Self-Attention Mechanism

Feature transformation improves predictive performance on tabular data by constructing informative abstractions from raw features. Recent generative approaches encode transformation knowledge into continuous embedding spaces for efficient exploration of candidate strategies, but face three key limitations: (1) overlooking hierarchical relationships between low-level features, operations, and high-level abstractions; (2) enforcing order-sensitive embeddings on inherently permutation-invariant transformation sequences, thereby introducing systematic bias; and (3) relying on gradient-based search, which is ill-suited to non-convex transformation spaces. We propose a framework with two complementary components. First, a permutation-invariant hierarchical module captures interactions across features, operations, and abstraction levels, with a self-attention pooling mechanism that maps semantically equivalent structures to consistent embeddings aligned with downstream performance. Second, a policy-guided multi-objective reinforcement learning strategy initializes the search from empirically strong seeds and jointly optimizes predictive accuracy and transformation efficiency. Extensive experiments on diverse tabular benchmarks demonstrate the effectiveness and robustness of our framework against strong baselines. Our code and data are publicly available at: https://github.com/RayLiu1103/PHER.

Graph modeling, a crucial task for representing complex relationships in graph-structured data, has achieved significant success in recent years. However, current graph modeling methods rely on traditional Graph Neural Networks and pre-training approaches to implicitly learn the underlying relational structure of graph data. Thus, these prior methods cannot capture the complex graph structure and correlations among inputs. In this paper, we introduce a novel Attention-based Normalizing Flow-based Approach\footnote{Our implementation and models will be released publicly for research reproducibility.} (ANFA or $α$) that provides an explicit, interpretable, and tractable Graph Modeling ($α$-Graph). In particular, we propose a new Unconditional Graph Normalizing Flow with an Invertible Attention Mechanism to capture the complex relational structure of graph data. To further enhance the expressiveness of the model, we introduce Conditional Graph Normalizing Flow with Learnable Queries that enables efficient modeling of correlations in graph-structured data. We show that our Conditional Graph Normalizing Flows behave similarly to Unconditional Graph Normalizing Flows, enhancing expressiveness while maintaining training stability and efficiency. Our experimental results on three benchmarks will illustrate the effectiveness and the state-of-the-art (SoTA) performance of the proposed $α$-Graph method.

Low-quality face recognition (LQFR) remains challenging due to the difficulty of matching degraded query (probe) images against low-quality (LQ) enrollment (gallery) imagery and the scarcity of training data for large-scale models. While recent face recognition (FR) models perform well on high-quality (HQ) imagery, their accuracy drops significantly on LQ images with extremely low signal-to-noise ratio (SNR). Moreover, fine-tuning HQ-pretrained models on LQ data often improves LQ recognition at the expense of HQ generalization. This trade-off becomes more pronounced in modern evaluation settings spanning multiple datasets with varying image quality levels. To address these limitations, we propose a unified framework that combines three main components: (1) Local Probability Margin (LPM), which estimates per-sample difficulty directly from the model's discriminative landscape; (2) Nested Attention Module (NAM), a new low-rank adapter module that embeds a self-attention mechanism within selected transformer layers; and (3) Quality Gating Protocol (QGP), where an off-the-shelf image quality estimator modulates the adapter contribution at test time, enabling a single model to handle the full quality spectrum without sacrificing HQ performance. Experiments on surveillance (TinyFace, SurvFace) and standard (IJB-B, IJB-C) face recognition benchmarks demonstrate consistent gains in both identification and verification. Code and models will be released at github.com/candllq/nam.

We present Turing-20B-A2B, a 20B-parameter Mixture-of-Experts language model that activates approximately 2B parameters per token, designed for long-context and latency-sensitive physical AI applications. The model adopts Quantile Routing in a dynamic top-k configuration, enabling token-adaptive expert allocation while maintaining balanced expert utilization and a controlled average compute budget. During deployment, we further apply capacity-constrained routing to prompt prefill for more regular and efficient expert execution, while retaining dropless routing during pretraining. Turing-20B-A2B also employs a hybrid attention architecture that combines Lightning Attention with a small number of full-attention layers for efficient long-context modeling. The model is pretrained with a progressive three-stage curriculum and extended to a native context length of 128K through continued pretraining, with further inference-time extension to 512K using YaRN. Despite its compact active-parameter budget, Turing-20B-A2B achieves, at the base-model stage, overall general capability exceeding Qwen3-8B Base and approaching Qwen3.5-9B Base, while maintaining strong long-context performance and favorable prefill-latency scaling. These results demonstrate an effective balance among model capability, long-context scalability, and practical inference efficiency.

We present the first systematic study of Massive activations (MAs) in layer-interleaved HLA LLMs and uncover two architecture-aligned morphologies: MAs consistently spike immediately before full attention layers, forming pre-attention spikes (PAS), and can persist through intervening linear attention layers, giving rise to inter-spike plateaus (ISP). As full attention becomes denser, successive PAS become increasingly connected through ISP, ultimately recovering the stable MA morphology of full attention LLMs. We establish the recurrence of this organization across five linear attention architectures, six hybridization configurations, five data domains, and representative open-source hybrid models spanning 1.2B to 397B total parameters. Controlled pretraining of GDN-based hybrids at scales up to 1.3B shows that both morphologies emerge early and respond asymmetrically to output gating: full attention output gating strongly attenuates their absolute magnitudes without eliminating their layerwise organization, whereas removing GDN gates yields comparatively modest amplification. Mechanistically, our systematic-outlier analysis supports a shared lifecycle account governed by the timing of MA cancellation. PAS follows a localized write-sink-cancel process, while the extended persistence of ISP is consistent with delayed cancellation. At the full attention limit, this account recovers the stable MA morphology characteristic of full attention LLMs. Our code is available at https://github.com/StartluxLabs/Massive-Activations-HLA.

Large language models (LLMs) have achieved remarkable breakthroughs across various applications. However, their architectures remain inefficient in pretraining due to two main limitations: (i) self-attention lacks an explicit inductive bias for locality, leading to redundant modeling of sequence-internal local information; (ii) mixture-of-experts (MoE) implicitly couples knowledge storage with computational pathways, hindering flexible access to sequence-external global knowledge. To overcome these limitations, we propose LoKiFormer, a novel LLM architecture that augments the standard decoder with two dedicated modules: 1) Local Fusion Attention (LFA), which incorporates a convolutional fusion to attention, explicitly capturing local patterns and allowing the attention to operate on more informative representations; 2) Knowledge Memory Module (KMM), which introduces a parametric key-value memory that explicitly stores global knowledge in addressable slots, decoupling storage from computation and enabling direct knowledge retrieval. Together, these modules enable LoKiFormer to achieve more efficient and effective integration of information at both levels. Experimental results show that LoKiFormer converges 1.33x faster in pre-training than baseline models, underscoring its superiority over existing LLM architectures.

A standard self-attention layer consists of two interacting circuits: the query-key circuit that governs attention allocation, and the output-value circuit that maps attended representations to predictions. Collapsed and factorized parameterizations of the query-key and output-value circuits lead to qualitatively different attention patterns. In particular, some parameterizations give sharper attention to task-relevant tokens, at a similar training loss. We analyze how the parameterizations of these circuits shape the parameter trajectories in single-layer self-attention models trained for next-token prediction. Through gradient-flow analysis, we show that factorization induces implicit rescaling of the two circuits' learning rates. We derive closed-form dynamics showing that output-value and query-key parameters move along a line, with relative speeds determined by their learning rates. Faster query-key learning relative to output-value learning thus produces sharper attention, as the model compensates for slower output-value learning by increasing attention mass on relevant tokens. Experiments show that differences in the relative learning rates of the two circuits govern attention concentration. This improves attention interpretability proxies while maintaining comparable predictive performance.

Attention-only dynamical theories model Transformer residual directions as particles aggregating on a sphere. We extend this framework by incorporating the feed-forward network (FFN) term as a local steering field acting on each token state. The resulting theory predicts that the tangential component of the FFN field is necessary for motion in residual-direction space, that critical residual directions correspond to nonlinear projective equilibria, and that a commutator defect determines when a finite attention--FFN block can be accurately approximated by a parallel, additive flow. Across GPT-2, Pythia, Mistral, and Llama models, the extended theory improves one-step angular prediction relative to an attention-only baseline, with the contribution of the FFN increasing from GPT-2 to Llama-3-8B. Intervention experiments show that retaining only the tangential FFN component preserves most model quality, whereas retaining only the radial component causes performance to collapse. The tangential component also preserves output diversity under aggregation pressure. As a practical application, layers with small commutator defects can be approximately parallelized with only a modest increase in loss, whereas layers with large defects degrade rapidly. These findings support the interpretation of FFN layers as directional steering fields that shape Transformer residual geometry and govern the feasibility of block-level interventions.

We convert 21 of 28 full-attention layers of Qwen3-0.6B-Base into KDA (Kimi Delta Attention) linear-attention layers on a single consumer-grade GPU budget, and ask a simple question: what exactly does the conversion break? After surgery, hidden-state alignment and end-to-end KL distillation drive the student close to its teacher in perplexity, yet multiple-choice accuracy stays near random chance (25-29% vs. the teacher's 50.6% on C-Eval). Using a four-permutation diagnostic that rotates answer options while holding content fixed, we show the model sticks to option labels (predicting "A" 81% of the time; 106/161 questions keep the same label under all four rotations) rather than following answer content -- an interface injury that standard distillation metrics cannot see. A 1,000-step format-targeted completion-only KL stage repairs the interface (+12.48 points on C-Eval, label-stickiness roughly halved), after which persona SFT and one round of on-policy DPO preserve benchmark scores within noise. We release code, weights, recipes, and the full audit trail, and distill the engineering lessons -- including an FP32-master failure mode in which bf16 optimizer updates are silently swallowed -- that made convergence possible at this budget.

We revisit the regret loss framework introduced in Park et al. (2025), which uses decision-theoretic regret as a direct loss function for training models to make better decisions, through the lens of probability-simplex policies. Our first result shows that a single-layer self-attention model trained with regret loss admits a stationary point whose forward-pass exactly matches smoothed fictitious play with the appropriate stepsize that ensures no-regret behavior-i.e., for any given policy input, the model outputs the same update that smoothed fictitious play would produce. In parallel, we also newly introduce a swap-regret loss function, which extends the regret-loss framework beyond external regret and enables models to directly optimize for swap-deviation robustness. We further show that this swap-regret loss admits a stationary point whose forward pass implements the corresponding swap-regret update induced by classical Blum-Mansour no-pass implementation algorithm, with each head implementing an external-regret update via smoothed fictitious play. Together, these results show that regret-trained attention can realize differentiable mechanisms whose deployment induces equilibrium behavior in games: external-regret dynamics lead to coarse correlated equilibrium, while swap-regret dynamics lead to correlated equilibrium. Thus, regret-based objectives steer minimal attention architectures toward online-learning dynamics with game-theoretic guarantees, without supervised traces of those algorithms.

Recent work has shown that large language models (LLMs) can iteratively improve their outputs by incorporating generated samples and their corresponding evaluation scores as in-context examples. Despite these empirical findings, the theoretical foundations underlying this phenomenon remain poorly understood. In this paper, we show that score-conditioned In-Context Learning (ICL) admits a structural correspondence to policy gradient optimization. We first provide a constructive proof that self-attention mechanisms can implement reward-weighted aggregation analogous to the REINFORCE algorithm under specific weight matrix configurations, and discuss the relationship between this construction and the behavior of pretrained transformers. The correspondence is directional in hidden-state space and holds exactly only under the stated simplifying conditions; we quantify its strength empirically. Within our simplified hidden-state model, we furthermore derive an exact upper bound on the distribution shift induced by a bounded attention update, yielding a trust-region-like analogy to KL-constrained policy optimization. We validate our theory through extensive experiments across multiple LLMs, demonstrating that LLMs effectively utilize score information to shift output distributions toward high-scoring exemplars, and that attention weights exhibit a strong correlation with example scores.

Video diffusion transformers (vDiTs) generate high quality video but introduce extremely high compute cost due to the long diffusion timesteps and self attention computation. As diffusion timesteps are reduced, the computation cost of self attention becomes the dominant bottleneck. Existing acceleration approaches largely inherit sparse attention techniques from large language models, which fail to consider the unique spatiotemporal correlation of video data. This paper presents Kaleido, an algorithm hardware codesign that accelerates all operations in vDiTs by exploiting channel-wise spatiotemporal correlations in latent space. Based on this insight, we propose a lightweight channelwise reuse algorithm that skips redundant computations by reusing partial results while preserving higher generative quality than prior methods (>17 dB). To efficiently support this algorithm, we design a systolic array like accelerator with reconfigurable processing elements and a lightweight data dispatcher to mitigate irregular sparsity and data access patterns introduced by our reuse algorithm. Evaluations across three mainstream vDiT models show that Kaleido achieves up to 5.9x speedup and 16.0x energy savings over state of the art accelerators.

The pursuit of autonomously self-improving models has attracted growing interest in the era of large-scale foundation models. Drawing inspiration from the concept of "enlightenment" or "aha moment" in human brain, we hypothesize that large models exhibit an analogous enlightenment phenomenon-a latent capacity for sudden capability boost. Then, we propose Enlightenment, a novel training-free post-tuning paradigm for large-scale models. Our approach modifies shortcuts for key modules/layers without weight updates, while existing training-free ones predominantly manipulate attention weights. We introduce two architecture-specific instantiations: i) For large language models, we propose attention head-mixing shortcuts that recalibrate attention weights by linking the initial attention head's output to all other target heads, modulated by an adaptive scaling factor initialization strategy. ii) For vision-language models, we apply a lightweight scalar-modulated factor to residual connections in the decoder layers, regulating information flow. Extensive experiments show that Enlightenment efficiently unlocks the latent potential of pre-trained networks, yielding remarkable performance improvements across diverse benchmarks and models.

Vision Transformers (ViTs) have demonstrated remarkable performance in computer vision tasks. However, their self-attention mechanism often diffuses focus across background regions, relying on spurious correlations rather than object-relevant cues. Inspired by inhibitory mechanisms observed in biological vision systems, we propose the Inhibited Self-Attention (ISA), a novel self-attention that integrates inhibitory signals to enhance feature selectivity and suppress spurious responses. In contrast to conventional self-attention, which relies solely on positive attention values due to softmax normalization, our approach retains and utilizes negative attention scores to suppress irrelevant features and sharpen focus on objects of interest. Experiments across multiple datasets, including ImageNet-1k and COCO, and several robustness benchmarks demonstrate that ISA enhances object-centric selectivity, reduces shortcut reliance, and improves out-of-distribution generalization. Our analysis of relevance maps confirms that ViTs with ISA exhibit sharper, more localized focus on object-relevant regions while reducing distractions from non-relevant (background) features, enabling more reliable models. We release our code at https://github.com/prdvanderwal/inhibited-self-attention

The $\mathcal{O}(N^2)$ complexity of attention over $N$ tokens remains a computational bottleneck in transformer models. Vector-Quantized (VQ) attention reduces this to $\mathcal{O}(MN)$ by representing keys with $M$ codewords, but applies uniform codebook capacity regardless of where attention mass concentrates: high-attention regions of key space may be coarsely approximated while low-attention regions waste representational capacity. We propose Adaptive Vector-Quantized (AVQ) Attention, which adaptively allocates codebook capacity based on attention importance. Starting from a small set of codewords, our method identifies the most important codes during the forward pass and refines them with pre-learned child codewords, achieving fine-grained quantization where it matters most while maintaining coarse quantization elsewhere. We develop an implementation using custom Triton kernels that enables the full adaptive refinement process, including importance scoring, child codeword insertion, and parent contribution replacement, to be carried out within the tiled computation paradigm of Flash Attention with minimal overhead. Our approach maintains $\mathcal{O}(MN)$ complexity while achieving improved accuracy-efficiency trade-offs compared to fixed-codebook VQ-attention.

The \textit{11th Affective Behaviour Analysis in-the-wild Competition} includes the Multi-Task Learning Challenge, where participants develop a unified framework for Valence-Arousal Estimation, Expression Recognition, and Action Unit Detection. The challenge lies in learning emotion-related representations that generalize across subjects while remaining robust to spurious factors such as identity, illumination, pose, and demographic variation. To aggregate features extracted by a pre-trained backbone into a compact representation for prediction, attention mechanisms selectively weight the most informative facial regions. However, these attention weights can still capture dataset-specific correlations rather than genuine affective cues. To address this limitation, we propose an attention pooling framework that combines causal supervision with cross-covariance regularization of attention components, encouraging subject-invariant attention and non-redundant representations that improve generalization. Our method achieves $CCC_{VA}=0.5123$ for VA estimation on the official validation set, together with $F_{EX}=0.3116$ and $F_{AU}=0.3974$ for expression recognition and action unit detection, respectively, resulting in an overall $P$ score (the sum of the individual task metrics) of $1.2214$.

Recent LLM-based agent systems continuously accumulate context across multi-turn interactions, tool invocations, and cross-session workflows. Replaying the full history for every request quickly becomes impractical: long contexts increase prefill cost, may exceed context limits, and often bury task-relevant evidence in irrelevant content, degrading both serving efficiency and output quality. We propose Akashic, a low-overhead memory system built around MemAttention, which organizes context into bounded chunks and models semantic relationships across chunks, preserving cross-chunk evidence without repeatedly rewriting the full history. Akashic further applies hardware-software co-designed memory placement to co-locate likely co-retrieved chunks, reducing retrieval fragmentation and I/O overhead. Across four representative workloads and three model sizes, Akashic improves task accuracy by up to 10.2 points, throughput by up to 1.21x, and sustainable request rate by up to 1.88x over strong prior memory baselines.

Hybrid attention models improve long-context efficiency by retaining only a subset of full-attention layers and replacing the remaining layers with linear attention. However, the effectiveness of Transformer-to-hybrid conversion critically depends on which layers preserve full attention. Existing hybrid layer selection methods typically rely on heuristic strategies such as fixed placement patterns or layerwise scoring, implicitly treating layer importance as isolated and overlooking the interdependent layer effect under a global hybrid configuration. In this work, we formulate hybrid layer selection as a budget-constrained subset optimization problem. We further propose FlashMorph (Fast LAyer Selection for Hybrid MORPHing), an effective, efficient and scalable layer selection method for Transformer-to-hybrid conversion. FlashMorph first constructs a morphable model by equipping each full-attention layer with a converted linear-attention branch. It then freezes all model weights and jointly optimizes layerwise gates on synthetic long-context retrieval data, with a linearization regularization that encourages the model to rely on linear attention for efficiency. The learned gates are discretized under a preset full-attention budget to instantiate the hybrid architecture, followed by standard logits distillation and long-context finetuning. Extensive experiments show that FlashMorph discovers more effective hybrid configurations, preserves strong long-context recall and general benchmark performance while substantially reducing layer selection cost compared with existing layer selection methods, demonstrating its effectiveness, efficiency, and scalability.

Hybrid attention models that mix full and sliding-window attention across layers offer a promising approach to efficient long-context inference, but the critical question of \emph{which layers} should retain full attention remains unsolved. Existing methods use either fixed periodic patterns or attention-based heuristics that may not capture what matters for downstream accuracy. We propose NLL-guided layer selection, a training-free method that directly measures each layer's importance by computing the negative log-likelihood degradation on answer tokens when that layer uses sliding-window instead of full attention. On LongMemEval with Qwen3-4B, our method achieves 64.6% accuracy using only 1/4 full-attention layers, matching the 1/2-FA periodic baseline (65.0%) while halving the computational budget. NLL-guided selection outperforms the SWAA-reported periodic 1/4-FA baseline by 10.4 percentage points and a matched LightTransfer-style baseline by 26.4 percentage points. De-confounding analysis shows the signal is consistent with long-range attention needs rather than generic layer sensitivity. The method requires only $\sim$15 minutes of one-time calibration, advancing the efficiency-accuracy Pareto frontier for long-context LLM deployment.

Lifelong continual learning remains an obstacle on the path to human-like intelligence. Modern transformers show sparks of intelligence with in-context learning. The quadratic nature of attention, however, prohibits transformers from performing this process on arbitrarily long sequences. In this work, we argue that extending in-context learning to lifelong settings is a practical solution for continual learning in AI agents. In particular, we argue that \emph{parametric forms of attention} are needed to understand a lifetime of context with transformers on a fixed hardware budget. These attention mechanisms learn the relationship between keys and their associated values at test-time with parametric regression. Our generalization of parametric approaches (linear attention, state-space models, fast weight programmers, and test-time training layers) contrasts with nonparametric counterparts like softmax attention. They replace the ever-growing key-value cache with an online-trainable neural network, maintaining a constant memory footprint. We highlight how parametric attention currently fall short of lifelong learning due to limited memory capacity or costly online updates. To address these issues, we pose a set of open questions with novel insights to guide the field toward long-horizon agents.

Large Transformer models function as Dense Associative Memories (DAMs), retrieving knowledge via high-dimensional attractor dynamics driven by the self-attention mechanism \citep{ramsauer2020hopfield, wu2024attention}. However, adapting these frozen memory systems to new tasks presents a fundamental ``Plasticity-Stability'' dilemma. Current methods either risk catastrophic interference by modifying synaptic weights directly (e.g., LoRA) \citep{hu2021lora} or degrade associative capacity by clogging the retrieval buffer with static prompt tokens (e.g., VPT) \citep{jia2022vpt}. In this work, we propose \textbf{H-Res} (Hierarchical Residual Steering), a mechanism that modulates the effective energy landscape of the Transformer without altering its global equilibrium or expanding its sequence length. By formulating adaptation as a control problem on the activation manifold \citep{chen2018neuralode}, H-Res learns a state-dependent vector field that steers token trajectories into task-specific basins of attraction. We formally prove that H-Res preserves the attention entropy of the foundation model and facilitates Neural Collapse \citep{papyan2020prevalence}. Empirically, Manifold Steering outperforms global weight modification by 26% on associative retrieval tasks and eliminates the computational overhead of prompt-based methods, scaling effectively to structured domains \citep{zha2023vtab}.

Multi-turn jailbreaks can evade turn-level moderation by spreading unsafe intent across a dialogue through gradual escalation, reframing, and role manipulation. We address multi-turn jailbreak detection as a conversation-level classification problem and introduce an efficient hierarchical detector that avoids expensive long-context concatenation while retaining cross-turn reasoning. The model encodes individual turns to form compact turn representations and applies a lightweight conversation module that captures dialogue dynamics and selectively attends to fine-grained evidence when needed. On a challenging evaluation benchmark of 14,038 conversations, our approach achieves an F1 of 0.9394, outperforming Claude Opus 4.7, the strongest competing baseline, by 0.07 while halving its false-positive rate. Ablation studies confirm that each architectural component contributes meaningfully, with combining cross-attention and self-attention in the conversation module yielding a 2.26 percentage point reduction in false-positive rate over the self-attention-only variant.

Modern language models increasingly adopt hybrid architectures that combine full attention with efficient attention modules, such as sliding-window attention (SWA) and recurrent sequence mixers. However, how these efficient modules shape model capabilities remains poorly understood. To address this gap, we conduct a systematic analysis across hybrid architectures from three perspectives: scaling behavior, mechanism analysis, and architecture design. First, from a scaling perspective, we find that efficient-attention design primarily affects how fast long-context capability emerges, while different hybrids eventually converge to comparable long-context performance under sufficient training. Second, mechanistically, we show that long-range retrieval is mainly carried by full attention, whereas efficient attention shapes its optimization trajectory. This explains a counter-intuitive phenomenon we call Large-Window Laziness: larger SWA windows can delay the formation of retrieval heads in full-attention layers. Third, guided by this mechanism, we show that applying NoPE to only the full-attention layers of a small-window SWA hybrid substantially improves long-context performance with negligible impact on short-context performance.

Transformer models are increasingly used as computational models of cognition and neural representation, so the mechanism implemented by self-attention is of interest beyond engineering performance. A complementary tradition in cognitive science models coordination, binding, and memory through dynamical interactions such as oscillator synchrony; we bring this mechanism into self-attention by introducing the Kuramoto Attention layer, whose value update is a synchronization step. Each token carries a bank of phase oscillators, so its hidden state lives on a high-dimensional torus. The attention weights form an adaptive coupling graph, and using the raw phase states as values makes the value update exactly the Kuramoto coupling direction for fixed attention weights. The softmax selects which oscillators couple, while the value path moves each token toward the attention-weighted circular mean of the tokens it selects. We train Kuramoto Attention on enwiki8 and CodeParrot against parameter-matched RoPE and SwiGLU transformers. At 5M parameters on CodeParrot, it improves on the transformer by both median and mean, with mean gaps of 0.012 validation and 0.010 test bits per byte. At 5M on enwiki8, all six runs have lower validation/test medians than the transformer and all-seed means within 0.01 BPC; five of six also form a tight lower-mean cluster. At 1M, it trails by about 0.02 BPC on enwiki8 and by 0.013-0.015 bits per byte on CodeParrot. Ablations and phase diagnostics show how the layer's synchronization and geometry-motivated components shape model performance. The result is a self-attention mechanism whose learned computation can be read directly as adaptive synchronization on phase states.

Pre-propagation graph neural networks (PPGNNs) push all graph-dependent computation into a preprocessing step and train only on the resulting dense hop features, which makes them highly scalable. A puzzle in this regime is that more complex hop aggregators do not reliably outperform simpler ones: on many benchmarks, a plain MLP-based aggregator matches or beats hop-attention variants. We revisit this behavior from a graph-filter perspective. Over a precomputed diffusion basis, existing PPGNNs differ mainly in how filter coefficients are shared across nodes and feature channels, rather than simply in raw aggregator capacity. MLP-based architectures learn channel-dependent filters that are largely shared across nodes, while hop-attention-based architectures learn node-dependent mixtures that are largely shared across channels. This reveals a missing regime in standard PPGNN designs: joint node- and channel-adaptive filtering under the pre-propagation computational contract. We propose FilterMoE, a mixture-of-experts PPGNN in which a small bank of learnable Chebyshev filter experts is routed jointly over nodes and channels by a 3D gating tensor. Across eleven homophilic and heterophilic benchmarks, FilterMoE outperforms strong PPGNN baselines on nine datasets and ranks first on all three large-scale benchmarks, improving the average test score by 1.53 points. These results establish joint node-channel filter routing as a robust alternative to dataset-specific hop-aggregator selection.

We use a mean-field-based transformer model to theoretically investigate how auxiliary variables, such as positional encoding, prevent mode collapse of self-attention mechanisms. The use of mean-field transformers to analyze the properties of self-attention mechanisms has garnered significant attention in recent years due to their ability to comprehensively analyze token interactions. However, analysis of this simple model suggests that mode collapse, where token distributions degenerate to a single point, occurs during long inferences (i.e., many layers), indicating a discrepancy with reality. This study investigates this mean-field transformer model and demonstrates that the introduction of auxiliary variables, such as positional encoding, acts as a counterforce against theoretical mode collapse. Specifically, we show that in the theoretical scheme, the energy-maximizing distribution does not degenerate to a single point; instead, it is characterized by a pushforward of the auxiliary variable distribution, thereby avoiding concentration in the Dirac measure. Our main examples are the positional encoding and the fixed prompt insertion treated as a parallel auxiliary-variable mechanism. Furthermore, we demonstrate that positional encoding and prompt insertion possess universality of representation in the limit, meaning that the limit distribution of inference can exactly represent a wide class of distributions. We also analyze several key properties of positional encoding and metastability, and validate our theoretical results through mathematical experiments.

Standard transformer architectures apply a single attention mechanism uniformly across all tokens and sequence positions, irrespective of local context or computational budget. We propose Meta-Attention, a framework that dynamically routes each token to the most appropriate attention strategy -- full softmax attention, linear (kernel) attention, or sliding-window local attention -- via a Bayesian Meta-Controller. Unlike prior routing approaches that use deterministic or prior-free learned routing, the Meta-Controller treats per-token mechanism selection as posterior inference under a compute-aware Dirichlet prior: routing weights are the output of an amortised variational posterior q(alpha | x_t; phi) trained with an Evidence Lower Bound (ELBO) objective that jointly encodes task performance and attention-mechanism cost. This design produces principled routing uncertainty estimates that govern the soft-to-hard routing transition, mitigates routing collapse without ad hoc load-balancing losses, and yields better compute-performance trade-offs than deterministic or prior-free learned routing at negligible overhead. Phase 1 empirical results on a Tiny LM benchmark confirm core predictions: the Bayesian controller's learned routing distribution implies a projected normalised FLOP cost of 25.1% under hard routing, vs. 59.3% for the prior-free baseline (-34.2 pp), and reduces routing entropy from 55.8% to 43.3% (-12.5 pp), demonstrating that the Dirichlet prior prevents routing collapse while the non-Bayesian model defaults to full attention. We present the Bayesian architecture, ELBO training objective, and a Phase 1 PyTorch prototype validating forward-pass correctness, posterior diversity, and a controlled ablation against a prior-free baseline. Code available at: https://github.com/KFEAL/meta-attention

Reinforcement Learning (RL) has long been a powerful solution to various problems in communication networks. However, traditional RL models still face with several limitations. Not only do they rely on large numbers of interactions with the environment, but they are also limited in terms of modeling long-term relationships and tackling partial observability. In recent years, the Transformer model has demonstrated the ability to enhance RL models, allowing them to overcome these issues. Particularly, the self-attention mechanism within the Transformer enables efficient modeling of long-range dependencies and global correlations, as well as accelerates training processes and handles heterogeneous data modalities. In this paper, we present a comprehensive survey of Transformer-based RL algorithms and their applications in communication networks. Specifically, the paper provides the mathematical background of RL and Transformer architectures, along with insights into key issues such as resource allocation, computation offloading, routing, and trajectory control, and network security. We conclude the paper by discussing challenges, open issues, and notable future research directions, including Transformer-enhanced DRL algorithms for semantic communication and network optimization.

A unified simulator that can model diverse physical phenomena without solver-specific redesign is a long-standing goal across simulation science. We present a learning-based particle simulator built on a single transformer architecture to model cloth, elastic solds, Newtonian and non-Newtonian fluids, granular materials, and molecular dynamics. Our model follows a prediction-correction design on a shared Lagrangian particle representation. An explicit predictor first advances particles under the known external forces, producing an intermediate state that captures externally driven motion but not inter-particle interactions. A learned corrector then predicts the residual position and velocity updates through three stages: a particle tokenizer that encodes local particle-particle, particle-boundary, and topology-guided interactions; a super-token encoder that hierarchically merges particle tokens into a compact set of super tokens via alternating self-attention and token merging; and a super-token decoder that lifts these super tokens back to particle resolution through cross-attention to predict per-particle position and velocity corrections. Progressive token merging reduces the attention cost at successive encoder layers by halving the token count at each level, and the decoder communicates through the compact super-token set rather than full particle-to-particle attention. Across the six dynamics categories, the same architecture generalizes to unseen materials, boundary configurations, initial conditions, and external forces. We further demonstrate downstream interactive control, inverse design, and learning from real-world manipulation data, reducing the need for per-phenomenon solver engineering.

Mixture-of-experts (MoE) architectures enable trillion-parameter LLMs with sparsely activated experts. Expert parallelism (EP) is a widely adopted MoE training strategy, but it suffers from severe all-to-all communication bottlenecks, which is exaggerated by the limited inter-node network bandwidth as the growing model size requires distributing experts across GPU nodes. Prior work focused on overlapping these all-to-all communications with feed-forward network (FFN) and self-attention computations, which often leaves residual network-bound stalls due to inherent imbalance in attention and FFN layers' computation-communication ratios. We present DisagMoE, a disaggregated MoE training system that jointly optimizes model placement and scheduling for maximal efficiency. DisagMoE separates attention and FFN layers into disjoint GPU groups, introduces a multi-stage pipeline with uni-directional, many-to-many communications, and employs a computation-communication roofline model to balance GPU and network bandwidth allocation among the attention and FFN groups. DisagMoE is implemented on Megatron-LM, and evaluation shows that DisagMoE improves training efficiency across multiple MoE models with up to 1.8x speedup on 16-node 8xH800 clusters.

Pareto set learning (PSL) is an emerging paradigm in multi-objective optimization that trains neural networks to map preference vectors to Pareto optimal solutions. However, existing PSL methods primarily focus on solving a single multi-objective optimization problem at a time. This limitation not only increases computational costs in multi-objective multitask optimization scenarios by requiring a separate model for each task, but also fails to exploit the inter-task correlations across tasks. To address this, we propose a Cross-tAsk correlation-aware Pareto Set Learning (CoAction) framework, which leverages task-aware transformer to handle multiple tasks simultaneously. Specifically, by assigning task-specific embedding vectors to individual tasks, the model effectively distinguishes between tasks while facilitating knowledge sharing among them. We utilize a Transformer encoder as the backbone architecture to leverage its self-attention mechanism for capturing complex task dependencies. The proposed approach is evaluated on comprehensive multitask test suites covering both benchmark problems and real-world applications, demonstrating effectiveness and competitive performance in Hypervolume, Range, and Sparsity.

Training stable biological foundation models requires rethinking attention mechanisms: we find that using sigmoid attention as a drop in replacement for softmax attention a) produces better learned representations: on six diverse single-cell datasets, sigmoid achieves 25% higher cell-type separation, better cell-type cohesion metrics, and lower validation loss, b) faster training, models with sigmoid attention train up to 10% faster than their softmax counterparts, and c) more stable training by eliminating inherent sources of instability in softmax attention. We establish that sigmoid attention has globally bounded derivatives ($\leq 0.25$) as opposed to softmax, and a diagonal Jacobian structure in contrast with softmax's dense coupling, which together help alleviate training instabilities. In stress tests on 160M-parameter bidirectional attention models trained without gradient clipping on 8K-token sequences, softmax diverges catastrophically, with gradients exploding by four orders of magnitude, while sigmoid remains stable. Finally, we implement and open-source TritonSigmoid, an efficient GPU kernel that achieves 515 TFLOPS on H100 GPUs, outperforming both FlashAttention-2 and FlashSigmoid, with native padding support, which is essential for biological sequences. Our results establish sigmoid attention as both theoretically grounded and empirically superior for biological foundation models. Code is available at https://github.com/MSDLLCpapers/triton-sigmoid

Over the past decade, hyperspectral image (HSI) classification has drawn considerable interest due to HSIs' ability to effectively distinguish terrestrial objects by capturing detailed, continuous spectral information. The strong performance of recent deep learning techniques in tasks like image classification and semantic segmentation has led to their growing use in HSI classification, due to their ability to capture complex spatial and spectral features more effectively than traditional methods. This paper presents MixerCA, a novel lightweight model for HSI classification that leverages depthwise convolution and a self-attention mechanism. MixerCA integrates depth-wise convolutions, token and channel mixing, and coordinate attention into a unified structure to decouple spatial and channel interactions, maintain consistent resolution throughout the network, and directly process HSI patches. Extensive experiments on four hyperspectral benchmark datasets reveal MixerCA's clear advantages over several competing algorithms, including 2D-CNN, 3D-CNN, Tri-CNN, HybridSN, ViT, and Swin Transformer. The source code is publicly available at https://github.com/mqalkhatib/MixerCA.

Vision Transformers (ViTs) have achieved remarkable success across vision tasks, yet recent studies show they remain vulnerable to backdoor attacks. Existing patch-wise attacks typically assume a single fixed trigger location during inference to maximize trigger attention. However, they overlook the self-attention mechanism in ViTs, which captures long-range dependencies across patches. In this work, we observe that a patch-wise trigger can achieve high attack effectiveness when activating backdoors across neighboring patches, a phenomenon we term the Trigger Radiating Effect (TRE). We further find that inter-patch trigger insertion during training can synergistically enhance TRE compared to single-patch insertion. Prior ViT-specific attacks that maximize trigger attention often sacrifice visual and attention stealthiness, making them detectable. Based on these insights, we propose PASTA, a twofold stealthy patch-wise backdoor attack in both pixel and attention domains. PASTA enables backdoor activation when the trigger is placed at arbitrary patches during inference. To achieve this, we introduce a multi-location trigger insertion strategy to enhance TRE. However, preserving stealthiness while maintaining strong TRE is challenging, as TRE is weakened under stealthy constraints. We therefore formulate a bi-level optimization problem and propose an adaptive backdoor learning framework, where the model and trigger iteratively adapt to each other to avoid local optima. Extensive experiments show that PASTA achieves 99.13% attack success rate across arbitrary patches on average, while significantly improving visual and attention stealthiness (144.43x and 18.68x) and robustness (2.79x) against state-of-the-art ViT defenses across four datasets, outperforming CNN- and ViT-based baselines.

For state-of-the-art image understanding, Vision Transformers (ViTs) have become the standard architecture but their processing diverges substantially from human attentional characteristics. We investigate whether this cognitive gap can be shrunk by fine-tuning the self-attention weights of Google's ViT-B/16 on human saliency fixation maps. To isolate the effects of semantically relevant signals from generic human supervision, the tuned model is compared against a shuffled control. Fine-tuning significantly improved alignment across five saliency metrics and induced three hallmark human-like biases: tuning reversed the baseline's anti-human large-object bias toward small-objects, amplified the animacy preference and diminished extreme attention entropy. Bayesian parity analysis provides decisive to very-strong evidence that this cognitive alignment comes at no cost to the model's original classification performance on in- (ImageNet), corrupted (ImageNet-C) and out-of-distribution (ObjectNet) benchmarks. An equivalent procedure applied to a ResNet-50 Convolutional Neural Network (CNN) instead degraded both alignment and accuracy, suggesting that the ViT's modular self-attention mechanism is uniquely suited for dissociating spatial priority from representational logic. These findings demonstrate that biologically grounded priors can be instilled as a free emergent property of human-aligned attention, to improve transformer interpretability.

The quadratic computational complexity of the standard attention mechanism constitutes a fundamental bottleneck for large language models in long-context inference. While existing KV cache compression methods alleviate memory pressure, they often sacrifice generation quality and fail to address the high overhead of floating-point arithmetic. This paper introduces DASH-KV, an innovative acceleration framework that reformulates attention as approximate nearest-neighbor search via asymmetric deep hashing. Under this paradigm, we design an asymmetric encoding architecture that differentially maps queries and keys to account for their distinctions in precision and reuse characteristics. To balance efficiency and accuracy, we further introduce a dynamic mixed-precision mechanism that adaptively retains full-precision computation for critical tokens. Extensive experiments on LongBench demonstrate that DASH-KV significantly outperforms state-of-the-art baseline methods while matching the performance of full attention, all while reducing inference complexity from O(N^2) to linear O(N).

Vision Transformer (ViT) has brought new breakthroughs to the field of image classification by introducing the self-attention mechanism and Graph Convolutional Networks(GCN) have been proposed and successfully applied in data representation and analysis. However, there are key challenges which limit their further development: (1) The patch size selected by ViT is crucial for accurate predictions, which raises a natural question: How to select the size of patches properly or how to comprehensively combine small patches and larger patches; (2) While the spatial structure information is important in vision tasks, the 1D position embeddings fails to capture the spatial structure information of patches more accurately; (3) The GCN can capture the local connectivity relationships between image nodes, but it lacks the ability to capture global graph structural information. On the contrary, the self-attention mechanism of ViT can draw the global relation on image patches, but it is unable to model the local structure of image. To overcome such limitations, we propose the Hierarchical Vision Transformer Enhanced by Graph Convolutional Network (GCN-HViT) for image classification. Specifically, the Hierarchical ViT we designed can model patch-wise information interactions on a global scale within each level and model hierarchical relationships between small patches and large patches across multiple levels. In addition, the proposed GCN method functions as a local feature extractor to obtain the local representation of each image patch which serves as a 2D position embedding of each patch in the 2D space. Meanwhile, it models patch-wise information interactions on a local scale within each level. Extensive experiments on 3 real-world datasets demonstrate that GCN-HViT achieves state-of-the-art performance.

Mixture-of-Experts (MoE) layers increase model capacity by activating only a small subset of experts per token, and typically rely on a learned router to map hidden states to expert assignments. In this work, we ask whether a dedicated learned router is strictly necessary for MoE routing. We propose Self-Routing, a parameter-free routing mechanism that uses a designated subspace of the token hidden state directly as expert logits, eliminating the router projection entirely while leaving the rest of the MoE layer unchanged. We evaluate Self-Routing on language modeling across different expert counts and model scales, and on ImageNet-1K classification by comparing it against a standard learned router, random-routing baselines, and dense non-MoE baselines. Our results show that Self-Routing remains competitive with the learned-router baseline while removing all dedicated routing parameters, and yields more balanced expert utilization, with about 17 % higher average normalized routing entropy and no explicit load-balancing loss. On ImageNet-1K with DeiT-S/16, Self-Routing also slightly improves over the corresponding learned-router MoE. These findings suggest that effective MoE routing can emerge from the hidden representation itself without requiring a separate learned router module.

Channel hopping (CS) communication systems must adapt to interference changes in the wireless network and to node mobility for maintaining throughput efficiency. Optimal scheduling requires up-to-date network state information (i.e., of channel occupancy) to select non-overlapping channels for links in interference regions. However, state sharing among nodes introduces significant communication overhead, especially as network size or node mobility scale, thereby decreasing throughput efficiency of already capacity-limited networks. In this paper, we eschew state sharing while adapting the CS schedule based on a learning-based channel occupancy prediction. We propose the MiLAAP attention-based prediction framework for machine learning models of spectral, spatial, and temporal dependencies among network nodes. MiLAAP uses a self-attention mechanism that lets each node capture the temporospectral CS pattern in its interference region and accordingly predict the channel occupancy state within that region. Notably, the prediction relies only on locally and passively observed channel activities, and thus introduces no communication overhead. To deal with node mobility, MiLAAP also uses a multi-head self-attention mechanism that lets each node locally capture the spatiotemporal dependencies on other network nodes that can interfere with it and accordingly predict the motion trajectory of those nodes. Detecting nodes that enter or move outside the interference region is used to further improve the prediction accuracy of channel occupancy. We show that for dynamic networks that use local CS sequences to support relatively long-lived flow traffics, the channel state prediction accuracy of MiLAAP is remarkably ~100% across different node mobility patterns and it achieves zero-shot generalizability across different periods of CS sequences.

This paper proposes a general machine learning framework called the localization method, which is fundamentally built on two core concepts: localization kernels and local means -- key components that underpin the self-attention mechanism. To establish a rigorous theoretical foundation, the framework is formally defined through two essential pillars: the formulation of the local(-ized) model and the localization trick. We systematically investigate the connections between the localization method and a wide range of existing machine learning models/methods, including (but not limited to) kernel methods, lazy learning, the MeanShift algorithm, relaxation labeling, Hopfield networks, local linear embedding (LLE), fuzzy inference, and denoising autoencoders (DAEs). By dissecting these relationships, we clarify the broader theoretical significance of the localization method and demonstrate its practical applicability across diverse machine learning tasks. Furthermore, we explore advanced extensions of the framework, such as adaptive kernels, hierarchical local models, and non-local models. Notably, we show that the Transformer -- a cornerstone of modern sequence modeling -- can be constructed using hierarchical local models, revealing the ability of the localization method to unify and generalize state-of-the-art architectures. This work not only provides a unified theoretical lens to reinterpret existing models but also offers new methodological tools for designing flexible, data-adaptive learning systems.