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.

Learning to Model the World with Language

To interact with humans in the world, agents need to understand the diverse types of language that people use, relate them to the visual world, and act based on them. While current agents learn to execute simple language instructions from task rewards, we aim to build agents that leverage diverse language that conveys general knowledge, describes the state of the world, provides interactive feedback, and more. Our key idea is that language helps agents predict the future: what will be observed, how the world will behave, and which situations will be rewarded. This perspective unifies language understanding with future prediction as a powerful self-supervised learning objective. We present Dynalang, an agent that learns a multimodal world model that predicts future text and image representations and learns to act from imagined model rollouts. Unlike traditional agents that use language only to predict actions, Dynalang acquires rich language understanding by using past language also to predict future language, video, and rewards. In addition to learning from online interaction in an environment, Dynalang can be pretrained on datasets of text, video, or both without actions or rewards. From using language hints in grid worlds to navigating photorealistic scans of homes, Dynalang utilizes diverse types of language to improve task performance, including environment descriptions, game rules, and instructions.

DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models

Learning from human feedback has been shown to improve text-to-image models. These techniques first learn a reward function that captures what humans care about in the task and then improve the models based on the learned reward function. Even though relatively simple approaches (e.g., rejection sampling based on reward scores) have been investigated, fine-tuning text-to-image models with the reward function remains challenging. In this work, we propose using online reinforcement learning (RL) to fine-tune text-to-image models. We focus on diffusion models, defining the fine-tuning task as an RL problem, and updating the pre-trained text-to-image diffusion models using policy gradient to maximize the feedback-trained reward. Our approach, coined DPOK, integrates policy optimization with KL regularization. We conduct an analysis of KL regularization for both RL fine-tuning and supervised fine-tuning. In our experiments, we show that DPOK is generally superior to supervised fine-tuning with respect to both image-text alignment and image quality.

Vision-Language Models as Success Detectors

Detecting successful behaviour is crucial for training intelligent agents. As such, generalisable reward models are a prerequisite for agents that can learn to generalise their behaviour. In this work we focus on developing robust success detectors that leverage large, pretrained vision-language models (Flamingo, Alayrac et al. (2022)) and human reward annotations. Concretely, we treat success detection as a visual question answering (VQA) problem, denoted SuccessVQA. We study success detection across three vastly different domains: (i) interactive language-conditioned agents in a simulated household, (ii) real world robotic manipulation, and (iii) "in-the-wild" human egocentric videos. We investigate the generalisation properties of a Flamingo-based success detection model across unseen language and visual changes in the first two domains, and find that the proposed method is able to outperform bespoke reward models in out-of-distribution test scenarios with either variation. In the last domain of "in-the-wild" human videos, we show that success detection on unseen real videos presents an even more challenging generalisation task warranting future work. We hope our initial results encourage further work in real world success detection and reward modelling.

Aligning Text-to-Image Models using Human Feedback

Deep generative models have shown impressive results in text-to-image synthesis. However, current text-to-image models often generate images that are inadequately aligned with text prompts. We propose a fine-tuning method for aligning such models using human feedback, comprising three stages. First, we collect human feedback assessing model output alignment from a set of diverse text prompts. We then use the human-labeled image-text dataset to train a reward function that predicts human feedback. Lastly, the text-to-image model is fine-tuned by maximizing reward-weighted likelihood to improve image-text alignment. Our method generates objects with specified colors, counts and backgrounds more accurately than the pre-trained model. We also analyze several design choices and find that careful investigations on such design choices are important in balancing the alignment-fidelity tradeoffs. Our results demonstrate the potential for learning from human feedback to significantly improve text-to-image models.

Guiding Pretraining in Reinforcement Learning with Large Language Models

Reinforcement learning algorithms typically struggle in the absence of a dense, well-shaped reward function. Intrinsically motivated exploration methods address this limitation by rewarding agents for visiting novel states or transitions, but these methods offer limited benefits in large environments where most discovered novelty is irrelevant for downstream tasks. We describe a method that uses background knowledge from text corpora to shape exploration. This method, called ELLM (Exploring with LLMs) rewards an agent for achieving goals suggested by a language model prompted with a description of the agent's current state. By leveraging large-scale language model pretraining, ELLM guides agents toward human-meaningful and plausibly useful behaviors without requiring a human in the loop. We evaluate ELLM in the Crafter game environment and the Housekeep robotic simulator, showing that ELLM-trained agents have better coverage of common-sense behaviors during pretraining and usually match or improve performance on a range of downstream tasks.

It Takes Four to Tango: Multiagent Selfplay for Automatic Curriculum Generation

We are interested in training general-purpose reinforcement learning agents that can solve a wide variety of goals. Training such agents efficiently requires automatic generation of a goal curriculum. This is challenging as it requires (a) exploring goals of increasing difficulty, while ensuring that the agent (b) is exposed to a diverse set of goals in a sample efficient manner and (c) does not catastrophically forget previously solved goals. We propose Curriculum Self Play (CuSP), an automated goal generation framework that seeks to satisfy these desiderata by virtue of a multi-player game with four agents. We extend the asymmetric curricula learning in PAIRED (Dennis et al., 2020) to a symmetrized game that carefully balances cooperation and competition between two off-policy student learners and two regret-maximizing teachers. CuSP additionally introduces entropic goal coverage and accounts for the non-stationary nature of the students, allowing us to automatically induce a curriculum that balances progressive exploration with anti-catastrophic exploitation. We demonstrate that our method succeeds at generating an effective curricula of goals for a range of control tasks, outperforming other methods at zero-shot test-time generalization to novel out-of-distribution goals.

Bayesian Imitation Learning for End-to-End Mobile Manipulation

In this work we investigate and demonstrate benefits of a Bayesian approach to imitation learning from multiple sensor inputs, as applied to the task of opening office doors with a mobile manipulator. Augmenting policies with additional sensor inputs, such as RGB + depth cameras, is a straightforward approach to improving robot perception capabilities, especially for tasks that may favor different sensors in different situations. As we scale multi-sensor robotic learning to unstructured real-world settings (e.g. offices, homes) and more complex robot behaviors, we also increase reliance on simulators for cost, efficiency, and safety. Consequently, the sim-to-real gap across multiple sensor modalities also increases, making simulated validation more difficult. We show that using the Variational Information Bottleneck (Alemi et al., 2016) to regularize convolutional neural networks improves generalization to held-out domains and reduces the sim-to-real gap in a sensor-agnostic manner. As a side effect, the learned embeddings also provide useful estimates of model uncertainty for each sensor. We demonstrate that our method is able to help close the sim-to-real gap and successfully fuse RGB and depth modalities based on understanding of the situational uncertainty of each sensor. In a real-world office environment, we achieve 96% task success, improving upon the baseline by +16%.

Practical Imitation Learning in the Real World via Task Consistency Loss

Recent work in visual end-to-end learning for robotics has shown the promise of imitation learning across a variety of tasks. Such approaches are expensive both because they require large amounts of real world training demonstrations and because identifying the best model to deploy in the real world requires time-consuming real-world evaluations. These challenges can be mitigated by simulation: by supplementing real world data with simulated demonstrations and using simulated evaluations to identify high performing policies. However, this introduces the well-known "reality gap" problem, where simulator inaccuracies decorrelate performance in simulation from that of reality. In this paper, we build on top of prior work in GAN-based domain adaptation and introduce the notion of a Task Consistency Loss (TCL), a self-supervised loss that encourages sim and real alignment both at the feature and action-prediction levels. We demonstrate the effectiveness of our approach by teaching a mobile manipulator to autonomously approach a door, turn the handle to open the door, and enter the room. The policy performs control from RGB and depth images and generalizes to doors not encountered in training data. We achieve 80% success across ten seen and unseen scenes using only ~16.2 hours of teleoperated demonstrations in sim and real. To the best of our knowledge, this is the first work to tackle latched door opening from a purely end-to-end learning approach, where the task of navigation and manipulation are jointly modeled by a single neural network.

Auto-Tuned Sim-to-Real Transfer

Policies trained in simulation often fail when transferred to the real world due to the `reality gap' where the simulator is unable to accurately capture the dynamics and visual properties of the real world. Current approaches to tackle this problem, such as domain randomization, require prior knowledge and engineering to determine how much to randomize system parameters in order to learn a policy that is robust to sim-to-real transfer while also not being too conservative. We propose a method for automatically tuning simulator system parameters to match the real world using only raw RGB images of the real world without the need to define rewards or estimate state. Our key insight is to reframe the auto-tuning of parameters as a search problem where we iteratively shift the simulation system parameters to approach the real-world system parameters. We propose a Search Param Model (SPM) that, given a sequence of observations and actions and a set of system parameters, predicts whether the given parameters are higher or lower than the true parameters used to generate the observations. We evaluate our method on multiple robotic control tasks in both sim-to-sim and sim-to-real transfer, demonstrating significant improvement over naive domain randomization. Project videos and code at https://yuqingd.github.io/autotuned-sim2real/