We propose a new finetuning method to provide pre-trained large language models (LMs) the ability to scale test-time compute through the diffusion framework. By increasing the number of diffusion steps, we show our finetuned models achieve monotonically increasing accuracy, directly translating to improved performance across downstream tasks. Furthermore, our finetuned models can expertly answer questions on specific topics by integrating powerful guidance techniques, and autonomously determine the compute required for a given problem by leveraging adaptive ODE solvers. Our method is universally applicable to any foundation model pre-trained with a cross-entropy loss and does not modify any of its original weights, fully preserving its strong single-step generation capabilities. We show our method is more effective and fully compatible with traditional finetuning approaches, introducing an orthogonal new direction to unify the strengths of the autoregressive and diffusion frameworks.

This dynamic adjustment has parallels to concepts like fast-weight memories, which enable networks to update weights in response to task demands (Schmid-huber, 1992; Gomez & Schmidhuber, 2005), and neural network weights being treated as dynamic programs (Schmidhuber, 2015). Recently, Panigrahi et al. (2023) introduces an approach where a smaller auxiliary transformer is updated dynamically within a larger model, aligning with the principles of self-adaptive behavior. This adaptation can be explored from two perspectives: a macroview, where multiple LLMs collaborate and/or compete, and a microview, where internal adaptations allow a single LLM to specialize in different tasks. Macroview: From this perspective, the system directs queries to LLMs with domain specific expertise, prioritizing outputs from expert models, thereby achieving higher accuracy and task-specific optimization. Such task-specific ensembles can be realized through various mechanisms: multiple LLMs playing distinct roles and coordinate toward a shared goal (Zhuge et al., 2023), engaging in mutual listening and debate (Du et al., 2023), or using meticulously crafted prompt constructions (Zhang et al., 2024) to integrate knowledge library and skill planning.

Prior methods propose to offset the escalating costs of modern foundation models by dropping specific parts of their contexts with hand-designed rules, while attempting to preserve their original performance. We overcome this trade-off with Neural Attention Memory Models (NAMMs), introducing a learned network for memory management that improves both the performance and efficiency of transformers. We evolve NAMMs atop pre-trained transformers to provide different latent contexts focusing on the most relevant information for individual layers and attention heads. NAMMs are universally applicable to any model using self-attention as they condition exclusively on the values in the produced attention matrices. Learning NAMMs on a small set of problems, we achieve substantial performance improvements across multiple long-context benchmarks while cutting the model's input contexts up to a fraction of the original sizes. We show the generality of our conditioning enables zero-shot transfer of NAMMs trained only on language to entirely new transformer architectures even across input modalities, with their benefits carrying over to vision and reinforcement learning.

The forward-backward representation (FB) is a recently proposed framework (Touati et al., 2023; Touati & Ollivier, 2021) to train behavior foundation models (BFMs) that aim at providing zero-shot efficient policies for any new task specified in a given reinforcement learning (RL) environment, without training for each new task. Here we address two core limitations of FB model training. First, FB, like all successor-feature-based methods, relies on a linear encoding of tasks: at test time, each new reward function is linearly projected onto a fixed set of pre-trained features. This limits expressivity as well as precision of the task representation. We break the linearity limitation by introducing auto-regressive features for FB, which let finegrained task features depend on coarser-grained task information. This can represent arbitrary nonlinear task encodings, thus significantly increasing expressivity of the FB framework. Second, it is well-known that training RL agents from offline datasets often requires specific techniques.We show that FB works well together with such offline RL techniques, by adapting techniques from (Nair et al.,2020b; Cetin et al., 2024) for FB. This is necessary to get non-flatlining performance in some datasets, such as DMC Humanoid. As a result, we produce efficient FB BFMs for a number of new environments. Notably, in the D4RL locomotion benchmark, the generic FB agent matches the performance of standard single-task offline agents (IQL, XQL). In many setups, the offline techniques are needed to get any decent performance at all. The auto-regressive features have a positive but moderate impact, concentrated on tasks requiring spatial precision and task generalization beyond the behaviors represented in the trainset.

For instance, TD3+Behavior Cloning (TD3+BC, Fujimoto and Gu, 2021) achieves this by regularizing the actor loss with the divergence between the learned policy and the data-generating policy, while Advantage Weighted Actor Critic (AWAC, Nair et al., 2020) seeks a policy maximizing the data likelihood weighted by its exponentiated advantage function. Later extensions of AWAC also modify the critic loss to avoid querying actions outside the given data by learning a value function, e.g., by expectile regression in Implicit Q-learning (IQL, Kostrikov et al., 2022) and Gumbel regression in Extreme Q-learning (XQL, Garg et al., 2023). This class of methods can be easily integrated with online fine-tuning, even leading to several successful applications for real-world tasks (Lu et al., 2022; Nair et al., 2023). However, current offline RL methods still fail in simple settings. Hong et al. (2023c,b) showed that if the data contains many low-return and few high-return trajectories, policy constrained methods are unnecessarily conservative and fail to learn good behavior. Singh et al. (2023) report a similar effect on heteroskedastic datasets where the variability of behaviors differs across different regions of the state space.

We propose a new class of deep reinforcement learning (RL) algorithms that model latent representations in hyperbolic space. Sequential decision-making requires reasoning about the possible future consequences of current behavior. Consequently, capturing the relationship between key evolving features for a given task is conducive to recovering effective policies. To this end, hyperbolic geometry provides deep RL models with a natural basis to precisely encode this inherently hierarchical information. However, applying existing methodologies from the hyperbolic deep learning literature leads to fatal optimization instabilities due to the non-stationarity and variance characterizing RL gradient estimators. Hence, we design a new general method that counteracts such optimization challenges and enables stable end-to-end learning with deep hyperbolic representations. We empirically validate our framework by applying it to popular on-policy and off-policy RL algorithms on the Procgen and Atari 100K benchmarks, attaining near universal performance and generalization benefits. Given its natural fit, we hope future RL research will consider hyperbolic representations as a standard tool.

Popular off-policy deep reinforcement learning algorithms compensate for overestimation bias during temporal-difference learning by utilizing pessimistic estimates of the expected target returns. In this work, we propose a novel learnable penalty to enact such pessimism, based on a new way to quantify the critic's epistemic uncertainty. Furthermore, we propose to learn the penalty alongside the critic with dual TD-learning, a strategy to estimate and minimize the bias magnitude in the target returns. Our method enables us to accurately counteract overestimation bias throughout training without incurring the downsides of overly pessimistic targets. Empirically, by integrating our method and other orthogonal improvements with popular off-policy algorithms, we achieve state-of-the-art results in continuous control tasks from both proprioceptive and pixel observations.

Human beings are able to understand objectives and learn by simply observing others perform a task. Imitation learning methods aim to replicate such capabilities, however, they generally depend on access to a full set of optimal states and actions taken with the agent's actuators and from the agent's point of view. In this paper, we introduce a new algorithm - called Disentangling Generative Adversarial Imitation Learning (DisentanGAIL) - with the purpose of bypassing such constraints. Our algorithm enables autonomous agents to learn directly from high dimensional observations of an expert performing a task, by making use of adversarial learning with a latent representation inside the discriminator network. Such latent representation is regularized through mutual information constraints to incentivize learning only features that encode information about the completion levels of the task being demonstrated. This allows to obtain a shared feature space to successfully perform imitation while disregarding the differences between the expert's and the agent's domains. Empirically, our algorithm is able to efficiently imitate in a diverse range of control problems including balancing, manipulation and locomotive tasks, while being robust to various domain differences in terms of both environment appearance and agent embodiment.