We study how to fine-tune LLMs using user-edit deployment data consisting of a set of context, an agent's response, and user edits. This deployment data is naturally generated by users in applications such as LLMs-based writing assistants and coding agents. The _natural_ origin of user edits makes it a desired source for adapting and personalizing of LLMs. In this setup, there emerges a unification of various feedback types namely preferences, supervised labels, and cost that are typically studied separately in the literature. In this paper, we initiate the theoretical investigation of learning from user edits. We first derive bounds for learning algorithms that learn from each of these feedback types. We prove that these algorithms have different trade-offs depending upon the user, data distribution, and model class. We then propose a simple ensembling procedure to jointly learn from these feedback types. On two domains from Gao et al. 2024, we show our ensembling procedure outperforms these methods that learn from individual feedback. Further, we show that our proposed procedure can robustly adapt to different user-edit distributions at test time.

We study pre-training representations for decision-making using video data, which is abundantly available for tasks such as game agents and software testing. Even though significant empirical advances have been made on this problem, a theoretical understanding remains absent. We initiate the theoretical investigation into principled approaches for representation learning and focus on learning the latent state representations of the underlying MDP using video data. We study two types of settings: one where there is iid noise in the observation, and a more challenging setting where there is also the presence of exogenous noise, which is non-iid noise that is temporally correlated, such as the motion of people or cars in the background. We study three commonly used approaches: autoencoding, temporal contrastive learning, and forward modeling. We prove upper bounds for temporal contrastive learning and forward modeling in the presence of only iid noise. We show that these approaches can learn the latent state and use it to do efficient downstream RL with polynomial sample complexity. When exogenous noise is also present, we establish a lower bound result showing that the sample complexity of learning from video data can be exponentially worse than learning from action-labeled trajectory data. This partially explains why reinforcement learning with video pre-training is hard. We evaluate these representational learning methods in two visual domains, yielding results that are consistent with our theoretical findings.

Transformer-based Large Language Models (LLMs) have become a fixture in modern machine learning. Correspondingly, significant resources are allocated towards research that aims to further advance this technology, typically resulting in models of increasing size that are trained on increasing amounts of data. This work, however, demonstrates the surprising result that it is often possible to significantly improve the performance of LLMs by selectively removing higher-order components of their weight matrices. This simple intervention, which we call LAyer-SElective Rank reduction (LASER), can be done on a model after training has completed, and requires minimal additional parameters and data. We show extensive experiments demonstrating the generality of this finding across language models and datasets, and provide in-depth analyses offering insights into both when LASER is effective and the mechanism by which it operates

We study interactive learning in a setting where the agent has to generate a response (e.g., an action or trajectory) given a context and an instruction. In contrast, to typical approaches that train the system using reward or expert supervision on response, we study _learning with hindsight labeling_ where a teacher provides an instruction that is most suitable for the agent's generated response. This hindsight labeling of instruction is often easier to provide than providing expert supervision of the optimal response which may require expert knowledge or can be impractical to elicit. We initiate the theoretical analysis of _interactive learning with hindsight labeling_. We first provide a lower bound showing that in general, the regret of any algorithm must scale with the size of the agent's response space. Next, we study a specialized setting where the underlying instruction-response distribution can be decomposed as a low-rank matrix. We introduce an algorithm called LORIL for this setting and show that it is a no-regret algorithm with the regret scaling with $\sqrt{T}$ and depends on the _intrinsic rank_ but does not depend on the agent's response space. We provide experiments showing the performance of LORIL in practice for 2 domains.

Imitation learning (IL) aims to mimic the behavior of an expert in a sequential decision making task by learning from demonstrations, and has been widely applied to robotics, autonomous driving, and autoregressive text generation. The simplest approach to IL, behavior cloning (BC) is thought to incur sample complexity with unfavorable quadratic dependence on the problem horizon, motivating a variety of different online algorithms that attain improved linear horizon dependence under stronger assumptions on the data and the learner’s access to the expert. We revisit the apparent gap between offline and online IL from a learning-theoretic perspective, with a focus on general policy classes up to and including deep neural networks. Through a new analysis of BC with the logarithmic loss, we show that it is possible to achieve horizon-independent sample complexity in offline IL whenever (i) the range of the cumulative payoffs is controlled, and (ii) an appropriate notion of supervised learning complexity for the policy class is controlled. Specializing our results to deterministic, stationary policies, we show that the gap between offline and online IL is not fundamental: (i) it is possible to achieve linear dependence on horizon in offline IL under dense rewards (matching what was previously only known to be achievable in online IL); and (ii) without further assumptions on the policy class, online IL cannot improve over offline IL with the logarithmic loss, even in benign MDPs. We complement our theoretical results with experiments on standard RL tasks and autoregressive language generation to validate the practical relevance of our findings.

We study interactive learning of language agents based on user edits made to the agent's output. In a typical setting such as writing assistants, the user interacts with a language agent to generate a response given a context, and may optionally edit the agent response to personalize it based on their latent preference, in addition to improving the correctness. The edit feedback is naturally generated, making it a suitable candidate for improving the agent's alignment with the user's preference, and for reducing the cost of user edits over time. We propose a learning framework, PRELUDE that infers a description of the user's latent preference based on historic edit data and using it to define a prompt policy that drives future response generation. This avoids fine-tuning the agent, which is costly, challenging to scale with the number of users, and may even degrade its performance on other tasks. Furthermore, learning descriptive preference improves interpretability, allowing the user to view and modify the learned preference. However, user preference can be complex and vary based on context, making it challenging to learn. To address this, we propose a simple yet effective algorithm named CIPHER that leverages a large language model (LLM) to infer the user preference for a given context based on user edits. In the future, CIPHER retrieves inferred preferences from the k-closest contexts in the history, and forms an aggregate preference for response generation. We introduce two interactive environments -- summarization and email writing, for evaluation using a GPT-4 simulated user. We compare with algorithms that directly retrieve user edits but do not learn descriptive preference, and algorithms that learn context-agnostic preference. On both tasks, CIPHER outperforms baselines by achieving the lowest edit distance cost. Meanwhile, CIPHER has a lower computational expense, as using learned preference results in a shorter prompt than directly using user edits. Our further analysis reports that the user preference learned by CIPHER shows significant similarity to the ground truth latent preference.

We introduce Language Feedback Models (LFMs) that identify desirable behaviour --- actions that help achieve tasks specified in the instruction - for imitation learning in instruction following. To train LFMs, we obtain feedback from Large Language Models (LLMs) on visual trajectories verbalized to language descriptions. First, by using LFMs to identify desirable behaviour to imitate, we improve in task-completion rate over strong behavioural cloning baselines on three distinct language grounding environments (Touchdown, ScienceWorld, and ALFWorld). Second, LFMs outperform using LLMs as experts to directly predict actions, when controlling for the number of LLM output tokens. Third, LFMs generalize to unseen environments, improving task-completion rate by 3.5-12.0% through one round of adaptation. Finally, LFMs can be modified to provide human-interpretable feedback without performance loss, allowing human verification of desirable behaviour for imitation learning.

Safety is a crucial necessity in many applications of reinforcement learning (RL), whether robotic, automotive, or medical. Many existing approaches to safe RL rely on receiving numeric safety feedback, but in many cases this feedback can only take binary values; that is, whether an action in a given state is safe or unsafe. This is particularly true when feedback comes from human experts. We therefore consider the problem of provable safe RL when given access to an offline oracle providing binary feedback on the safety of state, action pairs. We provide a novel meta algorithm, SABRE, which can be applied to any MDP setting given access to a blackbox PAC RL algorithm for that setting. SABRE applies concepts from active learning to reinforcement learning to provably control the number of queries to the safety oracle. SABRE works by iteratively exploring the state space to find regions where the agent is currently uncertain about safety. Our main theoretical results shows that, under appropriate technical assumptions, SABRE never takes unsafe actions during training, and is guaranteed to return a near-optimal safe policy with high probability. We provide a discussion of how our meta-algorithm may be applied to various settings studied in both theoretical and empirical frameworks.

Learning to control an agent from offline data collected in a rich pixel-based visual observation space is vital for real-world applications of reinforcement learning (RL). A major challenge in this setting is the presence of input information that is hard to model and irrelevant to controlling the agent. This problem has been approached by the theoretical RL community through the lens of *exogenous information*, i.e., any control-irrelevant information contained in observations. For example, a robot navigating in busy streets needs to ignore irrelevant information, such as other people walking in the background, textures of objects, or birds in the sky. In this paper, we focus on the setting with visually detailed exogenous information and introduce new offline RL benchmarks that offer the ability to study this problem. We find that contemporary representation learning techniques can fail on datasets where the noise is a complex and time-dependent process, which is prevalent in practical applications. To address these, we propose to use multi-step inverse models to learn Agent-Centric Representations for Offline-RL (ACRO). Despite being simple and reward-free, we show theoretically and empirically that the representation created by this objective greatly outperforms baselines.

We present a novel observation about the behavior of offline reinforcement learning (RL) algorithms: on many benchmark datasets, offline RL can produce well-performing and safe policies even when trained with "wrong" reward labels, such as those that are zero everywhere or are negatives of the true rewards. This phenomenon cannot be easily explained by offline RL's return maximization objective. Moreover, it gives offline RL a degree of robustness that is uncharacteristic of its online RL counterparts, which are known to be sensitive to reward design. We demonstrate that this surprising robustness property is attributable to an interplay between the notion of *pessimism* in offline RL algorithms and certain implicit biases in common data collection practices. As we prove in this work, pessimism endows the agent with a *survival instinct*, i.e., an incentive to stay within the data support in the long term, while the limited and biased data coverage further constrains the set of survival policies. Formally, given a reward class -- which may not even contain the true reward -- we identify conditions on the training data distribution that enable offline RL to learn a near-optimal and safe policy from any reward within the class. We argue that the survival instinct should be taken into account when interpreting results from existing offline RL benchmarks and when creating future ones. Our empirical and theoretical results suggest a new paradigm for offline RL, whereby an agent is "nudged" to learn a desirable behavior with imperfect reward but purposely biased data coverage. Please visit our website [https://survival-instinct.github.io](https://survival-instinct.github.io) for accompanied code and videos.

Noise contrastive learning is a popular technique for unsupervised representation learning. In this approach, a representation is obtained via reduction to supervised learning, where given a notion of semantic similarity, the learner tries to distinguish a similar (positive) example from a collection of random (negative) examples. The success of modern contrastive learning pipelines relies on many design decisions, such as the choice of data augmentation, the number of negative examples, and the batch size; however, there is limited understanding as to how these parameters interact and affect downstream performance. We focus on disambiguating the role of one of these parameters: the number of negative examples. Theoretically, we show the existence of a collision-coverage trade-off suggesting that the optimal number of negative examples should scale with the number of underlying concepts in the data. Empirically, we scrutinize the role of the number of negatives in both NLP and vision tasks.

In real-world reinforcement learning applications the learner’s observation space is ubiquitously high-dimensional with both relevant and irrelevant information about the task at hand. Learning from high-dimensional observations has been the subject of extensive investigation in supervised learning and statistics (e.g., via sparsity), but analogous issues in reinforcement learning are not well understood, even in finite state/action (tabular) domains. We introduce a new problem setting for reinforcement learning, the Exogenous Markov Decision Process (ExMDP), in which the state space admits an (unknown) factorization into a small controllable (or, endogenous) component and a large irrelevant (or, exogenous) component; the exogenous component is independent of the learner’s actions, but evolves in an arbitrary, temporally correlated fashion. We provide a new algorithm, OSSR, which learns a near-optimal policy with sample complexity polynomial in the size of the endogenous component and nearly independent of the size of the exogenous component, thereby offering a doubly-exponential improvement over off-the-shelf algorithms. Our results highlight for the first time that sample-efficient reinforcement learning is possible in the presence of exogenous information, and provide a simple, user-friendly benchmark for investigation going forward.

Contrastive learning is a popular form of self-supervised learning that encourages augmentations (views) of the same input to have more similar representations compared to augmentations of different inputs. Recent attempts to theoretically explain the success of contrastive learning on downstream classification tasks prove guarantees depending on properties of augmentations and the value of contrastive loss of representations. We demonstrate that such analyses, that ignore inductive biases of the function class and training algorithm, cannot adequately explain the success of contrastive learning, even provably leading to vacuous guarantees in some settings. Extensive experiments on image and text domains highlight the ubiquity of this problem – different function classes and algorithms behave very differently on downstream tasks, despite having the same augmentations and contrastive losses. Theoretical analysis is presented for the class of linear representations, where incorporating inductive biases of the function class allows contrastive learning to work with less stringent conditions compared to prior analyses.

We study reinforcement learning (RL) in settings where observations are high-dimensional, but where an RL agent has access to abstract knowledge about the structure of the state space, as is the case, for example, when a robot is tasked to go to a specific room in a building using observations from its own camera, while having access to the floor plan. We formalize this setting as transfer reinforcement learning from an "abstract simulator," which we assume is deterministic (such as a simple model of moving around the floor plan), but which is only required to capture the target domain's latent-state dynamics approximately up to unknown (bounded) perturbations (to account for environment stochasticity). Crucially, we assume no prior knowledge about the structure of observations in the target domain except that they can be used to identify the latent states (but the decoding map is unknown). Under these assumptions, we present an algorithm, called TASID, that learns a robust policy in the target domain, with sample complexity that is polynomial in the horizon, and independent of the number of states, which is not possible without access to some prior knowledge. In synthetic experiments, we verify various properties of our algorithm and show that it empirically outperforms transfer RL algorithms that require access to "full simulators" (i.e., those that also simulate observations).

We present a novel interactive learning protocol that enables training request-fulfilling agents by verbally describing their activities. Unlike imitation learning (IL), our protocol allows the teaching agent to provide feedback in a language that is most appropriate for them. Compared with reward in reinforcement learning (RL), the description feedback is richer and allows for improved sample complexity. We develop a probabilistic framework and an algorithm that practically implements our protocol. Empirical results in two challenging request-fulfilling problems demonstrate the strengths of our approach: compared with RL baselines, it is more sample-efficient; compared with IL baselines, it achieves competitive success rates without requiring the teaching agent to be able to demonstrate the desired behavior using the learning agent’s actions. Apart from empirical evaluation, we also provide theoretical guarantees for our algorithm under certain assumptions about the teacher and the environment.

We present an algorithm, HOMER, for exploration and reinforcement learning in rich observation environments that are summarizable by an unknown latent state space. The algorithm interleaves representation learning to identify a new notion of kinematic state abstraction with strategic exploration to reach new states using the learned abstraction. The algorithm provably explores the environment with sample complexity scaling polynomially in the number of latent states and the time horizon, and, crucially, with no dependence on the size of the observation space, which could be infinitely large. This exploration guarantee further enables sample-efficient global policy optimization for any reward function. On the computational side, we show that the algorithm can be implemented efficiently whenever certain supervised learning problems are tractable. Empirically, we evaluate HOMER on a challenging exploration problem, where we show that the algorithm is more sample efficient than standard reinforcement learning baselines.

We introduce a new problem setting for continuous control called the LQR with Rich Observations, or RichLQR. In our setting, the environment is summarized by a low-dimensional continuous latent state with linear dynamics and quadratic costs, but the agent operates on high-dimensional, nonlinear observations such as images from a camera. To enable sample-efficient learning, we assume that the learner has access to a class of decoder functions (e.g., neural networks) that is flexible enough to capture the mapping from observations to latent states. We introduce a new algorithm, RichID, which learns a near-optimal policy for the RichLQR with sample complexity scaling only with the dimension of the latent state space and the capacity of the decoder function class. RichID is oracle-efficient and accesses the decoder class only through calls to a least-squares regression oracle. To our knowledge, our results constitute the first provable sample complexity guarantee for continuous control with an unknown nonlinearity in the system model.

We study the problem of jointly reasoning about language and vision through a navigation and spatial reasoning task. We introduce the Touchdown task and dataset, where an agent must first follow navigation instructions in a Street View environment to a goal position, and then guess a location in its observed environment described in natural language to find a hidden object. The data contains 9326 examples of English instructions and spatial descriptions paired with demonstrations. We perform qualitative linguistic analysis, and show that the data displays a rich use of spatial reasoning. Empirical analysis shows the data presents an open challenge to existing methods.

Building perceptual systems for robotics which perform well under tight computational budgets requires novel architectures which rethink the traditional computer vision pipeline. Modern vision architectures require the agent to build a summary representation of the entire scene, even if most of the input is irrelevant to the agent's current goal. In this work, we flip this paradigm, by introducing EARLYFUSION vision models that condition on a goal to build custom representations for downstream tasks. We show that these goal specific representations can be learned more quickly, are substantially more parameter efficient, and more robust than existing attention mechanisms in our domain. We demonstrate the effectiveness of these methods on a simulated item retrieval problem that is trained in a fully end-to-end manner via imitation learning.

We propose to decompose instruction execution to goal prediction and action generation.We design a model that maps raw visual observations to goals using LINGUNET, a language-conditioned image generation network, and then generates the actions required to complete them.Our model is trained from demonstration only without external resources.To evaluate our approach, we introduce two benchmarks for instruction following: LANI, a navigation task; and CHAI, where an agent executes household instructions.Our evaluation demonstrates the advantages of our model decomposition, and illustrates the challenges posed by our new benchmarks.

Semantic parsing from denotations faces two key challenges in model training: (1) given only the denotations (e.g., answers), search for good candidate semantic parses, and (2) choose the best model update algorithm.We propose effective and general solutions to each of them.Using policy shaping, we bias the search procedure towards semantic parses that are more compatible to the text, which provide better supervision signals for training.In addition, we propose an update equation that generalizes three different families of learning algorithms, which enables fast model exploration.When experimented on a recently proposed sequential question answering dataset, our framework leads to a new state-of-theart model that outperforms previous work by 5.0% absolute on exact match accuracy.Question: what nation scored the most points

We examine the impact of learning Lipschitz continuous models in the context of model-based reinforcement learning. We provide a novel bound on multi-step prediction error of Lipschitz models where we quantify the error using the Wasserstein metric. We go on to prove an error bound for the value-function estimate arising from Lipschitz models and show that the estimated value function is itself Lipschitz. We conclude with empirical results that show the benefits of controlling the Lipschitz constant of neural-network models.

We propose to directly map raw visual observations and text input to actions for instruction execution.While existing approaches assume access to structured environment representations or use a pipeline of separately trained models, we learn a single model to jointly reason about linguistic and visual input.We use reinforcement learning in a contextual bandit setting to train a neural network agent.To guide the agent's exploration, we use reward shaping with different forms of supervision.Our approach does not require intermediate representations, planning procedures, or training different models.We evaluate in a simulated environment, and show significant improvements over supervised learning and common reinforcement learning variants.

Dipendra Kumar Misra, Kejia Tao, Percy Liang, Ashutosh Saxena. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.