Abstract:Transformers face quadratic complexity and memory issues with long sequences, prompting the adoption of linear attention mechanisms using fixed-size hidden states. However, linear models often suffer from limited recall performance, leading to hybrid architectures that combine linear and full attention layers. Despite extensive hybrid architecture research, the choice of linear attention component has not been deeply explored. We systematically evaluate various linear attention models across generations - vector recurrences to advanced gating mechanisms - both standalone and hybridized. To enable this comprehensive analysis, we trained and open-sourced 72 models: 36 at 340M parameters (20B tokens) and 36 at 1.3B parameters (100B tokens), covering six linear attention variants across five hybridization ratios. Benchmarking on standard language modeling and recall tasks reveals that superior standalone linear models do not necessarily excel in hybrids. While language modeling remains stable across linear-to-full attention ratios, recall significantly improves with increased full attention layers, particularly below a 3:1 ratio. Our study highlights selective gating, hierarchical recurrence, and controlled forgetting as critical for effective hybrid models. We recommend architectures such as HGRN-2 or GatedDeltaNet with a linear-to-full ratio between 3:1 and 6:1 to achieve Transformer-level recall efficiently. Our models are open-sourced at https://huggingface.co/collections/m-a-p/hybrid-linear-attention-research-686c488a63d609d2f20e2b1e.
Abstract:Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, especially when guided by explicit chain-of-thought (CoT) reasoning that verbalizes intermediate steps. While CoT improves both interpretability and accuracy, its dependence on natural language reasoning limits the model's expressive bandwidth. Latent reasoning tackles this bottleneck by performing multi-step inference entirely in the model's continuous hidden state, eliminating token-level supervision. To advance latent reasoning research, this survey provides a comprehensive overview of the emerging field of latent reasoning. We begin by examining the foundational role of neural network layers as the computational substrate for reasoning, highlighting how hierarchical representations support complex transformations. Next, we explore diverse latent reasoning methodologies, including activation-based recurrence, hidden state propagation, and fine-tuning strategies that compress or internalize explicit reasoning traces. Finally, we discuss advanced paradigms such as infinite-depth latent reasoning via masked diffusion models, which enable globally consistent and reversible reasoning processes. By unifying these perspectives, we aim to clarify the conceptual landscape of latent reasoning and chart future directions for research at the frontier of LLM cognition. An associated GitHub repository collecting the latest papers and repos is available at: https://github.com/multimodal-art-projection/LatentCoT-Horizon/.
Abstract:Neuromorphic engineering aims to advance computing by mimicking the brain's efficient processing, where data is encoded as asynchronous temporal events. This eliminates the need for a synchronisation clock and minimises power consumption when no data is present. However, many benchmarks for neuromorphic algorithms primarily focus on spatial features, neglecting the temporal dynamics that are inherent to most sequence-based tasks. This gap may lead to evaluations that fail to fully capture the unique strengths and characteristics of neuromorphic systems. In this paper, we present NeuroMorse, a temporally structured dataset designed for benchmarking neuromorphic learning systems. NeuroMorse converts the top 50 words in the English language into temporal Morse code spike sequences. Despite using only two input spike channels for Morse dots and dashes, complex information is encoded through temporal patterns in the data. The proposed benchmark contains feature hierarchy at multiple temporal scales that test the capacity of neuromorphic algorithms to decompose input patterns into spatial and temporal hierarchies. We demonstrate that our training set is challenging to categorise using a linear classifier and that identifying keywords in the test set is difficult using conventional methods. The NeuroMorse dataset is available at Zenodo, with our accompanying code on GitHub at https://github.com/Ben-E-Walters/NeuroMorse.