The Dynamical Memory Transformer (DMT) is a hybrid architecture that decouples sequence encoding from information retrieval. It employs multiple memory units—each a dynamical system with a complex diagonal recurrent matrix—updated via an input‑dependent leak rate, while the decoder uses attention to read these units without accessing the full token history. This design enables an associative scan with linear total work and logarithmic parallel depth, keeping recurrent states and cache sizes constant during autoregressive inference. DMT outperforms Transformers on several cognitive tasks and matches their language‑modeling performance, though it lags behind Mamba in language modeling; it also shows superior length extrapolation on selective copy tasks.
arXiv:2608. 02032v1 Announce Type: new Abstract: Modern language models are built primarily from Transformers, recurrent models, and their hybrid architectures.
Modern language models are built primarily from Transformers, recurrent models, and their hybrid architectures. Transformers rely on token-level attention memories, while recurrent models such as state space models (SSMs) and linear attention maintain compact recurrent states.
The paper proposes a lightweight recurrent memory module inserted between the lower and upper halves of a 6‑layer decoder‑only transformer. This module, which uses cross‑attention to observe hidden states, a GRU to update a persistent state, and gated addition to modulate subsequent layers, adds only 3.7% more parameters. It reduces evaluation loss by 28.5% and narrows the generalization gap, with ablations showing the benefit comes solely from the memory topology rather than auxiliary losses.
arXiv:2503. 18970v4 Announce Type: replace Abstract: Structured State Space Models (SSMs) have become a prominent class of sequence models, developed against two long-standing difficulties: the sequential computation and gradient propagation limits of Recurrent Neural Networks (RNNs), and the quadratic time and memory cost of self-attention in Transformers.
arXiv:2605.26797v2 Announce Type: replace Abstract: We study Latent Recurrent Transformer (LRT), a lightweight augmentation of autoregressive transformers that reuses a high-level source-layer hidden...
arXiv:2606. 10650v1 Announce Type: cross Abstract: The scalability of Large Language Models (LLMs) to long contexts is fundamentally constrained by the quadratic complexity of standard attention, motivating the adoption of linear attention mechanisms with sub-quadratic cost.
The paper compares three recurrent-depth transformer architectures—Block Universal Transformer (BUT), CoTFormer, and a new LSTM Universal Transformer (LSTM‑UT)—on cellular automata and delayed‑recall tasks. BUT consistently extrapolates to unseen depths better than CoTFormer, though its accuracy eventually falls. LSTM‑UT, which uses bounded gated memory, outperforms both baselines in depth extrapolation and delayed recall, demonstrating the value of gated memory for repeated computation and retrieval.
arXiv:2610.15715v1 Announce Type: new Abstract: Linear attention promises the efficiency needed for long-context language modelling, yet its fixed-size recurrent state must simultaneously support ong...
arXiv:2609.39661v1 Announce Type: new Abstract: Self-attention gives LLMs fine-grained, query-dependent access to context, but dense token interactions incur quadratic prefill cost and a key--value c...
The scalability of Large Language Models (LLMs) to long contexts is fundamentally constrained by the quadratic complexity of standard attention, motivating the adoption of linear attention mechanisms with sub-quadratic cost. To improve representation capacity under long contexts, recent approaches organize memory in a multi-state manner.
The paper investigates whether extra computation in recurrent Transformers should be allocated to more temporal steps or greater physical depth. Using Latent Recurrent Transformers (LRTs), the authors insert a latent thought token between vocabulary tokens, allowing each token to pass through the same $L$ layers twice while sharing parameters. Experiments on 16‑ and 20‑layer mixture‑of‑experts NanoChat backbones show that a single thought token brings a shallower model within 0.006–0.004 bits per byte of a double‑depth counterpart, recovering 67–81% of the improvement with roughly 48% fewer parameters.
arXiv:2606. 18587v1 Announce Type: cross Abstract: Decoder-only Transformers compute attention over the KV cache of preceding tokens.