The paper introduces the Graph Memory Transformer (GMT), a decoder‑only language model that replaces the traditional Feed‑Forward Network sublayer with an explicit learned memory graph. In GMT, token representations are routed through a bank of centroids linked by a directed transition matrix, using gravitational source routing and token‑conditioned target selection to produce gated displacements instead of retrieved values. The resulting model contains 82.2 M parameters, trains stably, and offers direct inspection of centroid usage and transition structure, though it slightly lags a larger GPT‑style baseline in validation loss and perplexity.
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...
The paper shows that a plain Transformer can serve as an effective graph learner by adding three lightweight modifications: simplified L₂ attention, adaptive RMS normalization, and an MLP-based positional encoding stem. These changes preserve token magnitude and enable the model to achieve high expressivity on graph benchmarks, outperforming more complex graph transformer variants. The results suggest that plain Transformers can act as a unified backbone for multimodal learning across language, vision, and graph domains.
arXiv:2604. 17324v2 Announce Type: replace-cross Abstract: Global self-attention drives modern graph transformers, yet the softmax at its core imposes a structural constraint rarely examined directly: every attention row is non-negative and sums to one, so each per-head output is a mass-conserving convex combination of value vectors.
arXiv:2502.09245v3 Announce Type: replace Abstract: In contrast to RNNs, which compress their history into a single hidden state, Transformers can attend to all past tokens directly. However, standar...
arXiv:2608. 08888v1 Announce Type: new Abstract: Autoregressive transformers compute along two axes: horizontally across generated tokens, and vertically through model depth.
arXiv:2605.10247v2 Announce Type: replace Abstract: Applying Large Language Models (LLMs) to graph-structured data usually involves multi-step pipelines in which textual node attributes are compresse...
Efficient Memory Crystallization (EMC) is a training‑free, test‑time framework that transforms each incoming graph domain into a compact, semantically faithful memory via a closed‑form distribution‑matching solution. By eliminating the need for auxiliary generative modules, EMC reduces redundant domain information and models inter‑domain dependencies through state‑evolving memories, yielding a tighter generalization error bound. Experiments show EMC outperforms state‑of‑the‑art baselines on non‑stationary graph data, cutting average runtime by 87.4% and GPU memory usage by 92.4% compared to a recent competitor.
The paper investigates how the choice of graph tokenization affects transformer expressivity. It analyzes three tokenization families—spectral, random‑walk, and adjacency—showing that each induces different depth requirements and that some tokenizations are inherently lossy or ill‑conditioned for certain tasks. The authors prove lower bounds and impossibility results for converting between tokenizations and validate these findings with experiments on synthetic and real‑world data.
The paper introduces Gated Recurrent Transformers, a depth‑sharing architecture that brackets a single shared core with fixed prelude and coda blocks and uses a lightweight projection and element‑wise update gate to modulate recurrent updates. This design allows functional specialization across recurrences while reducing memory footprint. Experiments show that, under equal FLOPs or parameter budgets, the recurrent model matches or surpasses deeper GPT‑2 Small baselines, achieving similar or better accuracy with fewer parameters and lower peak decoding memory.
arXiv:2606. 27538v1 Announce Type: cross Abstract: We introduce the context-ready transformer, a new recurrent neural network architecture built from a D-layer transformer block that pre-contextualizes each token before it enters the block.
arXiv:2608.15062v3 Announce Type: replace-cross Abstract: Scaling transformer language models creates an inherent tension between expressivity and memory efficiency. While unique weights across layer...
BF1 is a deterministic block‑aligned dyadic sparse‑attention retrofit designed to reduce the cost of causal attention in long‑context transformers. It combines a small exact local neighborhood, a global first block, and logarithmically spaced historical blocks, achieving O(n log n) token interactions per layer with O(log n) communication depth. On an NVIDIA RTX PRO 6000 Blackwell GPU, BF1 outperforms dense attention for 2K–4K tokens and delivers up to a 10.91× prefill speedup at 32K tokens, while retrofitting eight of 28 Qwen3‑0.6B layers reduces first‑token latency by up to 15.3% at 32K tokens and yields the lowest perplexity among compared sparse and dense training protocols.