Teaching Large Language Models to Reason with Reinforcement Learning

Reinforcement Learning from Human Feedback (\textbf{RLHF}) has emerged as a dominant approach for aligning LLM outputs with human preferences. Inspired by the success of RLHF, we study the performance of multiple algorithms that learn from feedback (Expert Iteration, Proximal Policy Optimization (\textbf{PPO}), Return-Conditioned RL) on improving LLM reasoning capabilities. We investigate both sparse and dense rewards provided to the LLM both heuristically and via a learned reward model. We additionally start from multiple model sizes and initializations both with and without supervised fine-tuning (\textbf{SFT}) data. Overall, we find all algorithms perform comparably, with Expert Iteration performing best in most cases. Surprisingly, we find the sample complexity of Expert Iteration is similar to that of PPO, requiring at most on the order of $10^6$ samples to converge from a pretrained checkpoint. We investigate why this is the case, concluding that during RL training models fail to explore significantly beyond solutions already produced by SFT models. Additionally, we discuss a trade off between maj@1 and pass@96 metric performance during SFT training and how conversely RL training improves both simultaneously. We then conclude by discussing the implications of our findings for RLHF and the future role of RL in LLM fine-tuning.

Rainbow Teaming: Open-Ended Generation of Diverse Adversarial Prompts

As large language models (LLMs) become increasingly prevalent across many real-world applications, understanding and enhancing their robustness to user inputs is of paramount importance. Existing methods for identifying adversarial prompts tend to focus on specific domains, lack diversity, or require extensive human annotations. To address these limitations, we present Rainbow Teaming, a novel approach for producing a diverse collection of adversarial prompts. Rainbow Teaming casts adversarial prompt generation as a quality-diversity problem, and uses open-ended search to generate prompts that are both effective and diverse. It can uncover a model's vulnerabilities across a broad range of domains including, in this paper, safety, question answering, and cybersecurity. We also demonstrate that fine-tuning on synthetic data generated by Rainbow Teaming improves the safety of state-of-the-art LLMs without hurting their general capabilities and helpfulness, paving the path to open-ended self-improvement.

Generalization to New Sequential Decision Making Tasks with In-Context Learning

Training autonomous agents that can learn new tasks from only a handful of demonstrations is a long-standing problem in machine learning. Recently, transformers have been shown to learn new language or vision tasks without any weight updates from only a few examples, also referred to as in-context learning. However, the sequential decision making setting poses additional challenges having a lower tolerance for errors since the environment's stochasticity or the agent's actions can lead to unseen, and sometimes unrecoverable, states. In this paper, we use an illustrative example to show that naively applying transformers to sequential decision making problems does not enable in-context learning of new tasks. We then demonstrate how training on sequences of trajectories with certain distributional properties leads to in-context learning of new sequential decision making tasks. We investigate different design choices and find that larger model and dataset sizes, as well as more task diversity, environment stochasticity, and trajectory burstiness, all result in better in-context learning of new out-of-distribution tasks. By training on large diverse offline datasets, our model is able to learn new MiniHack and Procgen tasks without any weight updates from just a handful of demonstrations.

Multi-Objective GFlowNets

In many applications of machine learning, like drug discovery and material design, the goal is to generate candidates that simultaneously maximize a set of objectives. As these objectives are often conflicting, there is no single candidate that simultaneously maximizes all objectives, but rather a set of Pareto-optimal candidates where one objective cannot be improved without worsening another. Moreover, in practice, these objectives are often under-specified, making the diversity of candidates a key consideration. The existing multi-objective optimization methods focus predominantly on covering the Pareto front, failing to capture diversity in the space of candidates. Motivated by the success of GFlowNets for generation of diverse candidates in a single objective setting, in this paper we consider Multi-Objective GFlowNets (MOGFNs). MOGFNs consist of a novel Conditional GFlowNet which models a family of single-objective sub-problems derived by decomposing the multi-objective optimization problem. Our work is the first to empirically demonstrate conditional GFlowNets. Through a series of experiments on synthetic and benchmark tasks, we empirically demonstrate that MOGFNs outperform existing methods in terms of Hypervolume, R2-distance and candidate diversity. We also demonstrate the effectiveness of MOGFNs over existing methods in active learning settings. Finally, we supplement our empirical results with a careful analysis of each component of MOGFNs.

Continual Learning In Environments With Polynomial Mixing Times

The mixing time of the Markov chain induced by a policy limits performance in real-world continual learning scenarios. Yet, the effect of mixing times on learning in continual reinforcement learning (RL) remains underexplored. In this paper, we characterize problems that are of long-term interest to the development of continual RL, which we call scalable MDPs, through the lens of mixing times. In particular, we establish that scalable MDPs have mixing times that scale polynomially with the size of the problem. We go on to demonstrate that polynomial mixing times present significant difficulties for existing approaches and propose a family of model-based algorithms that speed up learning by directly optimizing for the average reward through a novel bootstrapping procedure. Finally, we perform empirical regret analysis of our proposed approaches, demonstrating clear improvements over baselines and also how scalable MDPs can be used for analysis of RL algorithms as mixing times scale.

Compositional Attention: Disentangling Search and Retrieval

Multi-head, key-value attention is the backbone of the widely successful Transformer model and its variants. This attention mechanism uses multiple parallel key-value attention blocks (called heads), each performing two fundamental computations: (1) search - selection of a relevant entity from a set via query-key interactions, and (2) retrieval - extraction of relevant features from the selected entity via a value matrix. Importantly, standard attention heads learn a rigid mapping between search and retrieval. In this work, we first highlight how this static nature of the pairing can potentially: (a) lead to learning of redundant parameters in certain tasks, and (b) hinder generalization. To alleviate this problem, we propose a novel attention mechanism, called Compositional Attention, that replaces the standard head structure. The proposed mechanism disentangles search and retrieval and composes them in a dynamic, flexible and context-dependent manner through an additional soft competition stage between the query-key combination and value pairing. Through a series of numerical experiments, we show that it outperforms standard multi-head attention on a variety of tasks, including some out-of-distribution settings. Through our qualitative analysis, we demonstrate that Compositional Attention leads to dynamic specialization based on the type of retrieval needed. Our proposed mechanism generalizes multi-head attention, allows independent scaling of search and retrieval, and can easily be implemented in lieu of standard attention heads in any network architecture.

Curriculum in Gradient-Based Meta-Reinforcement Learning

Gradient-based meta-learners such as Model-Agnostic Meta-Learning (MAML) have shown strong few-shot performance in supervised and reinforcement learning settings. However, specifically in the case of meta-reinforcement learning (meta-RL), we can show that gradient-based meta-learners are sensitive to task distributions. With the wrong curriculum, agents suffer the effects of meta-overfitting, shallow adaptation, and adaptation instability. In this work, we begin by highlighting intriguing failure cases of gradient-based meta-RL and show that task distributions can wildly affect algorithmic outputs, stability, and performance. To address this problem, we leverage insights from recent literature on domain randomization and propose meta Active Domain Randomization (meta-ADR), which learns a curriculum of tasks for gradient-based meta-RL in a similar as ADR does for sim2real transfer. We show that this approach induces more stable policies on a variety of simulated locomotion and navigation tasks. We assess in- and out-of-distribution generalization and find that the learned task distributions, even in an unstructured task space, greatly improve the adaptation performance of MAML. Finally, we motivate the need for better benchmarking in meta-RL that prioritizes \textit{generalization} over single-task adaption performance.

Generating Automatic Curricula via Self-Supervised Active Domain Randomization

Goal-directed Reinforcement Learning (RL) traditionally considers an agent interacting with an environment, prescribing a real-valued reward to an agent proportional to the completion of some goal. Goal-directed RL has seen large gains in sample efficiency, due to the ease of reusing or generating new experience by proposing goals. In this work, we build on the framework of self-play, allowing an agent to interact with itself in order to make progress on some unknown task. We use Active Domain Randomization and self-play to create a novel, coupled environment-goal curriculum, where agents learn through progressively more difficult tasks and environment variations. Our method, Self-Supervised Active Domain Randomization (SS-ADR), generates a growing curriculum, encouraging the agent to try tasks that are just outside of its current capabilities, while building a domain-randomization curriculum that enables state-of-the-art results on various sim2real transfer tasks. Our results show that a curriculum of co-evolving the environment difficulty along with the difficulty of goals set in each environment provides practical benefits in the goal-directed tasks tested.