State Space Models (SSMs) have emerged as efficient alternatives to Transformers for sequential modeling, but their inability to leverage modality-specific features limits their performance in multi-modal pretraining. Here, we propose Mixture-of-Mamba, a novel SSM architecture that introduces modality-aware sparsity through modality-specific parameterization of the Mamba block. Building on Mixture-of-Transformers (W. Liang et al. arXiv:2411.04996; 2024), we extend the benefits of modality-aware sparsity to SSMs while preserving their computational efficiency. We evaluate Mixture-of-Mamba across three multi-modal pretraining settings: Transfusion (interleaved text and continuous image tokens with diffusion loss), Chameleon (interleaved text and discrete image tokens), and an extended three-modality framework incorporating speech. Mixture-of-Mamba consistently reaches the same loss values at earlier training steps with significantly reduced computational costs. In the Transfusion setting, Mixture-of-Mamba achieves equivalent image loss using only 34.76% of the training FLOPs at the 1.4B scale. In the Chameleon setting, Mixture-of-Mamba reaches similar image loss with just 42.50% of the FLOPs at the 1.4B scale, and similar text loss with just 65.40% of the FLOPs. In the three-modality setting, MoM matches speech loss at 24.80% of the FLOPs at the 1.4B scale. Our ablation study highlights the synergistic effects of decoupling projection components, where joint decoupling yields greater gains than individual modifications. These results establish modality-aware sparsity as a versatile and effective design principle, extending its impact from Transformers to SSMs and setting new benchmarks in multi-modal pretraining. Our code can be accessed at https://github.com/Weixin-Liang/Mixture-of-Mamba

The development of large language models (LLMs) has expanded to multi-modal systems capable of processing text, images, and speech within a unified framework. Training these models demands significantly larger datasets and computational resources compared to text-only LLMs. To address the scaling challenges, we introduce Mixture-of-Transformers (MoT), a sparse multi-modal transformer architecture that significantly reduces pretraining computational costs. MoT decouples non-embedding parameters of the model by modality -- including feed-forward networks, attention matrices, and layer normalization -- enabling modality-specific processing with global self-attention over the full input sequence. We evaluate MoT across multiple settings and model scales. In the Chameleon 7B setting (autoregressive text-and-image generation), MoT matches the dense baseline's performance using only 55.8\% of the FLOPs. When extended to include speech, MoT reaches speech performance comparable to the dense baseline with only 37.2\% of the FLOPs. In the Transfusion setting, where text and image are trained with different objectives, a 7B MoT model matches the image modality performance of the dense baseline with one third of the FLOPs, and a 760M MoT model outperforms a 1.4B dense baseline across key image generation metrics. System profiling further highlights MoT's practical benefits, achieving dense baseline image quality in 47.2\% of the wall-clock time and text quality in 75.6\% of the wall-clock time (measured on AWS p4de.24xlarge instances with NVIDIA A100 GPUs).

This work compares ways of extending Reinforcement Learning algorithms to Partially Observed Markov Decision Processes (POMDPs) with options. One view of options is as temporally extended action, which can be realized as a memory that allows the agent to retain historical information beyond the policy's context window. While option assignment could be handled using heuristics and hand-crafted objectives, learning temporally consistent options and associated sub-policies without explicit supervision is a challenge. Two algorithms, PPOEM and SOAP, are proposed and studied in depth to address this problem. PPOEM applies the forward-backward algorithm (for Hidden Markov Models) to optimize the expected returns for an option-augmented policy. However, this learning approach is unstable during on-policy rollouts. It is also unsuited for learning causal policies without the knowledge of future trajectories, since option assignments are optimized for offline sequences where the entire episode is available. As an alternative approach, SOAP evaluates the policy gradient for an optimal option assignment. It extends the concept of the generalized advantage estimation (GAE) to propagate option advantages through time, which is an analytical equivalent to performing temporal back-propagation of option policy gradients. This option policy is only conditional on the history of the agent, not future actions. Evaluated against competing baselines, SOAP exhibited the most robust performance, correctly discovering options for POMDP corridor environments, as well as on standard benchmarks including Atari and MuJoCo, outperforming PPOEM, as well as LSTM and Option-Critic baselines. The open-sourced code is available at https://github.com/shuishida/SoapRL.

Learned optimizers (LOs) can significantly reduce the wall-clock training time of neural networks, substantially reducing training costs. However, they often suffer from poor meta-generalization, especially when training networks larger than those seen during meta-training. To address this, we use the recently proposed Maximal Update Parametrization ($\mu$P), which allows zero-shot generalization of optimizer hyperparameters from smaller to larger models. We extend $\mu$P theory to learned optimizers, treating the meta-training problem as finding the learned optimizer under $\mu$P. Our evaluation shows that LOs meta-trained with $\mu$P substantially improve meta-generalization as compared to LOs trained under standard parametrization (SP). Notably, when applied to large-width models, our best $\mu$LO, trained for 103 GPU-hours, matches or exceeds the performance of VeLO, the largest publicly available learned optimizer, meta-trained with 4000 TPU-months of compute. Moreover, $\mu$LOs demonstrate better generalization than their SP counterparts to deeper networks and to much longer training horizons (25 times longer) than those seen during meta-training.

Understanding the training dynamics of deep neural networks is challenging due to their high-dimensional nature and intricate loss landscapes. Recent studies have revealed that, along the training trajectory, the gradient approximately aligns with a low-rank top eigenspace of the training loss Hessian, referred to as the dominant subspace. Given this alignment, this paper explores whether neural networks can be trained within the dominant subspace, which, if feasible, could lead to more efficient training methods. Our primary observation is that when the SGD update is projected onto the dominant subspace, the training loss does not decrease further. This suggests that the observed alignment between the gradient and the dominant subspace is spurious. Surprisingly, projecting out the dominant subspace proves to be just as effective as the original update, despite removing the majority of the original update component. Similar observations are made for the large learning rate regime (also known as Edge of Stability) and Sharpness-Aware Minimization. We discuss the main causes and implications of this spurious alignment, shedding light on the intricate dynamics of neural network training.

Many applications of reinforcement learning can be formalized as goal-conditioned environments, where, in each episode, there is a "goal" that affects the rewards obtained during that episode but does not affect the dynamics. Various techniques have been proposed to improve performance in goal-conditioned environments, such as automatic curriculum generation and goal relabeling. In this work, we explore a connection between off-policy reinforcement learning in goal-conditioned settings and knowledge distillation. In particular: the current Q-value function and the target Q-value estimate are both functions of the goal, and we would like to train the Q-value function to match its target for all goals. We therefore apply Gradient-Based Attention Transfer (Zagoruyko and Komodakis 2017), a knowledge distillation technique, to the Q-function update. We empirically show that this can improve the performance of goal-conditioned off-policy reinforcement learning when the space of goals is high-dimensional. We also show that this technique can be adapted to allow for efficient learning in the case of multiple simultaneous sparse goals, where the agent can attain a reward by achieving any one of a large set of objectives, all specified at test time. Finally, to provide theoretical support, we give examples of classes of environments where (under some assumptions) standard off-policy algorithms such as DDPG require at least O(d^2) replay buffer transitions to learn an optimal policy, while our proposed technique requires only O(d) transitions, where d is the dimensionality of the goal and state space. Code is available at https://github.com/alevine0/ReenGAGE.