reinforcement learning of style-conditioned policies using explicit style supervision via subtrajectory labeling functions. In this setting, aligning style with high task performance is particularly challenging due to distribution shift and inherent conflicts between style and reward. Existing methods, despite introducing numerous definitions of style, often fail to reconcile these objectives effectively. To address these challenges, we propose a unified definition of behavior style and instantiate it into a practical framework. Building on this, we introduce Style-Conditioned Implicit Q-Learning (SCIQL), which leverages offline goal-conditioned RL techniques, such as hindsight relabeling and value learning, and combine it with a new Gated Advantage Weighted Regression mechanism to efficiently optimize task performance while preserving style alignment. Experiments demonstrate that SCIQL achieves superior performance on both objectives compared to prior offline methods. Code, datasets and visuals are available in: https://sciql-iclr-2026.github.io/.

Grgur Kovač, Rémy Portelas, Masataka Sawayama et al.

The standard way to study Large Language Models (LLMs) with benchmarks or psychology questionnaires is to provide many different queries from similar minimal contexts (e.g. multiple choice questions). However, due to LLMs' highly context-dependent nature, conclusions from such minimal-context evaluations may be little informative about the model's behavior in deployment (where it will be exposed to many new contexts). We argue that context-dependence (specifically, value stability) should be studied as a specific property of LLMs and used as another dimension of LLM comparison (alongside others such as cognitive abilities, knowledge, or model size). We present a case-study on the stability of value expression over different contexts (simulated conversations on different topics) as measured using a standard psychology questionnaire (PVQ) and on behavioral downstream tasks. Reusing methods from psychology, we study Rank-order stability on the population (interpersonal) level, and Ipsative stability on the individual (intrapersonal) level. We consider two settings (with and without instructing LLMs to simulate particular personas), two simulated populations, and three downstream tasks. We observe consistent trends in the stability of models and model families - Mixtral, Mistral, GPT-3.5 and Qwen families are more stable than LLaMa-2 and Phi. The consistency of these trends implies that some models exhibit higher value stability than others, and that stability can be estimated with the set of introduced methodological tools. When instructed to simulate particular personas, LLMs exhibit low Rank-order stability, which further diminishes with conversation length. This highlights the need for future research on LLMs that coherently simulate different personas. This paper provides a foundational step in that direction, and, to our knowledge, it is the first study of value stability in LLMs.

Emilien Biré, Anthony Kobanda, Ludovic Denoyer et al.

Developing agents for complex and underspecified tasks, where no clear objective exists, remains challenging but offers many opportunities. This is especially true in video games, where simulated players (bots) need to play realistically, and there is no clear reward to evaluate them. While imitation learning has shown promise in such domains, these methods often fail when agents encounter out-of-distribution scenarios during deployment. Expanding the training dataset is a common solution, but it becomes impractical or costly when relying on human demonstrations. This article addresses active imitation learning, aiming to trigger expert intervention only when necessary, reducing the need for constant expert input along training. We introduce Random Network Distillation DAgger (RND-DAgger), a new active imitation learning method that limits expert querying by using a learned state-based out-of-distribution measure to trigger interventions. This approach avoids frequent expert-agent action comparisons, thus making the expert intervene only when it is useful. We evaluate RND-DAgger against traditional imitation learning and other active approaches in 3D video games (racing and third-person navigation) and in a robotic locomotion task and show that RND-DAgger surpasses previous methods by reducing expert queries. https://sites.google.com/view/rnd-dagger

Anthony Kobanda, Rémy Portelas, Odalric-Ambrym Maillard et al.

We consider a Continual Reinforcement Learning setup, where a learning agent must continuously adapt to new tasks while retaining previously acquired skill sets, with a focus on the challenge of avoiding forgetting past gathered knowledge and ensuring scalability with the growing number of tasks. Such issues prevail in autonomous robotics and video game simulations, notably for navigation tasks prone to topological or kinematic changes. To address these issues, we introduce HiSPO, a novel hierarchical framework designed specifically for continual learning in navigation settings from offline data. Our method leverages distinct policy subspaces of neural networks to enable flexible and efficient adaptation to new tasks while preserving existing knowledge. We demonstrate, through a careful experimental study, the effectiveness of our method in both classical MuJoCo maze environments and complex video game-like navigation simulations, showcasing competitive performances and satisfying adaptability with respect to classical continual learning metrics, in particular regarding the memory usage and efficiency.

Anthony Kobanda, Odalric-Ambrym Maillard, Rémy Portelas

Autonomous agents operating in domains such as robotics or video game simulations must adapt to changing tasks without forgetting about the previous ones. This process called Continual Reinforcement Learning poses non-trivial difficulties, from preventing catastrophic forgetting to ensuring the scalability of the approaches considered. Building on recent advances, we introduce a benchmark providing a suite of video-game navigation scenarios, thus filling a gap in the literature and capturing key challenges : catastrophic forgetting, task adaptation, and memory efficiency. We define a set of various tasks and datasets, evaluation protocols, and metrics to assess the performance of algorithms, including state-of-the-art baselines. Our benchmark is designed not only to foster reproducible research and to accelerate progress in continual reinforcement learning for gaming, but also to provide a reproducible framework for production pipelines -- helping practitioners to identify and to apply effective approaches.

Grgur Kovač, Jérémy Perez, Rémy Portelas et al.

Large language models (LLMs) are increasingly used in the creation of online content, creating feedback loops as subsequent generations of models will be trained on this synthetic data. Such loops were shown to lead to distribution shifts - models misrepresenting the true underlying distributions of human data (also called model collapse). However, how human data properties affect such shifts remains poorly understood. In this paper, we provide the first empirical examination of the effect of such properties on the outcome of recursive training. We first confirm that using different human datasets leads to distribution shifts of different magnitudes. Through exhaustive manipulation of dataset properties combined with regression analyses, we then identify a set of properties predicting distribution shift magnitudes. Lexical diversity is found to amplify these shifts, while semantic diversity and data quality mitigate them. Furthermore, we find that these influences are highly modular: data scrapped from a given internet domain has little influence on the content generated for another domain. Finally, experiments on political bias reveal that human data properties affect whether the initial bias will be amplified or reduced. Overall, our results portray a novel view, where different parts of internet may undergo different types of distribution shift.

Alexi Canesse, Mathieu Petitbois, Ludovic Denoyer et al.

Offline Reinforcement Learning (RL) has emerged as a powerful alternative to imitation learning for behavior modeling in various domains, particularly in complex navigation tasks. An existing challenge with Offline RL is the signal-to-noise ratio, i.e. how to mitigate incorrect policy updates due to errors in value estimates. Towards this, multiple works have demonstrated the advantage of hierarchical offline RL methods, which decouples high-level path planning from low-level path following. In this work, we present a novel hierarchical transformer-based approach leveraging a learned quantizer of the space. This quantization enables the training of a simpler zone-conditioned low-level policy and simplifies planning, which is reduced to discrete autoregressive prediction. Among other benefits, zone-level reasoning in planning enables explicit trajectory stitching rather than implicit stitching based on noisy value function estimates. By combining this transformer-based planner with recent advancements in offline RL, our proposed approach achieves state-of-the-art results in complex long-distance navigation environments.

Anthony Kobanda, Waris Radji, Mathieu Petitbois et al.

Offline Goal-Conditioned Reinforcement Learning seeks to train agents to reach specified goals from previously collected trajectories. Scaling that promises to long-horizon tasks remains challenging, notably due to compounding value-estimation errors. Principled geometric offers a potential solution to address these issues. Following this insight, we introduce Projective Quasimetric Planning (ProQ), a compositional framework that learns an asymmetric distance and then repurposes it, firstly as a repulsive energy forcing a sparse set of keypoints to uniformly spread over the learned latent space, and secondly as a structured directional cost guiding towards proximal sub-goals. In particular, ProQ couples this geometry with a Lagrangian out-of-distribution detector to ensure the learned keypoints stay within reachable areas. By unifying metric learning, keypoint coverage, and goal-conditioned control, our approach produces meaningful sub-goals and robustly drives long-horizon goal-reaching on diverse a navigation benchmarks.

Mathieu Petitbois, Rémy Portelas, Sylvain Lamprier et al.

Imitation Learning (IL) techniques aim to replicate human behaviors in specific tasks. While IL has gained prominence due to its effectiveness and efficiency, traditional methods often focus on datasets collected from experts to produce a single efficient policy. Recently, extensions have been proposed to handle datasets of diverse behaviors by mainly focusing on learning transition-level diverse policies or on performing entropy maximization at the trajectory level. While these methods may lead to diverse behaviors, they may not be sufficient to reproduce the actual diversity of demonstrations or to allow controlled trajectory generation. To overcome these drawbacks, we propose a different method based on two key features: a) Temporal Consistency that ensures consistent behaviors across entire episodes and not just at the transition level as well as b) Controllability obtained by constructing a latent space of behaviors that allows users to selectively activate specific behaviors based on their requirements. We compare our approach to state-of-the-art methods over a diverse set of tasks and environments. Project page: https://mathieu-petitbois.github.io/projects/swr/

Grgur Kovač, Rémy Portelas, Katja Hofmann et al.

Building embodied autonomous agents capable of participating in social interactions with humans is one of the main challenges in AI. Within the Deep Reinforcement Learning (DRL) field, this objective motivated multiple works on embodied language use. However, current approaches focus on language as a communication tool in very simplified and non-diverse social situations: the "naturalness" of language is reduced to the concept of high vocabulary size and variability. In this paper, we argue that aiming towards human-level AI requires a broader set of key social skills: 1) language use in complex and variable social contexts; 2) beyond language, complex embodied communication in multimodal settings within constantly evolving social worlds. We explain how concepts from cognitive sciences could help AI to draw a roadmap towards human-like intelligence, with a focus on its social dimensions. As a first step, we propose to expand current research to a broader set of core social skills. To do this, we present SocialAI, a benchmark to assess the acquisition of social skills of DRL agents using multiple grid-world environments featuring other (scripted) social agents. We then study the limits of a recent SOTA DRL approach when tested on SocialAI and discuss important next steps towards proficient social agents. Videos and code are available at https://sites.google.com/view/socialai.

Grgur Kovač, Rémy Portelas, Katja Hofmann et al.

Building embodied autonomous agents capable of participating in social interactions with humans is one of the main challenges in AI. This problem motivated many research directions on embodied language use. Current approaches focus on language as a communication tool in very simplified and non diverse social situations: the "naturalness" of language is reduced to the concept of high vocabulary size and variability. In this paper, we argue that aiming towards human-level AI requires a broader set of key social skills: 1) language use in complex and variable social contexts; 2) beyond language, complex embodied communication in multimodal settings within constantly evolving social worlds. In this work we explain how concepts from cognitive sciences could help AI to draw a roadmap towards human-like intelligence, with a focus on its social dimensions. We then study the limits of a recent SOTA Deep RL approach when tested on a first grid-world environment from the upcoming SocialAI, a benchmark to assess the social skills of Deep RL agents. Videos and code are available at https://sites.google.com/view/socialai01 .

Rémy Portelas, Clément Romac, Katja Hofmann et al.

A major challenge in the Deep RL (DRL) community is to train agents able to generalize their control policy over situations never seen in training. Training on diverse tasks has been identified as a key ingredient for good generalization, which pushed researchers towards using rich procedural task generation systems controlled through complex continuous parameter spaces. In such complex task spaces, it is essential to rely on some form of Automatic Curriculum Learning (ACL) to adapt the task sampling distribution to a given learning agent, instead of randomly sampling tasks, as many could end up being either trivial or unfeasible. Since it is hard to get prior knowledge on such task spaces, many ACL algorithms explore the task space to detect progress niches over time, a costly tabula-rasa process that needs to be performed for each new learning agents, although they might have similarities in their capabilities profiles. To address this limitation, we introduce the concept of Meta-ACL, and formalize it in the context of black-box RL learners, i.e. algorithms seeking to generalize curriculum generation to an (unknown) distribution of learners. In this work, we present AGAIN, a first instantiation of Meta-ACL, and showcase its benefits for curriculum generation over classical ACL in multiple simulated environments including procedurally generated parkour environments with learners of varying morphologies. Videos and code are available at https://sites.google.com/view/meta-acl .

Rémy Portelas, Katja Hofmann, Pierre-Yves Oudeyer

A major challenge in the Deep RL (DRL) community is to train agents able to generalize over unseen situations, which is often approached by training them on a diversity of tasks (or environments). A powerful method to foster diversity is to procedurally generate tasks by sampling their parameters from a multi-dimensional distribution, enabling in particular to propose a different task for each training episode. In practice, to get the high diversity of training tasks necessary for generalization, one has to use complex procedural generation systems. With such generators, it is hard to get prior knowledge on the subset of tasks that are actually learnable at all (many generated tasks may be unlearnable), what is their relative difficulty and what is the most efficient task distribution ordering for training. A typical solution in such cases is to rely on some form of Automated Curriculum Learning (ACL) to adapt the sampling distribution. One limit of current approaches is their need to explore the task space to detect progress niches over time, which leads to a loss of time. Additionally, we hypothesize that the induced noise in the training data may impair the performances of brittle DRL learners. We address this problem by proposing a two stage ACL approach where 1) a teacher algorithm first learns to train a DRL agent with a high-exploration curriculum, and then 2) distills learned priors from the first run to generate an "expert curriculum" to re-train the same agent from scratch. Besides demonstrating 50% improvements on average over the current state of the art, the objective of this work is to give a first example of a new research direction oriented towards refining ACL techniques over multiple learners, which we call Classroom Teaching.

Rémy Portelas, Cédric Colas, Lilian Weng et al.

Automatic Curriculum Learning (ACL) has become a cornerstone of recent successes in Deep Reinforcement Learning (DRL).These methods shape the learning trajectories of agents by challenging them with tasks adapted to their capacities. In recent years, they have been used to improve sample efficiency and asymptotic performance, to organize exploration, to encourage generalization or to solve sparse reward problems, among others. The ambition of this work is dual: 1) to present a compact and accessible introduction to the Automatic Curriculum Learning literature and 2) to draw a bigger picture of the current state of the art in ACL to encourage the cross-breeding of existing concepts and the emergence of new ideas.

Nicolas Lair, Cédric Colas, Rémy Portelas et al.

Autonomous reinforcement learning agents, like children, do not have access to predefined goals and reward functions. They must discover potential goals, learn their own reward functions and engage in their own learning trajectory. Children, however, benefit from exposure to language, helping to organize and mediate their thought. We propose LE2 (Language Enhanced Exploration), a learning algorithm leveraging intrinsic motivations and natural language (NL) interactions with a descriptive social partner (SP). Using NL descriptions from the SP, it can learn an NL-conditioned reward function to formulate goals for intrinsically motivated goal exploration and learn a goal-conditioned policy. By exploring, collecting descriptions from the SP and jointly learning the reward function and the policy, the agent grounds NL descriptions into real behavioral goals. From simple goals discovered early to more complex goals discovered by experimenting on simpler ones, our agent autonomously builds its own behavioral repertoire. This naturally occurring curriculum is supplemented by an active learning curriculum resulting from the agent's intrinsic motivations. Experiments are presented with a simulated robotic arm that interacts with several objects including tools.

Sébastien Forestier, Rémy Portelas, Yoan Mollard et al.

Intrinsically motivated spontaneous exploration is a key enabler of autonomous developmental learning in human children. It enables the discovery of skill repertoires through autotelic learning, i.e. the self-generation, self-selection, self-ordering and self-experimentation of learning goals. We present an algorithmic approach called Intrinsically Motivated Goal Exploration Processes (IMGEP) to enable similar properties of autonomous learning in machines. The IMGEP architecture relies on several principles: 1) self-generation of goals, generalized as parameterized fitness functions; 2) selection of goals based on intrinsic rewards; 3) exploration with incremental goal-parameterized policy search and exploitation with a batch learning algorithm; 4) systematic reuse of information acquired when targeting a goal for improving towards other goals. We present a particularly efficient form of IMGEP, called AMB, that uses a population-based policy and an object-centered spatio-temporal modularity. We provide several implementations of this architecture and demonstrate their ability to automatically generate a learning curriculum within several experimental setups. One of these experiments includes a real humanoid robot exploring multiple spaces of goals with several hundred continuous dimensions and with distractors. While no particular target goal is provided to these autotelic agents, this curriculum allows the discovery of diverse skills that act as stepping stones for learning more complex skills, e.g. nested tool use.