Attention Mechanisms

Vision-Language Models (VLMs) achieve strong cross-modal performance, yet recent evidence suggests they over-rely on textual descriptions while under-utilizing visual evidence -- a phenomenon termed ``text shortcut learning.'' We propose an adversarial evaluation framework that quantifies this cross-modal dependency by measuring accuracy degradation (Drop) when semantically conflicting text is paired with unchanged images. Four adversarial strategies -- shape_swap, color_swap, position_swap, and random_text -- are applied to a controlled geometric-shapes dataset ($n{=}1{,}000$). We compare three configurations: Baseline CLIP (ViT-B/32), LoRA fine-tuning, and LoRA Optimized (integrating Hard Negative Mining, Label Smoothing, layer-wise learning rates, Cosine Restarts, curriculum learning, and data augmentation). The optimized model reduces average Drop from 27.5% to 9.8% (64.4% relative improvement, $p{<}0.001$) while maintaining 97% normal accuracy. Attention visualization and embedding-space analysis confirm that the optimized model attends more to visual features and achieves tighter cross-modal alignment.

We present Open-TQ-Metal, the first implementation of fused compressed-domain attention on Apple Silicon, enabling 128K-context inference for Llama 3.1 70B on a single 64GB consumer Mac -- a configuration impossible with all existing inference frameworks. Open-TQ-Metal quantizes the KV cache to int4 on the fly and computes attention directly on the compressed representation via custom Metal compute shaders, eliminating all intermediate dequantization matrices. Across 330 experiments spanning two model families (Gemma 4 31B and Llama 3.1 70B), the fused sdpa_int4 kernel achieves 48x attention speedup at 128K context over the dequantize-then-attend baseline, reduces KV cache memory from 40 GB to 12.5 GB (3.2x compression), and maintains identical top-1 token predictions to FP16 inference. We further provide the first cross-architecture analysis of KV cache quantization methods, revealing that the attention scale factor -- not model size -- determines whether angular quantization schemes like PolarQuant succeed or fail, with Gemma 4's attn_scale=1.0 amplifying directional error 25-100x more than Llama's standard 1/sqrt(d) scaling.

In long-context decoding for LLMs and LMMs, attention becomes increasingly memory-bound because each decoding step must load a large amount of KV-cache data from GPU memory. Existing acceleration strategies often trade efficiency for accuracy by relying on heuristic pruning that may discard useful information. At a deeper level, they also tend to indiscriminately preserve all high-scoring tokens, treat early tokens as indispensable anchors, or rely on heuristic head routing, reflecting an insufficient mechanistic understanding of the attention sink phenomenon. In this paper, we show that the attention sink phenomenon corresponds to a stable, reachable, and error-controllable fixed point constructed during training. Based on this insight, we propose SinkRouter, a training-free selective routing framework that detects the sink signal and skips computations that would otherwise produce near-zero output. To translate this mechanism into real-world acceleration, we develop a hardware-aware Triton kernel with block-level branching and Split-K parallelism. We conduct extensive evaluations on a diverse suite of long-context benchmarks, including LongBench, InfiniteBench, CVBench, MileBench, and MMVP, using both text-only and multimodal backbones such as Llama-3.1-8B, Llama-3.1-70B, Yi-9B-200K, LLaVA-1.5-7B, and LLaVA-1.5-13B. Across these settings, SinkRouter consistently improves decoding efficiency while maintaining competitive accuracy, and reaches 2.03x speedup with a 512K context.

Large language models (LLMs) increasingly rely on chain-of-thought (CoT) reasoning to solve complex tasks. Yet ensuring that the reasoning trace both contributes to and faithfully reflects the processes underlying the model's final answer, rather than merely accompanying it, remains challenging. We introduce AtManRL, a method that leverages differentiable attention manipulation to learn more faithful reasoning through reinforcement learning. By training an additive attention mask that identifies tokens in the CoT crucial for producing correct answers, we derive a saliency reward signal that encourages the model to generate reasoning traces that genuinely influence its final predictions. We integrate this saliency reward with outcome-based rewards within the GRPO framework to jointly optimize for correctness and interpretability. Experiments on GSM8K and MMLU with Llama-3.2-3B-Instruct demonstrate that our approach can identify influential reasoning tokens and enable training more transparent reasoning models.

The significant computational demands of large language models have increased interest in distilling reasoning abilities into smaller models via Chain-of-Thought (CoT) distillation. Current CoT distillation methods mainly focus on transferring teacher-generated rationales for complex reasoning to student models. However, they do not adequately explore teachers' dynamic attention toward critical information during reasoning. We find that language models exhibit progressive attention shifts towards key information during reasoning, which implies essential clues for drawing conclusions. Building on this observation and analysis, we introduce a novel CoT distillation framework that transfers the teacher's stepwise attention on key information to the student model. This establishes structured guidance for the student's progressive concentration on key information during reasoning. More importantly, we develop a Mixture of Layers module enabling dynamic alignment that adapts to different layers between the teacher and student. Our method achieves consistent performance improvements across multiple mathematical and commonsense reasoning datasets. To our knowledge, it is the first method to leverage stepwise attention within CoT distillation to improve small model reasoning.

Current large language models reason in isolation. Although it is common to sample multiple reasoning paths in parallel, these trajectories do not interact, and often fail in the same redundant ways. We introduce LACE, a framework that transforms reasoning from a collection of independent trials into a coordinated, parallel process. By repurposing the model architecture to enable cross-thread attention, LACE allows concurrent reasoning paths to share intermediate insights and correct one another during inference. A central challenge is the absence of natural training data that exhibits such collaborative behavior. We address this gap with a synthetic data pipeline that explicitly teaches models to communicate and error-correct across threads. Experiments show that this unified exploration substantially outperforms standard parallel search, improving reasoning accuracy by over 7 points. Our results suggest that large language models can be more effective when parallel reasoning paths are allowed to interact.

Sparse attention has been proposed as a way to alleviate the quadratic cost of transformers, a central bottleneck in long-context training. A promising line of work is $α$-entmax attention, a differentiable sparse alternative to softmax that enables input-dependent sparsity yet has lagged behind softmax due to the computational overhead necessary to compute the normalizer $τ$. In this paper, we introduce AdaSplash-2, which addresses this limitation through a novel histogram-based initialization that reduces the number of iterations needed to compute $τ$ to typically 1--2. The key idea is to compute a coarse histogram of attention scores on the fly and store it in on-chip SRAM, yielding a more accurate initialization that enables fast forward and backward computation. Combined with a sparsity-aware GPU implementation that skips zero blocks with low overhead, AdaSplash-2 matches or improves per-step training time relative to FlashAttention-2 when block sparsity is moderate-to-high (e.g., $>$60%), which often occurs at long-context lengths. On downstream tasks, models trained with our efficient $α$-entmax attention match softmax baselines at short-context lengths and achieve substantial gains in long-context settings.

Token pruning methods for Vision Transformers (ViTs) promise quadratic reductions in attention FLOPs by dropping uninformative patches. Yet standard variable-length attention APIs -- including FlashAttention-2's varlen and PyTorch's NestedTensor SDPA -- fail to translate these savings into proportional wall-clock gains at the short post-pruning sequence lengths typical of ViTs ($\leq$197 tokens). We identify a dispatch-overhead bottleneck: at these lengths, host-side kernel dispatch consumes ${\sim}$50,$μ$s regardless of workload, exceeding the actual GPU compute time at moderate-to-high pruning rates. We present a lightweight bidirectional Triton attention kernel whose dispatch floor is ${\sim}$24,$μ$s -- roughly 2.17$\times$ lower than FlashAttention-2 varlen -- allowing pruning savings to become visible in wall-clock time. Integrated into a complete pack-attend-unpack pipeline and evaluated on an NVIDIA RTX 4000 Ada Generation GPU, our system achieves 1.88$\times$ end-to-end throughput over padded PyTorch SDPA at standard 224$\times$224 inputs, scaling to 2.51$\times$ at 384$\times$384. Against FlashAttention-2 varlen -- the strongest baseline -- our kernel delivers 9-12% higher throughput at serving batch sizes (BS=1-4), and 2.17$\times$ lower kernel latency at 80% token pruning. Numerical correctness is verified with max absolute logit difference $<$0.004 and bit-exact top-1 predictions.

When do transformers commit to a decision, and what prevents them from correcting it? We introduce \textbf{prolepsis}: a transformer commits early, task-specific attention heads sustain the commitment, and no layer corrects it. Replicating \citeauthor{lindsey2025biology}'s (\citeyear{lindsey2025biology}) planning-site finding on open models (Gemma2 2B, Llama3.2 1B), we ask five questions. (Q1)~Planning is invisible to six residual-stream methods; CLTs are necessary. (Q2)~The planning-site spike replicates with identical geometry. (Q3)~Specific attention heads route the decision to the output, filling a gap flagged as invisible to attribution graphs. (Q4)~Search requires ${\leq}16$ layers; commitment requires more. (Q5)~Factual recall shows the same motif at a different network depth, with zero overlap between recurring planning heads and the factual top-10. Prolepsis is architectural: the template is shared, the routing substrates differ. All experiments run on a single consumer GPU (16,GB VRAM).

To quantify the geometric expressivity of transformers, we introduce a tropical geometry framework to characterize their exact spatial partitioning capabilities. By modeling self-attention as a vector-valued tropical rational map, we prove it evaluates exactly to a Power Voronoi Diagram in the zero-temperature limit. Building on this equivalence, we establish a combinatorial rationale for Multi-Head Self-Attention (MHSA): via the Minkowski sum of Newton polytopes, multi-head aggregation expands the polyhedral complexity to $\mathcal{O}(N^H)$, overcoming the $\mathcal{O}(N)$ bottleneck of single heads. Extending this to deep architectures, we derive the first tight asymptotic bounds on the number of linear regions in transformers ($Θ(N^{d_{\text{model}}L})$), demonstrating a combinatorial explosion driven intrinsically by sequence length $N$, ambient embedding dimension $d_{\text{model}}$, and network depth $L$. Importantly, we guarantee that this idealized polyhedral skeleton is geometrically stable: finite-temperature soft attention preserves these topological partitions via exponentially tight differential approximation bounds.

Transformers commonly exhibit an attention sink: disproportionately high attention to the first position. We study this behavior in GPT-2-style models with learned query biases and absolute positional embeddings. Combining structural analysis with causal interventions, validated across natural-language, mathematical, and code inputs, we find that the sink arises from the interaction among (i) a learned query bias, (ii) the first-layer MLP transformation of the positional encoding, and (iii) structure in the key projection. Crucially, each component we identify is individually dispensable: architectures omitting each of them robustly exhibit sinks. This indicates that attention sinks may arise through distinct circuits across architectures. These findings inform mitigation of sinks, and motivate broader investigation into why sinks emerge.

Multiplicative gating is widely used in neural architectures and has recently been applied to attention layers to improve performance and training stability in large language models. Despite the success of gated attention, the mathematical implications of gated attention mechanisms remain poorly understood. We study attention through the geometry of its representations by modeling outputs as mean parameters of Gaussian distributions and analyzing the induced Fisher--Rao geometry. We show that ungated attention operator is restricted to intrinsically flat statistical manifolds due to its affine structure, while multiplicative gating enables non-flat geometries, including positively curved manifolds that are unattainable in the ungated setting. These results establish a geometric expressivity gap between ungated and gated attention. Empirically, we show that gated models exhibit higher representation curvature and improved performance on tasks requiring nonlinear decision boundaries whereas they provide no consistent advantage on tasks with linear decision boundaries. Furthermore, we identify a structured regime in which curvature accumulates under composition, yielding a systematic depth amplification effect.

Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works--specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit. Freezing all attention scores during steering drops performance by only 8.83% across three model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 85-96% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.

The self-attention mechanism is central to the success of Transformer architectures. However, standard row-stochastic attention has been shown to suffer from significant signal degradation across layers. In particular, it can induce rank collapse, resulting in increasingly uniform token representations, as well as entropy collapse, characterized by highly concentrated attention distributions. Recent work has highlighted the benefits of doubly stochastic attention as a form of entropy regularization, promoting a more balanced attention distribution and leading to improved empirical performance. In this paper, we study rank collapse across network depth and show that doubly stochastic attention matrices normalized with Sinkhorn algorithm preserve rank more effectively than standard softmax row-stochastic ones. As previously shown for softmax, skip connections are crucial to mitigate rank collapse. We empirically validate this phenomenon on both sentiment analysis and image classification tasks. Moreover, we derive a theoretical bound for the pure self-attention rank decay when using Sinkhorn normalization and find that rank decays to one doubly exponentially with depth, a phenomenon that has already been shown for softmax.

The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8$\times$A800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to $2.8\times$ and $2.0\times$ in the prefill and decode stages.

Scaled dot-product attention forms its queries and keys as independent linear projections, so the two never interact before the dot product that scores them. We study coupled query-key dynamics, a pre-scoring transformation that evolves each token's query and key jointly through a shared invertible coupling before standard scoring. We realize it as an alternating affine map in the style of real non-volume-preserving flows: the coupling is the identity at initialization, adds a small fraction of parameters per head, and leaves the softmax and surrounding architecture unchanged. We place coupling on top of existing attention methods rather than replacing them, and ask whether that composition helps. On WikiText-103, adding coupling to Differential Attention improves on it at both 150M and 455M parameters. At 455M the gain is significant at sequence length 512 (p=0.003, six seeds), survives a Bonferroni correction and replicates on a held-out test split; it also holds across rotary-embedding training lengths 512, 1024, and 2048. The same additive direction appears when coupling is added to query-key normalization (significant at 150M) and Multi-Token Attention (directional). Matched controls attribute the gain to the joint pre-scoring coupling rather than to added capacity, and show that removing the invertibility guarantee preserves the 455M gain yet is far worse than the base method at 150M, so invertibility is what makes the coupling reliable across scales. On its own, coupling lowers perplexity at 60M and 150M (one-sided Welch tests, p<0.05) but the gain narrows with scale and is not significant at 455M. We relate the construction to the expressivity of coupling flows, use an associative-recall study to map where coupling helps and where it degrades sharp retrieval, and conclude that coupling is most useful in composition with a scoring-stage method rather than as a standalone change.

We call query--key relevance absolute when its values lie on a fixed bounded scale, depend on neither competing keys nor sequence length, require no sequence-length-dependent calibration, and can all be zero. To realize this notion, we introduce screening, whose explicit threshold transforms bounded query--key similarities into relevance values, enabling exact rejection, empty selection, and direct inspection on a common scale. In a controlled comparison of 12 attention mechanisms on a matched Transformer backbone, only screening maintains both low long-context perplexity and robust retrieval beyond the training context; notably, it does so without inference-time scaling. Building on screening, we introduce Multiscreen, a language-model architecture composed of parallel gated screening tiles. Multiscreen retains these long-context gains while achieving greater parameter efficiency, stronger general zero-shot downstream performance, lower training cost at larger scales, and lower model-side time to first token than Transformer baselines. We further develop a normalization design that keeps Multiscreen training stable even at a learning rate of $1$ and show that an adapted version likewise stabilizes Transformer at the same learning rate.

Large Vision Language Models show impressive performance across image and video understanding tasks, yet their computational cost grows rapidly with the number of visual tokens. Existing token pruning methods mitigate this issue through empirical approaches while overlooking the internal mechanism of attention. In this paper, we propose a novel training free token pruning framework grounded in the dual form perspective of attention. We reformulate attention as an implicit linear layer whose weight matrix is the sum of rank 1 outer products, each generated by a single token's key value pair. Token pruning thus reduces to selecting an optimal subset of these rank 1 updates that best approximates the original dual weight matrix. Extending this perspective to standard softmax attention in LVLMs, we derive a novel metric quantifying both a token's information magnitude and information duplication. To efficiently select the subset with the proposed metric, we introduce Progressive Chunked Maximal Marginal Relevance. Extensive experiments demonstrate that our method achieves a better trade off between performance and efficiency, while providing another perspective on existing pruning approaches.

Attention sinks are defined as tokens that attract disproportionate attention. While these have been studied in single modality transformers, their cross-modal impact in Large Vision-Language Models (LVLM) remains largely unexplored: are they redundant artifacts or essential global priors? This paper first categorizes visual sinks into two distinct categories: ViT-emerged sinks (V-sinks), which propagate from the vision encoder, and LLM-emerged sinks (L-sinks), which arise within deep LLM layers. Based on the new definition, our analysis reveals a fundamental performance trade-off: while sinks effectively encode global scene-level priors, their dominance can suppress the fine-grained visual evidence required for local perception. Furthermore, we identify specific functional layers where modulating these sinks most significantly impacts downstream performance. To leverage these insights, we propose Layer-wise Sink Gating (LSG), a lightweight, plug-and-play module that dynamically scales the attention contributions of V-sink and the rest visual tokens. LSG is trained via standard next-token prediction, requiring no task-specific supervision while keeping the LVLM backbone frozen. In most layers, LSG yields improvements on representative multimodal benchmarks, effectively balancing global reasoning and precise local evidence.

Hallucinations in large ASR models present a critical safety risk. In this work, we propose the \textit{Spectral Sensitivity Theorem}, which predicts a phase transition in deep networks from a dispersive regime (signal decay) to an attractor regime (rank-1 collapse) governed by layer-wise gain and alignment. We validate this theory by analyzing the eigenspectra of activation graphs in Whisper models (Tiny to Large-v3-Turbo) under adversarial stress. Our results confirm the theoretical prediction: intermediate models exhibit \textit{Structural Disintegration} (Regime I), characterized by a $13.4\%$ collapse in Cross-Attention rank. Conversely, large models enter a \textit{Compression-Seeking Attractor} state (Regime II), where Self-Attention actively compresses rank ($-2.34\%$) and hardens the spectral slope, decoupling the model from acoustic evidence.

Token-level sparse attention mechanisms, exemplified by DeepSeek Sparse Attention (DSA), achieve fine-grained key selection by scoring every historical key for each query through a lightweight indexer, then computing attention only on the selected subset. While the downstream sparse attention itself scales favorably, the indexer must still scan the entire prefix for every query, introducing an per-layer bottleneck that grows prohibitively with context length. We propose HISA (Hierarchical Indexed Sparse Attention), a plug-and-play replacement for the indexer that rewrites the search path from a flat token scan into a two-stage hierarchical procedure: (1) a block-level coarse filtering stage that scores pooled block representations to discard irrelevant regions, followed by (2) a token-level refinement stage that applies the original indexer exclusively within the retained candidate blocks. HISA preserves the identical token-level top-sparse pattern consumed by the downstream Sparse MLA operator and requires no additional training. On kernel-level benchmarks, HISA achieves up to speedup at 64K context. On Needle-in-a-Haystack and LongBench, we directly replace the indexer in DeepSeek-V3.2 and GLM-5 with our HISA indexer, without any finetuning. HISA closely matches the original DSA in quality, while substantially outperforming block-sparse baselines.

Groundwater represents a key element of the water cycle, yet it exhibits complex and context-dependent relationships that make its modeling challenging. Theory-based models have been the cornerstone of scientific understanding. However, their computational cost, simplifying assumptions, and calibration requirements limit their use. In recent years, data-driven models have emerged as powerful alternatives. In particular, deep learning has proven to be a promising approach for its design flexibility and ability to learn complex relationships directly from the data without requiring extensive domain information. We proposed an attention-based pure deep learning model, named STAINet, to predict weekly groundwater levels in Piedmont (Italy), leveraging both irregular groundwater time series and weather image sequences. To enhance the model's trustworthiness and generalization ability, we merged the theory and data-driven approaches by considering physics-guided strategies to inject the groundwater flow equation into the model. Firstly, we restructured the tail of the architecture to predict the three terms of the governing equation, named the autoregressive, diffusion, and residual components - we thus obtained the PSTAINet-IB. Then, we further injected physics priors by adding loss terms related to the estimated equation components, obtaining the PSTAINet-ILB model. Lastly, we developed the PSTAINet-ILRB by imposing a loss term specific to the residual component, which forces the groundwater recharge to occur within the groundwater body recharge zone, which is identified by domain experts. The models were evaluated both by feeding true lagged values as input and by iterating their own predictions (rollouts) over the whole test set. The PSTAINet-ILB model performed the best, achieving remarkable test performance, and generating equation components in line with domain experts' expectations.

Language models struggle to generalize beyond pretraining context lengths, limiting long-horizon reasoning and retrieval. Continued pretraining on long-context data can help but is expensive due to the quadratic scaling of Attention. We observe that most tokens do not require (Global) Attention over the entire sequence and can rely on local context. Based on this, we propose L2A (Learning To Attend), a layer that enables conditional (token-wise) long-range memory access by deciding when to invoke global attention. We evaluate L2A on Qwen 2.5 and Qwen 3 models, extending their effective context length from 32K to 128K tokens. L2A matches the performance of standard long-context training to within 3% while skipping Global Attention for $\sim$80% of tokens, outperforming prior baselines. We also design custom Triton kernels to efficiently implement this token-wise conditional Attention on GPUs, achieving up to $\sim$2$\times$ improvements in training throughput and time-to-first-token over FlashAttention. Moreover, L2A enables post-training pruning of highly sparse Global Attention layers, reducing KV cache memory by up to 50% with negligible performance loss. Our code is released under Apache 2.0 at https://github.com/awslabs/hybrid-model-factory/tree/main/examples/research/L2A.

Recent decoder-only autoregressive text-to-speech (AR-TTS) models produce high-fidelity speech, but their memory and compute costs scale quadratically with sequence length due to full self-attention. In this paper, we propose WAND, Windowed Attention and Knowledge Distillation, a framework that adapts pretrained AR-TTS models to operate with constant computational and memory complexity. WAND separates the attention mechanism into two: persistent global attention over conditioning tokens and local sliding-window attention over generated tokens. To stabilize fine-tuning, we employ a curriculum learning strategy that progressively tightens the attention window. We further utilize knowledge distillation from a full-attention teacher to recover high-fidelity synthesis quality with high data efficiency. Evaluated on three modern AR-TTS models, WAND preserves the original quality while achieving up to 66.2% KV cache memory reduction and length-invariant, near-constant per-step latency.

Retrieval-augmented generation (RAG) enhances zero-shot time series (TS) forecasting by leveraging external knowledge bases, yet existing approaches overlook input-level relevance when fusing retrieved samples with the query. We argue that not all retrievals are equally useful, and irrelevant ones can degrade performance. To this end, we propose Cross-RAG, a zero-shot RAG-based forecasting framework that selectively attends to query-relevant retrieved samples via query--retrieval cross-attention. By modeling input-level relevance between the query and retrieved samples, Cross-RAG jointly incorporates three sources of information: 1) the query itself, 2) the retrieved samples, and 3) their relational interactions. In particular, this input-aware design enables Cross-RAG to remain stable as the number of retrieved samples $k$ grows, whereas prior methods without cross-attention require careful $k$ tuning to avoid degradation from irrelevant retrievals. Extensive experiments demonstrate that Cross-RAG consistently improves zero-shot forecasting performance across multiple TSFM backbones and various RAG methods, with additional analyses confirming its effectiveness across various retrieval scenarios. Code is available at https://github.com/seunghan96/cross-rag/.

Sequence modeling tasks can involve intrinsic higher-order dependencies, while standard self-attention assigns scores to token pairs and does not explicitly parameterize such interactions. We introduce Higher-Order Modular Attention (HOMA), which fuses pairwise attention with an explicit triadic attention pathway made tractable through overlapping blocks, local windows, and a low-rank projection. We compare HOMA with matched pairwise and purely triadic baselines on controlled PARITY and MATCH3 tasks, as well as TAPE benchmarks. HOMA is competitive with or outperforms the baselines, with its clearest advantages when the underlying dependencies extend beyond the explicitly modeled triadic order. These advantages are accompanied in several settings by faster convergence and improved parameter efficiency, with the learned nonlinear fusion providing an effective mechanism for combining the pairwise and triadic representations. Overall, our results provide empirical evidence that HOMA is an effective attention design when task structure extends beyond pairwise interactions.

Recent token merging techniques for Vision Transformers (ViTs) provide substantial speedups by reducing the number of tokens processed by self-attention, often without retraining. However, their direct application to the Segment Anything Model (SAM) family is nontrivial: SAM's image encoder mixes windowed and global attention, and its mask decoder relies on dense, prompt-conditioned features for precise boundary prediction. We systematically evaluate representative token-merging methods on SAM and Medical SAM in a strict off-the-shelf setting, and find that existing destination-selection heuristics can erode boundaries and leak prompt information as merge rates increase. We propose \textbf{StructSAM}, a resolution-preserving merge-unmerge framework tailored to SAM. StructSAM computes a lightweight token-energy score from first-order feature gradients, uses grid-based flatness screening to protect boundary and prompt regions, and merges tokens within flat areas toward low-energy destinations with explicit token recovery. We further provide a spectral graph coarsening view showing that score-guided merging yields bounded Laplacian spectral distortion compared to random or window-restricted baselines. Across eight natural and medical benchmarks, StructSAM reduces encoder FLOPs by 25-30% (up to 40%+ with prompt-aware merging) with minor drops in mIoU/Dice, consistently outperforming ToMe, PiToMe, ToMeSD, VidToMe, and ALGM at the same compute.

How natural speech is represented in the brain constitutes a major challenge for cognitive neuroscience. Reconstructing the speech envelope and Mel spectrogram from EEG and MEG provides a time-resolved way to study its temporal and spectral structure. Speech-related neural activity spans sensors and temporal scales; extracting these representations while adapting the use of context to each acoustic target is a central problem in speech reconstruction. We propose SHINE, a Sequential Hierarchical Integration Network for EEG and MEG. A residual sensor adapter unifies input dimensions, intermediate dilated-block states retain temporal depth, and a target- and time-dependent gate fuses local hierarchical and attention-enhanced context predictions. Across two EEG and two MEG datasets, SHINE has the highest mean envelope and mean-Mel Pearson correlations among nine local baseline implementations on all eight dataset-metric combinations. SHINE also placed second in the speech-detection Extended Track of the NeurIPS 2025 PNPL Competition. Code will be released at https://github.com/xuxiran/SHINE.

Recent work has identified a subset of attention heads in Transformer as retrieval heads, which are responsible for retrieving information from the context. In this work, we first investigate retrieval heads in multilingual contexts. In multilingual language models, we find that retrieval heads are often shared across multiple languages. Expanding the study to cross-lingual setting, we identify Retrieval-Transition heads(RTH), which govern the transition to specific target-language output. Our experiments reveal that RTHs are distinct from retrieval heads and more vital for Chain-of-Thought reasoning in multilingual LLMs. Across four multilingual benchmarks (MMLU-ProX, MGSM, MLQA, and XQuaD) and two model families (Qwen-2.5 and Llama-3.1), we demonstrate that masking RTH induces bigger performance drop than masking Retrieval Heads (RH). Our work advances understanding of multilingual LMs by isolating the attention heads responsible for mapping to target languages.

This paper studies simple transformers trained on a high-order Markov chain, where the model must incorporate information from multiple past positions, each with different statistical importance. We show that transformers learn the task incrementally, with each stage corresponding to learning how to copy information from a subset of positions via a sparse attention pattern. Notably, the learning dynamics transition from a competitive phase, where all heads focus on the statistically most important positions, to a cooperative phase, where different heads specialize in different patterns. We model these dynamics with simplified differential equations and prove stage-wise convergence of the resulting system. Functionally, these stages correspond to a sequence of increasingly expressive misspecified models, with the full model class reached only at the end. Overall, we give a theoretical account of how structured attention patterns and head specialization emerge in stages without an explicit curriculum, with implications for generalization in sequential tasks.