Despite being one of the earliest neural network layers, the Multilayer Perceptron (MLP) is arguably one of the least understood parts of the transformer architecture due to its dense computation and lack of easy visualization. This paper seeks to understand the MLP layers in dense LLM models by hypothesizing that these layers secretly approximately perform a sparse computation -- namely, that they can be well approximated by sparsely-activating Mixture of Experts (MoE) layers. Our hypothesis is based on a novel theoretical connection between MoE models and Sparse Autoencoder (SAE) structure in activation space. We empirically validate the hypothesis on pretrained LLMs, and demonstrate that the activation distribution matters -- these results do not hold for Gaussian data, but rather rely crucially on structure in the distribution of neural network activations. Our results shine light on a general principle at play in MLP layers inside LLMs, and give an explanation for the effectiveness of modern MoE-based transformers. Additionally, our experimental explorations suggest new directions for more efficient MoE architecture design based on low-rank routers.

In Transformer language models, activation vectors transform from current token embeddings to next token predictions as they pass through the model. To isolate a minimal form of this transformation, we identify language model subnetworks that make bigram predictions, naive next token predictions based only on the current token. We find that bigram subnetworks can be found in fully trained language models up to 1B parameters, and these subnetworks are critical for model performance even when they consist of less than 0.2% of model parameters. Bigram subnetworks are concentrated in the first Transformer MLP layer, and they overlap significantly with subnetworks trained to optimally prune a given model. Mechanistically, the bigram subnetworks often recreate a pattern from the full models where the first layer induces a sharp change that aligns activations with next token predictions rather than current token representations. Our results demonstrate that bigram subnetworks comprise a minimal subset of parameters that are both necessary and sufficient for basic next token predictions in language models, and they help drive the transformation from current to next token activations in the residual stream. These subnetworks can lay a foundation for studying more complex language model circuits by building up from a minimal circuit.

Large language models (LLMs) have significantly advanced generative applications in natural language processing (NLP). Recent trends in model architectures revolve around efficient variants of transformers or state-space/gated-recurrent models (SSMs, GRMs). However, prevailing SSM/GRM-based methods often emulate only a single attention head, potentially limiting their expressiveness. In this work, we propose MossNet, a novel mixture-of-state-space-experts architecture that emulates a linear multi-head attention (MHA). MossNet leverages a mixture-of-experts (MoE) implementation not only in channel-mixing multi-layered perceptron (MLP) blocks but also in the time-mixing SSM kernels to realize multiple "attention heads." Extensive experiments on language modeling and downstream evaluations show that MossNet outperforms both transformer- and SSM-based architectures of similar model size and data budgets. Larger variants of MossNet, trained on trillions of tokens, further confirm its scalability and superior performance. In addition, real-device profiling on a Samsung Galaxy S24 Ultra and an Nvidia A100 GPU demonstrate favorable runtime speed and resource usage compared to similarly sized baselines. Our results suggest that MossNet is a compelling new direction for efficient, high-performing recurrent LLM architectures.

This article will be presented at the IEEE MTT -S International Microwave Symposium (IMS 2025), San Francisco, CA, USA, USA 15-20, 2025 and accepted to the IEEE Microwave and Wireless Technology Letters (MWTL). Abstract --Digital Predistortion (DPD) is a popular technique to enhance signal quality in wideband RF power amplifiers (PAs). With increasing bandwidth and data rates, DPD faces significant energy consumption challenges during deployment, contrasting with its efficiency goals. State-of-the-art DPD models rely on recurrent neural networks (RNN), whose computational complexity hinders system efficiency. This paper introduces DeltaDPD, exploring the dynamic temporal sparsity of input signals and neuronal hidden states in RNNs for energy-efficient DPD, reducing arithmetic operations and memory accesses while preserving satisfactory linearization performance. Applying a TM3.1a 200MHz-BW 256-QAM OFDM signal to a 3.5 GHz GaN Doherty RF PA, DeltaDPD achieves -50.03 dBc in Adjacent Channel Power Ratio (ACPR), -37.22 dB in Normalized Mean Square Error (NMSE) and -38.52 dBc in Error V ector Magnitude (EVM) with 52% temporal sparsity, leading to a 1.8 reduction in estimated inference power .

--Over the past few years, the size of language models has grown exponentially, as has the computational cost to train these large models. This rapid growth has motivated researchers to develop new techniques aimed at enhancing the efficiency of the training process. Despite these advancements, optimally predicting the model size or allocating optimal resources remains a challenge. Several efforts have addressed the challenge by proposing different scaling laws, but almost all of them are architecture-specific (dense or sparse). In this work we revisit existing scaling laws and propose a generalized scaling law to provide a unified framework that is applicable to both dense and sparse large language models. We evaluate and compare our proposed scaling law with existing scaling laws to demonstrate its effectiveness. In recent years, transformer architectures [1] have revolutionized the deep learning approach. These architectures are now the foundation for the majority of popular large language models (LLMs).

We introduce and train distributed neural architectures (DNA) in vision and language domains. DNAs are initialized with a proto-architecture that consists of (transformer, MLP, attention, etc.) modules and routers. Any token (or patch) can traverse any series of modules in any order. DNAs are a natural generalization of the sparse methods such as Mixture-of-Experts, Mixture-of-Depths, parameter sharing, etc. Computation and communication patterns of DNA modules are learnt end-to-end during training and depend on the content and context of each token (or patch). These patterns can be shaped by further requirements added to the optimization objective such as compute/memory efficiency or load balancing. We empirically show that (i) trained DNAs are competitive with the dense baselines in both domains and (ii) compute efficiency/parameter sharing can be learnt from data. Next, we analyze the emergent connectivity and computation patterns in the trained DNAs. We find that the paths that tokens take through the models are themselves distributed according to a power-law. We show that some paths (or, equivalently, groups of modules) show emergent specialization. Finally, we demonstrate that models learn to allocate compute and active parameters in an interpretable way.

Mixture-of-Experts (MoE) layers are increasingly central to frontier model architectures. By selectively activating parameters, they reduce computational cost while scaling total parameter count. This paper investigates the impact of the number of active experts, termed granularity, comparing architectures with many (e.g., 8 per layer in DeepSeek) to those with fewer (e.g., 1 per layer in Llama-4 models). We prove an exponential separation in network expressivity based on this design parameter, suggesting that models benefit from higher granularity. Experimental results corroborate our theoretical findings and illustrate this separation.

Mixture-of-Experts (MoE) Large Language Models (LLMs) suffer from severely sub-optimal expert pathways-our study reveals that naive expert selection learned from pretraining leaves a surprising 10-20% accuracy gap for improvement. Motivated by this observation, we develop a novel class of test-time optimization methods to re-weight or "re-mixing" the experts in different layers jointly for each test sample. Since the test sample's ground truth is unknown, we propose to optimize a surrogate objective defined by the sample's "successful neighbors" from a reference set of samples. We introduce three surrogates and algorithms based on mode-finding, kernel regression, and the average loss of similar reference samples/tasks. To reduce the cost of optimizing whole pathways, we apply our algorithms merely to the core experts' mixing weights in critical layers, which enjoy similar performance but save significant computation. This leads to "Critical-Layer, Core-Expert, Collaborative Pathway Optimization (C3PO)". We apply C3PO to two recent MoE LLMs and examine it on six widely-used benchmarks. It consistently improves the base model by 7-15% in accuracy and outperforms widely used test-time learning baselines, e.g., in-context learning and prompt/prefix tuning, by a large margin. Moreover, C3PO enables MoE LLMs with 1-3B active parameters to outperform LLMs of 7-9B parameters, hence improving MoE's advantages on efficiency. Our thorough ablation study further sheds novel insights on achieving test-time improvement on MoE.

Meta has released the first two models from its multimodal Llama 4 suite: LLama 4 Scout and Llama 4 Maverick. Maverick is "the workhorse" of the two and excels at image and text understanding for "general assistant and chat use cases," the company said in a blog post, while the smaller model Scout could tackle things like "multi-document summarization, parsing extensive user activity for personalized tasks, and reasoning over vast codebases." The company also introduced Llama 4 Behemoth, an upcoming model it says is "among the world's smartest LLMs" -- and CEO Mark Zuckerberg said we'll be hearing about a fourth model, LLama 4 Reasoning, "in the next month." Both Maverick and Scout are available to download now from the LLama website and Hugging Face, and they've been added to Meta AI, including for WhatsApp, Messenger and Instagram DMs. Scout has 17 billion active parameters with 16 experts, Meta says.

Sparse Mixture of Expert (MoE) models are popular foundational architectures at large scale, however, under-explored at smaller sizes. Here, we show how to enable Compact Sparse Mixture of Experts (CoSMoEs) for on-device inference. Specifically, we tackle the three main on-device dimensions: Quality, Memory and Latency. Along the quality axis, we show that in a fair evaluation (removing confounding factors) MoE architectures outperform FLOP-aligned dense models at on-device scale. We introduce weight-decomposed experts, further improving the MoE model performance. Regarding model memory and latency, we significantly improve model offloading efficiency and, in turn, reduce model inference latency.