Causal Attention

Clinical data are split across incomplete cohorts with partially overlapping variables. Causal discovery under missing data and multi-source integration have developed separately, and tabular foundation models cannot take a causal graph without retraining. We make two contributions. First, causal graph fusion learns a partial ancestral graph on the shared variables, fuses it into each cohort's graph under temporal background knowledge and lets Markov blankets guide re-imputation, so missing data are handled inside discovery. Second, causal attention steering compiles the learned graph into the attention of a frozen tabular foundation model at inference. Against known ground truth, fusion dominates pooled and per-cohort baselines in every missingness and heterogeneity condition tested (p<0.001), and steering improves discrimination under distribution shift (p<10^{-5}), a gain that a random graph does not produce. In two spinal cord injury cohorts, fusion orients age, albumin and hemoglobin as ancestors of length of stay, a set no single cohort orients in full, and yields compact, temporally ordered predictor sets that match or exceed automatic feature selection on four of five outcomes.

Causal masking obstructs the column balancing used by optimal-transport (OT) attention. We study a prefix-consistent unbalanced operator and characterize the positional prior it induces. Under bounded scores, for any fixed balancing strength $ρ<1$, its fixed point dilutes attention: every fixed window receives a vanishing share as the context grows. For constant scores, we derive an exact scale-invariant prior. We then classify positional encodings by whether their distant attention mass remains bounded away from one. Discrete window-mass bounds quantify the finite-length dilution of summable log-distance priors near exponent one. Preregistered experiments with 25M- and 100M-parameter language models establish improved loss extrapolation over softmax; exploratory comparisons favor an explicit distance bias at a fraction of the cost. CABLE outperforms RoPE with evaluation-time YaRN on every seed at three training lengths. Changing only the bias beyond the training length moves long-context loss in the predicted direction on all 30 tested checkpoints. Even with exponents constrained above one, many heads approach the summability boundary during training. A margin selected by prior-only calculations for a specified window, length range and growth tolerance controls distant-mass growth in the tested setting. These results connect loss extrapolation to positional decay over the lengths actually evaluated.

Agentic AI drives significant productivity gains across diverse domains, but concurrent long-context prefills from multiple agents can saturate GPU compute resources. Although CPUs offer a viable alternative, existing CPU attention kernels fall short in adapting to multi-tier cache hierarchies and lack effective thread scheduling. To address this, we introduce CACPU, a high-performance attention kernel featuring tiered cache blocking and an in-kernel task dispatching mechanism designed for causal full-sequence prefilling. CACPU achieves up to a 2.01x speedup over PyTorch SDPA, FlexAttention, and SGLang attention kernels. Integrated into SGLang, it reduces Time to First Token (TTFT) of Llama-3.1-8B by up to 1.74x while preserving model accuracy.

Conventional causal attention aggregates values at discrete memory positions. We introduce Continuum Attention (CA), in which neighboring key and value vectors define a continuous, piecewise linear memory trajectory and each query retrieves information by normalized integration over its visible prefix. The segment integrals have closed forms in exponential moments, yielding analytically differentiable attention without numerical quadrature. The resulting operator preserves causal masking, normalized nonnegative effective weights, and the standard Transformer decoder structure. Controlled byte level experiments favor exact continuous integration over matched softmax attention. We additionally evaluate data efficiency, out of domain language modeling, zero-shot multiple choice prediction, robustness, and multilingual generalization. In multilingual evaluation, CA consistently outperforms all baselines across every evaluated language, achieving an average 5.8% reduction in bits per byte (BPB) relative to the strongest baseline.

We present CAST-ASD (Causal Attention over Speakers and Time), a causal joint model for active speaker detection (ASD). Recent ASD models improve performance by jointly processing all faces in a scene, but operate offline, with the full clip available before any frame is labelled. Low-latency causal detectors instead process each face independently. CAST-ASD combines joint face reasoning with fully causal inference. It applies scene-level attention across faces within the current frame and causal temporal attention over each face's history, enabling information sharing across speakers with zero lookahead. Its parameter count does not grow with the number of faces. On MSDWild, where most negative face-frames contain a visible competing speaker, CAST-ASD reduces competing-speaker false alarms by 32% relative to the causal per-face detector it extends, with larger gains as scene occupancy increases. It outperforms all causal baselines on both validation splits while using only one-tenth the inter-speaker parameters of a three-block fixed-roster convolution. Code and models are available at https://github.com/prathamg007/CAST-ASD.

Universal approximation is a necessary qualitative property of learning architectures to benefit from scaling laws. While it is generically verified on a variety of neural architectures and random feature models, it typically involves infinite width limits. In this work, we focus on deep self-attention models and consider instead the `dual' regime, where approximation power is enabled entirely by depth, and featuring strong parameter sharing across layers, motivated by recent models such as the Looped Transformers. More specifically, we ask whether one can find a predefined finite set of parameters, each defining an attention block, such that the resulting finite set of transformations can map any collection of $N$ sequences of $n$ tokens to any other collection of $N$ sequences of $n$ tokens. Crucially, these transformations are fixed independently of the input and output collections: only the order in which the blocks are applied, their signs, and their durations depend on the particular interpolation task. Our main result establishes it for residual softmax attention using only two frozen single-head blocks with Gaussian-initialized projection matrices. The result holds at both continuous and finite depth. We also characterize the restrictions imposed by causal masking and establish corresponding universal interpolation guarantees.

FullAttn repeatedly exposes the complete causal history to every attention head, creating substantial redundant computation and memory traffic even with IO-efficient dense kernels. We introduce CoWA, a structured attention architecture that distributes access to the causal history across KV heads. All heads share near-diagonal and prefix-sink windows, while complementary long-range windows partition the remaining history. Their union provides full causal coverage although each head attends sparsely to distant tokens. This position-defined attention pattern requires no learned router or indexer, is used consistently during training and inference, and aligns with KV-head tensor parallelism. A window-matched ablation at 8K isolates the effect of complementary long-range allocation: CoWA with 100% collective coverage reaches 89.73% accuracy, compared with 89.97% for FullAttn, while duplicated long-range windows perform substantially worse. Across a broader controlled associative-recall comparison with matched token budgets, CoWA closely tracks FullAttn as the context grows, whereas other sparse patterns lose a substantial fraction of the associations. In an attention-operator benchmark at 128K tokens with tensor parallelism, CoWA reduces forward and backward latency during training by 7.4x and 8.6x and decoding latency during inference by 3.0x over FullAttn. Its per-rank peak operator memory matches FullAttn during training and is 7.6x lower during decoding. Across scaling-law training from 0.6B to 14B parameters, CoWA closely tracks FullAttn in perplexity while reducing total training FLOPs. The resulting 14B models and 32B models from separate continued training achieve comparable knowledge, reasoning, and long-context retrieval scores to FullAttn. These results show that full causal coverage can be a collective property of the head ensemble rather than a duplicated property of every head.

Classical constitutive modeling of path-dependent inelastic materials relies on internal state variables whose evolution equations must be postulated based on domain knowledge and calibrated against experimental data. However, in many practical settings, the relevant internal variables are typically not measurable in experiments, and the constitutive response must be inferred entirely from measured strain-stress data without any prior knowledge of the material's internal state. We propose a data-driven constitutive modeling framework based on the concept of a material operator, which treats a deforming material as a functional mapping from its entire strain history to the corresponding stress response. In contrast to traditional autoregressive or recurrent formulations, the model is trained directly on full loading paths as function-to-function mappings, predicting complete stress trajectories in a single parallel forward pass. Temporal path dependence is enforced through a causally masked attention mechanism embedded within the operator, which restricts the model's attention to past material states while preserving computational parallelizability. Spectral convolutions provide discretization-invariant representations in the frequency domain, while causal attention captures highly adaptive, non-local history dependence. Furthermore, sinusoidal activation functions are used to resolve the strong nonlinear transitions inherent in inelastic regimes. The framework is evaluated across multidimensional, rate-independent material models exhibiting complex phenomena, with an emphasis on nonlinear plasticity and ductile damage accumulation. The results demonstrate accurate and robust predictions of irreversible deformation mechanisms while simultaneously achieving resolution invariance and excellent parallel efficiency.

Video carries the temporal structure of the physical world, yet learning representations from it has remained computationally expensive: prevailing self-supervised methods either prevent representation collapse through architectural asymmetries, coupling an exponential-moving-average target encoder, a stop-gradient, and a capacity-limited predictor, or circumvent it by reconstructing masked content in pixel space. We introduce LeVJEPA, the first video encoder trained under LeJEPA's collapse-free objective, which dispenses with both. A single encoder is trained with an invariance loss over global and local views of a clip, regularized by SIGReg, which excludes collapse with a provable guarantee. The architecture reduces to an encoder and a projector, and the objective to a single hyperparameter. This formulation admits two properties. First, the cost of pretraining is governed by the number of tokens the encoder observes; uniform random token dropping renders this number small while simultaneously improving downstream accuracy. At matched epochs on identical data, LeVJEPA matches or surpasses V-JEPA 2 across ViT-S/B/L at 5.6 to 20.8x less pretraining compute, and at matched total FLOPs it exceeds the strongest video baseline by 7.6 points on ImageNet-1K while remaining competitive on motion-centric benchmarks. Second, since no asymmetry between branches is required, the encoder can be trained with block-causal attention at no measurable accuracy cost: temporal ordering becomes a property of the encoder itself. Against a compute-matched DINOv2 trained on frames of the same videos, LeVJEPA approaches the image-pretrained encoder on appearance-centric evaluation while nearly doubling its motion-centric accuracy. These results indicate that, once its computational overhead is removed, video becomes a viable and in several respects preferable substrate for general-purpose visual pretraining.

We present the first causal mechanistic analysis of a tabular foundation model, investigating how TabPFN 2.5's feature wise attention heads distribute computation across layers. Using activation patching, ablation, and attention entropy across two synthetic regression datasets, we find clear temporal specialisation: one head's causal necessity dominates that of the others by 2 to 5 times at peak layer, with its dominant layer shifting across tasks of different complexity, while the remaining heads exhibit symmetric late layer profiles. Attention entropy and patching provide convergent evidence for the computationally active layers of the dominant head. We additionally investigate inference time steerability via contrastive activation steering, which fails to transfer across samples. We attribute this result to TabPFN's in context learning mechanism, which encodes task structure through context dependent attention rather than the stable parametric directions that make steering tractable in language models.

We introduce a new tool, Express, for converting a non-causal attention approximation into a causal approximation with matching approximation guarantees. When combined with the state-of-the-art Thinformer approximation, Express improves upon the best known causal attention guarantees, delivering $\log^{3/2}(n)/s$ approximation error with only $O(s)$ memory and $O(s^2 \log^2(n))$ compression overhead for a sequence of length $n$. We pair these developments with an efficient I/O-aware Triton implementation, demonstrate substantial speedups over FlashAttention 2, and use Express to overcome four resource bottlenecks in the language modeling pipeline: long-context prefill, KV cache compression, long-form memory-constrained decoding, and long-form compute-constrained decoding.

Multivariate time series (MTS) classification is foundational to pervasive computing and financial analysis, yet existing multi-scale paradigms are often constrained by suboptimal representation fidelity. We identify two critical bottlenecks: temporal non-causality in standard encoders that induces temporal confounding in non-stationary dynamics, and the absence of explicit channel saliency mechanisms that allows noise to contaminate the latent space. To address these challenges, we propose the Causal Attention and Spatio-temporal Encoder Network (CASE-NET), an architecture designed for structural manifold pre-conditioning. CASE-NET synergizes a Causal Temporal Encoder, which enforces physical arrow-of-time constraints via masked self-attention and causal convolutions, with an Adaptive Channel Recalibration module functioning as an information bottleneck to suppress detrimental noise. Comprehensive evaluations across six heterogeneous domains demonstrate that CASE-NET establishes new state-of-the-art benchmarks on four tasks, achieving a peak accuracy of 98.6% on the AWR dataset and superior robustness in non-stationary regimes.

Accurate state-of-charge (SOC) estimation is critical for the safe and efficient operation of lithium-ion batteries in battery management systems (BMS). Although data-driven approaches can effectively capture nonlinear battery dynamics, many existing methods rely on long historical input sequences, resulting in high computational cost and introducing padding-induced positional bias at the beginning of drive cycles. To address these limitations, we propose C2L-Net, a novel context-to-latest data-driven framework for realistic online SOC estimation using only a short historical window (20 s). Unlike existing short-receptive-field or long-history models, the proposed framework explicitly separates contextual encoding from latest-measurement updating, enabling both efficient temporal modeling and rapid adaptation to dynamic battery states. The proposed model incorporates a chunk-based feature extraction mechanism that combines Theta Attention Pooling with a Fourier-based Seasonality Basis to capture local temporal patterns while reducing sequence length. A causal context encoder, integrating a gated recurrent unit (GRU) with Causal Cosine Attention, models temporal dependencies without information leakage. Furthermore, a latest-measurement decoder, inspired by recursive filtering, updates the contextual state using the most recent measurement, enhancing responsiveness to dynamic operating conditions. Extensive experiments on a public lithium-ion battery drive-cycle dataset under multiple fixed-temperature conditions demonstrate that the proposed method achieves state-of-the-art or competitive accuracy while significantly improving computational efficiency. In particular, C2L-Net achieves up to 60 times faster inference and requires fewer parameters than recent data-driven baselines, while maintaining robust performance across unseen driving profiles.

Leveraging deep learning for causal discovery in time series remains challenging because existing neural methods predominantly rely on component-wise architectures that fail to capture shared system dynamics or employ decoupled post-hoc graph extraction that risks overfitting to spurious correlations. We propose $\textbf{Mask2Cause}$, an end-to-end framework that recovers the underlying causal graph directly during the forecasting forward pass. Our approach introduces an Inverted Variable Embedding and an Adjacency-Constrained Masked Attention mechanism, trained with homoscedastic or heteroscedastic objectives to capture causal influences in both mean and variance. Empirical results on diverse benchmarks, from synthetic chaotic dynamics to realistic biological simulations, demonstrate state-of-the-art causal discovery with significantly reduced parameter complexity compared to standard baselines. We further show that inferred causal structures can be used to reduce parameter count of forecasting models by more than 70% on average while maintaining predictive accuracy.

Training causal transformers at extreme sequence lengths is bottlenecked by the quadratic time and memory of scaled dot-product attention (SDPA). In this work, we propose Lighthouse Attention, a training-only symmetrical selection-based hierarchical attention algorithm that wraps around ordinary SDPA and can be easily removed towards the end of the training. Our hierarchical selection is also gradient-free, which exempts us from dealing with a complicated and potentially inefficient backward pass kernel. Our contribution is three-fold: (i) A subquadratic hierarchical pre- and post-processing step that does adaptive compression and decompression of the sequence. (ii) A symmetrical compression strategy that pools queries, keys and values at the same time, while preserving left-to-right causality, which greatly improves parallelism. (iii) A two stage training approach which we pre-train for the majority of the time with Lighthouse Attention and recover a full attention model at the end with a short training. We run preliminary small scale LLM pre-training experiments that show the effectiveness of our method compared to full attention training with all other settings matched, where we achieve a faster total training time and lower final loss after the recovery phase. Full code is available at: https://github.com/ighoshsubho/lighthouse-attention

Generative Recommender Systems (GR) increasingly model user behavior as a sequence generation task by interleaving item and action tokens. While effective, this formulation introduces significant structural and computational inefficiencies: it doubles sequence length, incurs quadratic overhead, and relies on implicit attention to recover the causal relationship between an item and its associated action. Furthermore, interleaving heterogeneous tokens forces the Transformer to disentangle semantically incompatible signals, leading to increased attention noise and reduced representation efficiency.In this work, we propose a principled reformulation of generative recommendation that aligns sequence modeling with underlying causal structures and attention theory. We demonstrate that current interleaving mechanisms act as inefficient proxies for similarity-weighted action pooling. To address this, we introduce two novel architectures that eliminate interleaved dependencies to reduce sequence complexity by 50%: Attention-based Late Fusion for Actions (AttnLFA) and Attention-based Mixed Value Pooling (AttnMVP). These models explicitly encode the $i_n \rightarrow a_n$ causal dependency while preserving the expressive power of Transformer-based sequence modeling.We evaluate our framework on large-scale product recommendation data from a major social network. Experimental results show that AttnLFA and AttnMVP consistently outperform interleaved baselines, achieving evaluation loss improvements of 0.29% and 0.80%, and significant gains in Normalized Entropy (NE). Crucially, these performance gains are accompanied by training time reductions of 23% and 12%, respectively. Our findings suggest that explicitly modeling item-action causality provides a superior design paradigm for scalable and efficient generative ranking.

The quadratic computational complexity of self-attention remains a fundamental bottleneck for scaling Large Language Models (LLMs) to long contexts, particularly during the pre-filling phase. In this paper, we rethink the causal attention mechanism from the perspective of information flow. Due to causal constraints, tokens at initial positions participate in the aggregation of every subsequent token. However, existing sparse methods typically apply a uniform top-k selection across all token positions within a layer, ignoring the cumulative dependency of token information inherent in causal architectures. To address this, we propose Stem, a novel, plug-and-play sparsity module aligned with information flow. First, Stem employs the Token Position-Decay strategy, applying position-dependent top-k within each layer to retain initial tokens for recursive dependencies. Second, to preserve information-rich tokens, Stem utilizes the Output-Aware Metric. It prioritizes high-impact tokens based on approximate output magnitude. Extensive evaluations demonstrate that Stem achieves superior accuracy with reduced computation and pre-filling latency.

Video polyp segmentation (VPS) is an important task in computer-aided colonoscopy, as it helps doctors accurately locate and track polyps during examinations. However, VPS remains challenging because polyps often look similar to surrounding mucosa, leading to weak semantic discrimination. In addition, large changes in polyp position and scale across video frames make stable and accurate segmentation difficult. To address these challenges, we propose a robust VPS framework named CMSA-Net. The proposed network introduces a Causal Multi-scale Aggregation (CMA) module to effectively gather semantic information from multiple historical frames at different scales. By using causal attention, CMA ensures that temporal feature propagation follows strict time order, which helps reduce noise and improve feature reliability. Furthermore, we design a Dynamic Multi-source Reference (DMR) strategy that adaptively selects informative and reliable reference frames based on semantic separability and prediction confidence. This strategy provides strong multi-frame guidance while keeping the model efficient for real-time inference. Extensive experiments on the SUN-SEG dataset demonstrate that CMSA-Net achieves state-of-the-art performance, offering a favorable balance between segmentation accuracy and real-time clinical applicability.