Reinforcement Learning for Aligning Large Language Models Agents with Interactive Environments: Quantifying and Mitigating Prompt Overfitting

Reinforcement learning (RL) is a promising approach for aligning large language models (LLMs) knowledge with sequential decision-making tasks. However, few studies have thoroughly investigated the impact on LLM agents capabilities of fine-tuning them with RL in a specific environment. In this paper, we propose a novel framework to analyze the sensitivity of LLMs to prompt formulations following RL training in a textual environment. Our findings reveal that the performance of LLMs degrades when faced with prompt formulations different from those used during the RL training phase. Besides, we analyze the source of this sensitivity by examining the model's internal representations and salient tokens. Finally, we propose to use a contrastive loss to mitigate this sensitivity and improve the robustness and generalization capabilities of LLMs.

SAC-GLAM: Improving Online RL for LLM agents with Soft Actor-Critic and Hindsight Relabeling

The past years have seen Large Language Models (LLMs) strive not only as generative models but also as agents solving textual sequential decision-making tasks. When facing complex environments where their zero-shot abilities are insufficient, recent work showed online Reinforcement Learning (RL) could be used for the LLM agent to discover and learn efficient strategies interactively. However, most prior work sticks to on-policy algorithms, which greatly reduces the scope of methods such agents could use for both exploration and exploitation, such as experience replay and hindsight relabeling. Yet, such methods may be key for LLM learning agents, and in particular when designing autonomous intrinsically motivated agents sampling and pursuing their own goals (i.e. autotelic agents). This paper presents and studies an adaptation of Soft Actor-Critic and hindsight relabeling to LLM agents. Our method not only paves the path towards autotelic LLM agents that learn online but can also outperform on-policy methods in more classic multi-goal RL environments.

Bridging Environments and Language with Rendering Functions and Vision-Language Models

Vision-language models (VLMs) have tremendous potential for grounding language, and thus enabling language-conditioned agents (LCAs) to perform diverse tasks specified with text. This has motivated the study of LCAs based on reinforcement learning (RL) with rewards given by rendering images of an environment and evaluating those images with VLMs. If single-task RL is employed, such approaches are limited by the cost and time required to train a policy for each new task. Multi-task RL (MTRL) is a natural alternative, but requires a carefully designed corpus of training tasks and does not always generalize reliably to new tasks. Therefore, this paper introduces a novel decomposition of the problem of building an LCA: first find an environment configuration that has a high VLM score for text describing a task; then use a (pretrained) goal-conditioned policy to reach that configuration. We also explore several enhancements to the speed and quality of VLM-based LCAs, notably, the use of distilled models, and the evaluation of configurations from multiple viewpoints to resolve the ambiguities inherent in a single 2D view. We demonstrate our approach on the Humanoid environment, showing that it results in LCAs that outperform MTRL baselines in zero-shot generalization, without requiring any textual task descriptions or other forms of environment-specific annotation during training. Videos and an interactive demo can be found at https://europe.naverlabs.com/text2control

Physics-Informed Model and Hybrid Planning for Efficient Dyna-Style Reinforcement Learning

Applying reinforcement learning (RL) to real-world applications requires addressing a trade-off between asymptotic performance, sample efficiency, and inference time. In this work, we demonstrate how to address this triple challenge by leveraging partial physical knowledge about the system dynamics. Our approach involves learning a physics-informed model to boost sample efficiency and generating imaginary trajectories from this model to learn a model-free policy and Q-function. Furthermore, we propose a hybrid planning strategy, combining the learned policy and Q-function with the learned model to enhance time efficiency in planning. Through practical demonstrations, we illustrate that our method improves the compromise between sample efficiency, time efficiency, and performance over state-of-the-art methods.

Single-Reset Divide & Conquer Imitation Learning

Demonstrations are commonly used to speed up the learning process of Deep Reinforcement Learning algorithms. To cope with the difficulty of accessing multiple demonstrations, some algorithms have been developed to learn from a single demonstration. In particular, the Divide & Conquer Imitation Learning algorithms leverage a sequential bias to learn a control policy for complex robotic tasks using a single state-based demonstration. The latest version, DCIL-II demonstrates remarkable sample efficiency. This novel method operates within an extended Goal-Conditioned Reinforcement Learning framework, ensuring compatibility between intermediate and subsequent goals extracted from the demonstration. However, a fundamental limitation arises from the assumption that the system can be reset to specific states along the demonstrated trajectory, confining the application to simulated systems. In response, we introduce an extension called Single-Reset DCIL (SR-DCIL), designed to overcome this constraint by relying on a single initial state reset rather than sequential resets. To address this more challenging setting, we integrate two mechanisms inspired by the Learning from Demonstrations literature, including a Demo-Buffer and Value Cloning, to guide the agent toward compatible success states. In addition, we introduce Approximate Goal Switching to facilitate training to reach goals distant from the reset state. Our paper makes several contributions, highlighting the importance of the reset assumption in DCIL-II, presenting the mechanisms of SR-DCIL variants and evaluating their performance in challenging robotic tasks compared to DCIL-II. In summary, this work offers insights into the significance of reset assumptions in the framework of DCIL and proposes SR-DCIL, a first step toward a versatile algorithm capable of learning control policies under a weaker reset assumption.

A Definition of Open-Ended Learning Problems for Goal-Conditioned Agents

A lot of recent machine learning research papers have "Open-ended learning" in their title. But very few of them attempt to define what they mean when using the term. Even worse, when looking more closely there seems to be no consensus on what distinguishes open-ended learning from related concepts such as continual learning, lifelong learning or autotelic learning. In this paper, we contribute to fixing this situation. After illustrating the genealogy of the concept and more recent perspectives about what it truly means, we outline that open-ended learning is generally conceived as a composite notion encompassing a set of diverse properties. In contrast with these previous approaches, we propose to isolate a key elementary property of open-ended processes, which is to always produce novel elements from time to time over an infinite horizon. From there, we build the notion of open-ended learning problems and focus in particular on the subset of open-ended goal-conditioned reinforcement learning problems, as this framework facilitates the definition of learning a growing repertoire of skills. Finally, we highlight the work that remains to be performed to fill the gap between our elementary definition and the more involved notions of open-ended learning that developmental AI researchers may have in mind.

A Simple Open-Loop Baseline for Reinforcement Learning Locomotion Tasks

In search of the simplest baseline capable of competing with Deep Reinforcement Learning on locomotion tasks, we propose a biologically inspired model-free open-loop strategy. Drawing upon prior knowledge and harnessing the elegance of simple oscillators to generate periodic joint motions, it achieves respectable performance in five different locomotion environments, with a number of tunable parameters that is a tiny fraction of the thousands typically required by RL algorithms. Unlike RL methods, which are prone to performance degradation when exposed to sensor noise or failure, our open-loop oscillators exhibit remarkable robustness due to their lack of reliance on sensors. Furthermore, we showcase a successful transfer from simulation to reality using an elastic quadruped, all without the need for randomization or reward engineering.

Utility-based Adaptive Teaching Strategies using Bayesian Theory of Mind

Good teachers always tailor their explanations to the learners. Cognitive scientists model this process under the rationality principle: teachers try to maximise the learner's utility while minimising teaching costs. To this end, human teachers seem to build mental models of the learner's internal state, a capacity known as Theory of Mind (ToM). Inspired by cognitive science, we build on Bayesian ToM mechanisms to design teacher agents that, like humans, tailor their teaching strategies to the learners. Our ToM-equipped teachers construct models of learners' internal states from observations and leverage them to select demonstrations that maximise the learners' rewards while minimising teaching costs. Our experiments in simulated environments demonstrate that learners taught this way are more efficient than those taught in a learner-agnostic way. This effect gets stronger when the teacher's model of the learner better aligns with the actual learner's state, either using a more accurate prior or after accumulating observations of the learner's behaviour. This work is a first step towards social machines that teach us and each other, see https://teacher-with-tom.github.io.

Enhancing Agent Communication and Learning through Action and Language

We introduce a novel category of GC-agents capable of functioning as both teachers and learners. Leveraging action-based demonstrations and language-based instructions, these agents enhance communication efficiency. We investigate the incorporation of pedagogy and pragmatism, essential elements in human communication and goal achievement, enhancing the agents' teaching and learning capabilities. Furthermore, we explore the impact of combining communication modes (action and language) on learning outcomes, highlighting the benefits of a multi-modal approach.

Grounding Large Language Models in Interactive Environments with Online Reinforcement Learning

Recent works successfully leveraged Large Language Models' (LLM) abilities to capture abstract knowledge about world's physics to solve decision-making problems. Yet, the alignment between LLMs' knowledge and the environment can be wrong and limit functional competence due to lack of grounding. In this paper, we study an approach to achieve this alignment through functional grounding: we consider an agent using an LLM as a policy that is progressively updated as the agent interacts with the environment, leveraging online Reinforcement Learning to improve its performance to solve goals. Using an interactive textual environment designed to study higher-level forms of functional grounding, and a set of spatial and navigation tasks, we study several scientific questions: 1) Can LLMs boost sample efficiency for online learning of various RL tasks? 2) How can it boost different forms of generalization? 3) What is the impact of online learning? We study these questions by functionally grounding several variants (size, architecture) of FLAN-T5.