Transformers

Choosing a deep learning architecture for label-free single-cell classification remains an open question, with microscopy benchmarks reporting conflicting conclusions about CNNs versus transformers. We present a controlled benchmark on LIVECell phase-contrast microscopy data using source-image-disjoint train/validation/test splits to prevent parent-image leakage and matched optimisation, augmentation, and evaluation protocols across EfficientNet, Vision Transformer (ViT), and EVA-02 models. This allows the effects of architecture, pretraining, fine-tuning, tokenisation, and distillation to be disentangled. We find that the previously reported CNN advantage is largely explained by pretraining rather than architecture: the smallest pretrained model outperforms the strongest model trained from scratch despite far fewer parameters. Pretraining improves macro-F1 by 3-4 points, while the gap between the best pretrained CNN and transformer is below 0.5 points. Architectural choices nevertheless matter: ViT-S/8 outperforms ViT-S/16 and matches the four-times-larger ViT-B/16 at a quarter of the parameters, showing that finer tokenisation benefits small cell crops. Conversely, layer-wise learning-rate decay, central to the EVA-02 fine-tuning recipe, degrades performance, highlighting that transfer heuristics from natural-image recognition may not generalise to microscopy. Finally, knowledge distillation substantially improves the deployment frontier: compact EfficientNet-B0 students distilled from teacher councils outperform every individually trained backbone, including the EfficientNet-B5 and EVA-02 teachers. Overall, our results show that rigorous control of pretraining and evaluation is essential for interpreting biomedical architecture benchmarks, while distillation may be a more effective route to practical single-cell classification than architecture choice alone.

The KV cache is a primary bottleneck for Transformer decoding: its memory footprint and cache-read traffic grow with sequence length. Grouped-query attention (GQA) reduces this cost by sharing key-value heads, but still stores both a key and a value at every step. We introduce Grouped Value Attention (GVA), which stores grouped values and reconstructs content keys with a learned linear map. At inference, the map can be absorbed into the query, eliminating the need to materialize content keys in the intended decode path. A small shared decoupled RoPE channel retains positional information through a separately cached positional key. For the configurations studied, this representation reduces persistent cache scalars by approximately 45-47% relative to matched GQA. At the 350M-parameter scale with 30B FineWeb-Edu tokens, the 16-dimensional positional variant reaches 44.18 average accuracy across five tasks, compared with 44.36 for GQA and 43.88 for MLA. These results demonstrate near-GQA benchmark accuracy with a more compact cache representation. To translate this compact representation into faster autoregressive inference, we have developed custom decoding kernels and are currently evaluating their end-to-end inference performance with an open-source release planned soon.

Recent advancements in transformer length generalization theory enable us to reliably predict when a transformer can learn to solve a task. In particular, the C-RASP hypothesis (a formalized version of the so-called RASP-l conjecture) posits that transformers length-generalize on a task if and only if a solution is expressible in the C-RASP language. While this hypothesis has strong empirical validation, theoretical problems arise from the fact that no computable length generalization bounds exist for C-RASP, alongside the discovery of seemingly contradictory experiments. To address these problems, we refine the C-RASP hypothesis utilizing the recently-proposed fragments C-RASP+ and C-RASP1. These fragments have computable length generalization bounds, though in the worst case requiring an extremely large (double exponential) sample size. It is an open question whether these sample size bounds are tight. In this paper, we resolve this open question by providing an exponentially tighter bound. In doing so, we show a polynomial length generalization bound for transformers if we adopt compressed strings, via a novel connection to power words. As an application, we show how this yields a fine-grained analysis of the C-RASP conjecture that resolves contradicting experimental evidence against it.

Recent state-space models (SSMs) such as Mamba achieve language modeling performance comparable to transformers despite relying on fundamentally different architectures. This raises an important question: how do these structural differences influence the geometry and functional nature of their internal representations? We study this question through a multi-scale analysis of representations in transformers, SSMs, and hybrid architecture. First, we find that SSMs distribute their representational information evenly across all dimensions, whereas transformer representations are heavily dominated by a single principal direction. By evaluating hybrid architectures, we observe that the representation space becomes increasingly skewed toward a single dominant direction after each attention layer. Next, we explore how the different geometric spread of representations impacts representational capacity through compressibility. Surprisingly, we find that despite their contrasting geometric structures, both architectures exhibit tightly matched effective capacities. We further investigate whether this skewed geometry affects how concepts are encoded. Using rank-constrained probes, we demonstrate that both architectures encode concepts in subspaces of surprisingly similar dimensionality. Furthermore, we demonstrate that the transformers' dominant principal direction does not inherently encode more conceptual information. Finally, we zoom in and examine the alignment between manifolds, either by analyzing representations of specific topics or by looking at the nearest neighborhoods of tokens, and find that they are highly aligned. Ultimately, our analysis suggests that while transformers and SSMs induce different usage of latent space, they display a striking functional convergence at the level of local semantic manifolds.

Dimensional attention in learning is often implemented as a globally shared attention vector, where each stimulus dimension corresponds to a single scalar. These scalars are learned by models through gradient-descent on error, where predictive features acquire more salience. We show that under multi-outcome learning, where models predict more than one outcome, this shared vector becomes unstable; it collapses to its bounds and prevents the models from learning meaningful attentional tunings for learning and generalization. We address this by introducing an outcome-indexed attentional matrix that converts globally shared attentional tuning into an outcome-indexed representation. We present an analysis of the unstable shared vectors and derive the conditions under which it holds. Empirically, three synthetic experiments benchmark the proposed attention matrices and show that they converge to meaningful representations, something shared attention vectors fail to do. These results suggest that outcome-indexed attentional matrices are a general fix for gradient-based attentional processes, which improves models of learning under multi-outcome conditions.

Understanding why individuals respond differently to psilocybin requires modeling the drug's transcriptional perturbation signature at the cell-type level. I present a Transformer-based delta expression encoder that learns to classify differential gene expression status - upregulated, downregulated, or neutral - from single-nucleus RNA-sequencing data, without supervision from pathway annotations or prior biological knowledge. The model is trained on pseudobulk profiles from 623 examples spanning 18 cell types, 2 drug conditions, and 6 timepoints derived from the Liao et al. 2025 dataset, and achieves 69.4% weighted classification accuracy. Three principal findings are reported, alongside one direct test of a published hypothesis that returned a result inconsistent with that hypothesis. First, per-cell-type classification accuracy ranges from 28.3% (L2/3 IT, a primary HTR2A-expressing psilocybin target) to 99.6% (endothelial cells), consistent with known psilocybin response biology. Second, psilocybin-induced transcriptional downregulation is significantly more stereotyped across individuals than upregulation (Mann-Whitney U=18615.0, p<0.0001), a novel finding with a cortical depth gradient across excitatory subtypes. Third, attention-guided gene co-regulation analysis recovers drug-specific modules without pathway supervision. Separately, a direct test of whether baseline HTR2A expression predicts drug-response separability across cell types found a significant negative correlation (Spearman r = -0.7088, p = 0.0021), the opposite of what a simple HTR2A-gating account would predict.

Key-value (KV) cache eviction is essential for scaling long-context inference in Large Language Models. However, existing policies predominantly rely on empirical heuristics, lacking a rigorous characterization of token utility under the inherently nonlinear softmax attention mechanism. In this work, we rethink KV cache eviction through the lens of local information geometry, modeling the attention process as a nonlinear Gaussian communication channel. By performing a first-order Taylor expansion of the attention mapping, we derive the Jacobian Information Capacity, a novel objective that explicitly captures query relevance, softmax sensitivity, and structural diversity. Guided by this theory, we introduce Jacap, a capacity-aware eviction method that utilizes softmax-aware importance weighting and statistical leverage scores for subset selection. Extensive experiments across diverse architectures and benchmarks demonstrate that \textsc{Jacap} delivers superior performance in most scenarios, particularly in high-compression regimes.

MASTER's inter-stock multi-head attention -- the module responsible for modeling cross-sectional stock relationships -- accounts for 42.5% of model parameters and 25% of predictive value. We systematically decompose this module and uncover a surprising structure: the learned attention is near-uniform (perplexity 278/300), yet forcing exact uniformity eliminates all cross-sectional discrimination. Spectral analysis resolves this paradox: the deviation from uniformity is low-rank (effective rank ~65, top-10 modes capture 96.5% of energy), explaining why sparse approximations consistently fail while Nystrom low-rank attention (m=32 landmarks) matches full O(N^2) attention at O(mN) cost -- certified equivalent via TOST at both N=300 (5 seeds, Rank IC p=0.003) and N=800 (10 seeds, Rank IC p=0.034). Additional findings include: (i) attention anti-correlates with return similarity (Spearman rho = -0.614; on the industry-labeled subset, -0.645 unconditionally and -0.627 after controlling for industry, beta, and volatility), suggesting complementarity-seeking rather than correlation mining; (ii) all graph-based alternatives degrade performance, with hard masking worse than complete module removal; and (iii) at N ~ 3,500 with adapted architectures, no cross-stock module (GCN, Nystrom, or MASTER-style pipeline) significantly outperforms a per-stock LSTM baseline (n=4 seeds), indicating that the benefits observed at smaller scales do not trivially transfer. These results establish that the inter-stock attention's value resides in a compressible, dynamic, near-global redistribution that rewards low-rank approximation but resists sparsification.

Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linear, semantically meaningful features. Such methods are commonly interpreted as identifying features that persist in the residual stream and that subsequent layers build upon. We challenge this view by identifying the Transformer Layer Correction Mechanism (TLCM), wherein adjacent transformer layers systematically counteract portions of each other's contributions. TLCM appears in 5 out of 7 major open-source model families and activates across nearly all tokens in diverse texts. We show that TLCM emerges during pretraining, operates most strongly on contextually dependent tokens, and adaptively calibrates its correction strength based on the preceding layer's output. Using the layer Jacobian, we further show that TLCM selectively corrects specific subspaces while reinforcing others, which we interpret through a ``propose-and-reject'' framework in which layers propose candidate features and subsequent layers selectively remove inappropriate ones. This dynamic suggests that the residual stream at any layer contains transient proposals alongside persistent features, helping explain why SAE feature descriptions often have low specificity, why effective model steering requires extreme feature amplification, and why transcoders hold a theoretical advantage over SAEs.

Dynamic sparse attention reduces long-context prefill cost by routing each query chunk to a small set of key chunks at every Transformer layer. The sparse attention kernel avoids most token interactions, but the router still rebuilds a chunk--chunk score matrix layer after layer, even when the selected routes change little. We introduce RouteRelay, a router-agnostic method that reuses only route metadata across depth while continuing to compute attention with the current layer's queries, keys, and values. Anchor layers perform full routing. Intermediate layers rescore the previous top-$k$ route and a compact sentinel set of near-miss and randomly probed chunks. A query row is rerouted only when a sentinel challenges its weakest selected chunk. We give a top-$k$ stability condition, a probabilistic bound on missed challengers, and a row-selective GPU execution design. In a reproducible empirical evaluation, RouteRelay retains at least 99.99% route recall while rerouting 25.0%, 55.4%, and 78.2% of rows under low, moderate, and high cross-layer drift, respectively. Across routing scales, RouteRelay retains 100.0% recall while evaluating 38.4--51.6% of full-routing score pairs as the key-chunk count grows from 128 to 1024. Its unfused CPU execution remains slower than dense matrix multiplication, exposing row compaction and ledger updates as the main kernel-engineering targets.

Inference energy per token drives the cost and carbon footprint of deployed transformers. It is dominated by dense matrix products that incur fused multiply-accumulate (FMA) operations and memory traffic. To reduce these computations while retaining dense tensors for high GPU throughput, we develop a methodology from first principles to adapt structural complexity during training to maximize inference utility per unit compute. Channel penalties drive entire tensor slices to zero to enable physical removal while preserving density and the network function. The natural approach, penalizing the norm of operator components acting through each channel, is provably destabilized by gauge freedom. We resolve this pathology with GaugeLasso: additive symmetric group-lasso penalties that recover a monotone function of product-norms when the network converges to gauge balance. Our equilibrium analysis enables per-channel calibration to correctly suppress slices that under-perform in inference utility per unit compute. Under adaptive pressure, the network reorganizes into depth-dependent structural profiles that can be far smaller than the architecture required to learn the task. On polynomial long division over $\mathbb{F}_{31}$, compute compresses from 148 to 255 times with perfect accuracy. On character-level language modeling, compressed models outperform the hand-designed baseline at equal FMA. On masked autoencoding, a compression trial exposes which axes were over-provisioned and which saturated, guiding a better second design. Compaction also accelerates training monotonically as the model progresses. Post-hoc pruning with the same utility ranking cannot reach these structures, showing that sustained pressure is central to discovery of efficient models. Retraining a discovered architecture recovers baseline quality on our statistical tasks, but fails on our exact algorithmic task.

Post-hoc sparse attention accelerates long-context prefill by routing each query to a small set of token-level interactions. Hard selection, however, assigns zero probability to every omitted chunk: a routing miss cannot be recovered, and a fixed expansion budget spends the same work on easy and ambiguous queries. We introduce Coarse-to-fine Error-aware Dynamic Attention Routing (CEDAR), a coarse-to-fine method that keeps the language model frozen while preserving global coverage. Each semantic chunk contributes a cheap key--value summary to a residual attention path; chunks with high estimated approximation error are then expanded to exact token attention. Exact and summarized contributions are combined in a single softmax normalization, so refinement replaces, rather than duplicates, coarse evidence. We derive an output-error bound governed by within-chunk key/value dispersion and use it to allocate a variable refinement budget. A controlled clustered-attention study shows that residual summaries reduce reconstruction error by more than 98% relative to hard dropping at equal exact-chunk budgets. Experiments on long-context benchmarks demonstrate that CEDAR recovers most of the quality lost by hard sparse routing while maintaining approximately $3\times$ kernel speedup at 128K context.

A transformer can make an attribute linearly decodable in its residual stream at a depth where that attribute does not yet influence the output. This gap between where information is readable and where it is used has been shown for attributes stated directly in the input. We ask whether it also holds for an attribute the model must infer gradually over a conversation, namely how expert its dialogue partner is. Using ExpertCollab, a corpus of multi-turn research-planning dialogues between model-played personas at four expertise levels, we find that partner expertise is most decodable in the early layers and falls to near chance before the midpoint of the network. Counterfactual patching shows that injecting the expertise difference at the layer of peak decodability barely changes a fixed late-layer readout, whereas the same difference injected past the midpoint propagates almost completely, a separation of more than an order of magnitude. A content-matched random control and a probe-free diagnostic place the transition at the same early layer, and a statically specified control attribute stays decodable throughout. An inferred relational attribute is therefore represented well before it becomes causally active, which bounds where any attempt to read out or steer partner-conditioned behavior must intervene. We use one model on a synthetic corpus as an initial demonstration.

Recent image-to-3D generation models built on flow-matching diffusion Transformers (DiT) can produce high-fidelity meshes, yet their post-training strategy remains largely unexplored. There exist several critical bottlenecks in reinforcement learning: the inherent difficulty of defining comprehensive rewards for 3D geometric quality, and the gradient interference that arises when jointly optimizing heterogeneous objectives. Inspired by the practicability of on-policy distillation (OPD) in large language models and image generation, we propose \textbf{Flow3D-OPD}, a two-stage post-training framework that introduces multi-teacher distillation into 3D geometry generation. In the first stage, we utilize the semi-policy to enhance the foundational capability of the pretrained model and then design an agentic verifier for 3D geometric quality evaluation. Based on the verifier, we could cultivate domain-specialized teacher models via direct preference optimization (DPO). In the second stage, we consolidate heterogeneous expertise into a unified student model through on-policy distillation with hard task-routing sampling and gradient accumulation, which could mitigate the gradient interference in joint optimization. Without relying on elaborate modifications, our straightforward yet effective design achieves consistent improvements across all geometric quality dimensions and surpasses all teacher models in the average metric. Extensive experiments demonstrate that our approach provides an effective paradigm for reinforcement learning in 3D generation.

Cardiac, neural, behavioral, and speech measurements from wearable and mobile devices provide partial, noise-sensitive views of physiological state. BioSync combines these measurements into the \textbf{BioSync Index (BSI)}, a continuous composite digital biomarker defined under the BEST framework. The model applies multi-head self-attention to modality tokens and adds a linear branch whose hypothesis class includes standard feature concatenation. This architecture is motivated by latent-variable measurement theory and by the possibility that joint observations contain information unavailable from individual modalities. We evaluated BioSync on two literature-informed synthetic cohorts: a four-modality cognitive-decline cohort using HRV, EEG, actigraphy, and speech, and a metabolic-autonomic cohort structured around the public AI-READI wearable schema. In the cognitive cohort, BioSync and concatenation obtained AUCs of 0.928 and 0.926, respectively. In the metabolic cohort, BioSync obtained accuracy/F1 of 0.764/0.766, compared with 0.756/0.758 for concatenation. The BSI correlated with latent severity in both cohorts ($r=0.91$ and $r=0.68$). A pure-attention ablation obtained cognitive-cohort AUC 0.911, locating the increase to 0.928 in the combined wide-and-deep architecture. With matched modality-dropout training, BioSync led concatenation at five of six cognitive-cohort corruption rates and at the highest metabolic-cohort rate. Its cognitive-cohort AUC was also higher than five published digital-biomarker reference values, although differences in datasets and tasks preclude a controlled benchmark claim. Comparison with single-modality, early-fusion, and late-fusion designs across six prespecified criteria identifies the model's computational properties; validation on real cohorts remains necessary.

Blackwell's 4-bit floating-point (FP4) tensor cores do not automatically make attention faster because softmax conversion and on-chip dependencies dominate once its matrix products shrink. We address this with \emph{Direct-P} for noncausal inference and a causal path that passes the forward quantization directly into backward. Direct-P maps scores directly to FP4 probabilities and reaches up to 2.13$\times$ the bfloat16 (BF16) forward throughput on an NVIDIA GB200. The causal path reconstructs probabilities from saved quantized queries and keys and uses 8-bit floating-point (FP8) gradient operands, accelerating a complete single-GPU 8-billion-parameter update by up to 1.14$\times$. Matched distributed training retains FP8 probabilities and values; every tested MXFP4 probability/value training trajectory diverges.

Hybrid LLMs pair softmax attention with linear-attention layers such as Gated DeltaNet (GDN), whose recurrent state summarizes the context in fixed size. Early community 4-bit quantizations of Qwen3.8-27B (48 GDN layers, 16 attention layers) left the GDN block in 8- or 16-bit precision -- especially its decay and write-strength gates -- on the intuition that errors in a recurrence accumulate over long contexts. We test that intuition by building Minima: NVFP4 W4A4 on all 496 linear layers, GDN included. Across perplexity at 4K/32K, MMLU-Pro, GSM8K, AIME'25, GPQA-Diamond, LiveCodeBench, and RULER retrieval to 64K, Minima matches BF16 within seed noise (5-task average -0.52) while being the smallest (17.5 GiB) and fastest-prefill (+14-19%) recipe we compare, and its 32K perplexity gap shrinks with position. A four-part mechanism study explains why: (i) NVFP4's 16-element block scaling localizes the residual stream's extreme outliers, equalizing activation error across layer roles; (ii) the supposedly fragile gate projections are the least sensitive -- softplus/exponential and sigmoid parameterizations compress ~11% GEMM error to ~2% output error; (iii) the delta-rule recurrence holds injected noise at a flat plateau over 32K tokens and forgets a state impulse within hundreds of steps, because each write overwrites the state along the current key direction; (iv) the per-token quantization cost washes out with context instead of compounding. We also repair a global-scale mismatch that arises when per-module-calibrated NVFP4 checkpoints are served by kernels that fuse those modules into one GEMM, and show calibrated FP8 KV-cache scales are performance-free. The result: a practical recipe -- quantize everything, ship KV scales -- and a mechanistic account of why the recurrent half of a hybrid LLM is the easy half to quantize. Checkpoint: https://huggingface.co/minima-ai/mnma_qwen3.8_27b_nvfp4

What can a single layer of self-attention compute? We study head complexity: the minimum number of attention heads required to compute a function in a one-layer attention-only model. We establish an exact hierarchy under this measure: $k$ heads compute $k$-bit parity but cannot compute $(k+1)$-bit parity. The lower bound is unconditional in the two resources a transformer might otherwise exploit; it holds at unbounded embedding dimension and unbounded numerical precision. The proof rests on an alternating-sum obstruction: after clearing the softmax denominators, every monomial in the resulting decision polynomial omits at least one of the $k+1$ input bits, forcing its correlation with parity to vanish. The same obstruction yields lower bounds for related tasks, including the well-studied multi-hop induction-head task. We also establish compactness bounds for embedding dimension and numerical precision. Specifically, a compactness theorem shows that any function computable at all can be computed with embedding dimension and precision bounded by the discrete data of the task, namely, head count, alphabet size, and length. Thus, potentially unbounded dimension or precision provably cannot substitute for heads. Finally, we derive nearly matching universal bounds for general binary functions: $2^n$ heads suffice to compute every $n$-bit binary function, with one head per monomial in its multilinear expansion, while a counting argument shows almost all such functions require $Ω(2^n/n^2)$ heads. This lower bound matches the upper bound to within a $\operatorname{poly}(n)$ factor, even when dimension and precision are unbounded. Together, these results characterize head requirements for Boolean computation in this model.

A long-lived KV cache must be compressed before the queries that will read it exist. Selection by observed attention collapses there: on a NoPE-MLA model, H2O and SnapKV retrieve 0.00 and 0.33 of needles at 8x compression, because a token's importance has not yet been observed. VestigeKV instead derives a sparse attention pattern from a signal the cache already carries, occupying the sparse-attention literature's one unoccupied quadrant: training-free and query-independent. In NoPE-MLA the 64-dimensional decoupled branch is a vestige of RoPE that training repurposes into a salience channel; reading 11% of each row, it partitions the cache into an attended tier and a GPU-resident archive that no row ever leaves, reachable each step by a certified, query-adaptive trigger. Nothing is trained and cache rows are never quantized, so every quality effect attributes to selection and scheduling. On Kimi Linear 48B, retrieval holds at 1.00 under 8x and 0.96 under 32x from 8k to 65k context, with zero gap to full-row selection, and the recall tier holds 128x at 1.00 (8k). Both tiers stay on the GPU, so the win is speed, not memory: the per-step scan reads ~26% of the bytes dense attention would, and on a two-node sglang deployment the crossover sits at ~40k context, reaching 1.18x at 256k and 1.39x at 496k. The mechanism is exclusive to NoPE: the identical operator on a RoPE MLA collapses to 0.08, query-independent salience exists only without rotation, and query-universal exact merging is provably impossible under RoPE. All thresholds were frozen before their data; 20 archived verdicts and 8 closed routes accompany the paper.

Decoding-time KV cache compression research focuses heavily on designing better token scoring functions, while the temporal rule that aggregates scores across decode steps is often treated as an implementation detail. Under aggressive KV compression, we find that exponential-moving-average (EMA) aggregation makes approximately order-preserving scorer modifications largely indistinguishable at the eviction-set level. Value-norm and entropy variants remain highly correlated with attention and produce nearly unchanged retention sets, whereas KeyDiff, key norm, recency, and a learned scorer alter the ranking and degrade substantially. We associate this stability with the evaluated aggregation, which couples layer weighting and temporal retention. Building on this observation, we introduce InertiaKV, an EMA-based decoding-time eviction method, and InertiaKV-Lazy, its periodic-refresh variant, which yields 1.34-1.46x decode throughput relative to full refresh InertiaKV. We also study Score-Free decoding as a separate empirical operating point: it scores the full context once at the first decode step, freezes that ranking, and incurs an average quality change of +0.03 while removing all subsequent scoring. Across six open-weight backbones and the LongBench, LongBench-v2, and RULER benchmarks, the results identify temporal aggregation and ranking preservation as distinct, consequential design factors; they do not imply that scoring quality is irrelevant in general.

Large language models achieve superior performance on tasks that require extended reasoning, but long chains of thought make the KV cache a severe memory bottleneck. Existing KV cache compression methods share one paradigm: score each cached token by some estimate of how much it will matter later, and keep the top-scoring ones. We show that the selection signal contributes almost nothing. Random Attention keeps the prompt and evicts uniformly at random within each attention head, computing no score at all; across four models and six reasoning tasks it matches the strongest prior evictor while serving 32-43% higher throughput than it in vLLM deployment. Controlled experiments explain this by showing that 1) the prompt is the fragile part of the cache, and most of the gap between selectors is just whether their selection signal happened to keep it; 2) the reasoning trace protects itself against eviction with redundancy at two levels, in the text (the model restates what it still needs as it works) and across attention heads (each keeps its own copy of the trace), so once the prompt is safe, a random draw retains enough copies of what the model still needs, and no score is required to pick them. Our code is publicly available at https://github.com/SalesforceAIResearch/Random-Attention.

Hybrid architectures combining Full Attention (FA) and Linear Attention (LA) are increasingly prominent, yet their allocation remains heuristic. We seek an evidence-grounded basis in head-level functional organization learned by RoPE-based Transformers. Behavioral probes do not yield a complete taxonomy, so we propose two intervention metrics: RoPE Frequency Importance Score (RFIS), measuring how each frequency affects a head's attention distribution, and RoPE Positional Dependence (RPD), isolating dependence on rotary positional modulation. On Qwen3-series models and Llama3.1, RFIS suggests and RPD verifies a complete taxonomy of retrieval and positional heads separated by a salient mid-low-frequency band. Controlled Transformers show that this boundary follows the training-length positional scale; we term it the Global Positional Band (GPBand). The analysis suggests a potential cause of zero-shot length-extrapolation failure and yields two principles: positional modeling should operate only locally, with global access through position-independent retrieval; and both functions should be assigned at head granularity with layer-specific allocation. We instantiate them in Head-wise Hybrid Architecture (HwH), using NoPE FA for global retrieval and LA for local positional modeling. With an FA-to-LA ratio below 1:3, HwH retains strong language modeling and commonsense reasoning while improving retrieval and substantially strengthening zero-shot long-context extrapolation over Transformer, LA, and a layer-wise hybrid baseline. Ablations validate both principles and component roles, highlighting principled hybrid architecture design as a promising route toward future foundation models.

Modern supply chain operations can require coordinating replenishment across thousands of heterogeneous items under correlated stochastic demand, heterogeneous lead times, and shared fixed ordering costs, yielding observation spaces exceeding $10^4$ dimensions. At this scale, rolling-horizon stochastic mixed-integer linear programs (MILPs) become prohibitively slow, while standard reinforcement learning (RL) methods face increasingly challenging credit assignment in high-dimensional action spaces. We introduce OR-Transformer, a deep reinforcement learning framework for joint replenishment under stochastic demand, with an item-permutation-equivariant Transformer architecture and pathwise-gradient training through the inventory dynamics. Across problem sizes up to 1,024 inventory items, OR-Transformer increasingly outperforms learning-based and rolling-horizon MILP baselines as scale grows. It also reduces online decision-making time by over 4 million times relative to MILP solvers, enabling real-time, large-scale deep RL in supply chain operations.

The attention prefilling phase of long-context LLM inference scales quadratically, making self-attention a severe computational bottleneck. Traditional sparse attention methods mitigate this through fixed patterns or offline profiling, but lack the flexibility to adapt to input-dependent attention structure. Recent dynamic methods address this by routing heads to sparse patterns in real-time, but rely on indirect routing proxies with overhead and budget allocation mechanisms that overlook the post-softmax mass hierarchy. We present CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling), which identifies and addresses two structural challenges in this dynamic routing paradigm. First, we show that the routing decision can be read directly off the structure of the proxy attention map. We replace the Jensen-Shannon Divergence (JSD) routing with C_struct, a structural proxy that measures mass at Vertical-Slash compatible positions and reproduces JSD's routing decisions while eliminating both the pooled matmul and subsequent KL divergence overhead. Second, we formalize the post-softmax mass cliff and demonstrate theoretically that strictly cumulative coverage thresholds accumulate O(n) background noise at long contexts. CRISP navigates this via a sink-aware threshold grounded in the noise floor. Empirically, across InfiniteBench, RULER and LongBench on two model families, CRISP is the strongest sparse method overall and matches or exceeds exact dense attention on retrieval-heavy benchmarks, recovering up to +28.0 pp on retrieval tasks over baselines and achieving up to a 5.30x attention speedup at 512k tokens, driven primarily by our O(n) noise elimination during selection while preserving structural integrity.

Automatic Target Recognition (ATR) in Synthetic Aperture Sonar (SAS) is a task largely dominated by deep neural networks (DNNs). Most SAS-ATR models use convolutional neural network (CNN) architectures whereas transformer-based architectures have had much less representation in the literature despite being state of the art in general computer vision (CV) research. Additionally, researchers have had mixed results in attempting to overcome challenges presented by a scarcity of labeled training data by using methods such as data augmentation and the use of pretrained weights from a variety of imaging modalities. In this work, we compare the performance of modern CNN and transformer-based DNNs to determine which architecture and training configurations elicit the highest performance in SAS-ATR. We investigate how network size, architecture, pretraining method, data augmentation and other forms of regularization affect SAS-ATR performance with a focus on producing the highest-performing model and providing a roadmap for training state-of-the-art SAS-ATR models.

Vision Transformers (ViT) excel in semantic understanding but fail to discriminate between object instances (e.g., identical embeddings for two dogs), limiting their use in instance-level tasks such as object detection and instance segmentation. We propose Contrastive Vision Transformer (CoViT), a self-supervised learning framework that injects instance-awareness into ViT through geometry-guided contrastive learning. CoViT uniquely coordinates ViT's attention maps and embeddings by constructing triplets: (1) Attention-guided masking: Refine multi-head attention via adaptive thresholding and morphological operations to generate instance masks, identifying foreground anchors; (2) Hardest contrastive mining: For each anchor, computing pairwise embedding similarities to select the intra-instance hardest positive (least similar patch within its mask) and inter-instance hardest negative (most similar patch from other instances), with intra-instance regions masked during negative search. These triplets drive a contrastive loss that simultaneously compresses intra-instance variance and expands inter-instance margins, forcing ViT to discern subtle geometric and appearance differences between instances. CoViT consistently achieves stable performance gains of over 2 AP points across multiple instance-level perception tasks by using ViT as backbone architecture. Notably, CoViT requires no extra decoders or labels, demonstrating that a pure ViT can learn instance-aware representations via inherent attention priors and targeted contrastive constraints. Code and models will be released.

We extend recent work establishing an equivalence between one-layer transformers and nearest-neighbor classifiers in the binary setting to the multiclass case. By leveraging the simplex encoding, we show that one-layer transformers with an argmax classification head behave identically to a one-nearest-neighbor classifier in the multiclass setting. This closes a gap left by prior work, whose multiclass result relied on a non-standard rounding-based approach rather than the typical argmax head used in practice.

Scaling Transformers has driven large gains in language modeling, but transplanting this to behavior-sequence modeling in production ranking is challenging: recommendation differs in signal quality, where behavior sequences are noisy, temporally irregular, and sparsely supervised, and in computation asymmetry, where each request scores many candidates against one shared user history under tight latency budgets. We propose ReST, a recommendation-native Transformer scaling framework. For signal quality, it introduces a sequence encoder with dual-gated attention, rotary positional and temporal embedding, stabilized residual normalization, and training-only auxiliary objectives. For computation asymmetry, it factorizes ranking into a heavy reusable encoder and a lightweight cross decoder with projection-free KV attention and token-specific parameterization, coupling user-level shared-prefix training with shared-prefix serving for compute-once, decode-many-times ranking. Across industrial and public benchmarks, ReST achieves higher accuracy and scales more consistently along sequence length, depth, and width, where LLM-style Transformer blocks saturate. A one-week online A/B test on a production advertising platform improves online AUC by 1.31% and lifts a core revenue metric by 11.93% within a 50 ms P99 budget; ReST has since been fully deployed in production, showing that behavior-sequence scaling remains a promising, under-exploited axis for production ranking.

With the rapid and continuous growth in the incorporation of machine learning models based on the Transformer architecture, capable deployment is in high demand. In this context, capable deployment refers to operational performance aspects, e.g., throughput and latency, as well as efficiency aspects, e.g., energy consumption. When it comes to the task of inference using such models, purpose-built hardware accelerators provide a lucrative alternative to common deployment choices, such as Central Processing Units (CPUs) and Graphics Processing Units (GPUs). The Field Programmable Gate Array (FPGA) platforms category is an example of such alternative accelerators, promising implementation flexibility, energy efficiency, improved latency and suitability for on-site deployment. We investigate the most recent advances, trends, and design choices for Transformer inference on FPGA platforms. We perform a systematic literature review, extracting and delving into preferred techniques for implementation and optimisation. This study and the provided taxonomy of topics could act as a guide for researchers from the academia and industry alike.

We identify a recurrent algebraic regularity in Transformer attention: a sparse subset of effective OV operators $T=OV^\top$ nearly closes under composition, $T^2\approxαT$. Across six pretrained endpoints spanning 2.8B--235B parameters, 3.98--8.00% of heads reach squared closure alignment $\mathcal{P}\geq0.9$, while no matched within-layer O/V mismatch does. An exact principal-coordinate factorization, $T=Q_OKQ_V^\top$ and $T^2=Q_O(KDK)Q_V^\top$, separates within-support transport from read--write return geometry. Across all 7,304 heads in nine MHA/GQA models, scrambling only the orientation of $K$ while preserving singular values, norms, factor spans, and principal angles reduces median closure from 0.336 to $1.04\times10^{-4}$; trained orientation wins for 98.64% of heads and in every layer. Constructive searches show that high closure is feasible in every surveyed layer, but usually not attained. Retrospective trajectories in three independently trained lineages further separate broadly available capacity from the orientations attained by final strong heads. Under exact value sharing, headwise closure extends to a right-action algebra, $T_iT_j=α_jT_i$. Seven-model experiments verify the approximate law and reveal distinct oblique projections with a shared value-defined kernel. These results characterize scaled idempotence as a sparse trained orientation within broadly available geometric capacity and show how value sharing extends a headwise relation into a local operator algebra.

Denoising diffusion models are the dominant architecture for image generation, whereas most natural language generation and modeling are primarily handled by well-known transformer architectures employing attention mechanism. Here, we show that diffusion models also inherently use an attention mechanism very similar to that of transformers. Therefore, attention emerges as a universal machine learning principle, based on a general training objective. We also show similarities in basic functional principle of auto-encoders and attention-based models. These equivalences allows us to interchange these designs based on practical requirements. As an example, we can reformulate the diffusion framework to reduce the lengthy training process and computation-intensive image generation. Using this approach, a simplified algorithm is proposed for image generation which is based on attention mechanism. Results show that the attention-based implementation achieves comparable performance with significantly less effort and computational resources.

A transformer carries each token's state from layer to layer, and the whole vocabulary carried together forms a cloud that moves with depth. We ask whether a trained network moves this cloud the way optimal transport would: at the cheapest cost, and along the map that pairs each token with its optimal destination. We measure both on Pythia-160m and Pythia-410m, with an exact assignment between consecutive layer clouds, a measured sampling floor, calibration on couplings known to be optimal, and a split of the cost into the common shift of the cloud and the token-specific moves. At the last layer, both models move their tokens where the optimal-transport map sends them, at the optimal cost for Pythia-410m and slightly above it for Pythia-160m. At the first layer they do not. In between, single layers can be judged on cost at only two of ten transitions, and blocks of several layers move the cloud at close to the optimal cost. The agreement at the last layer is much weaker at initialisation (0.64 against 0.86) and grows with training.

This work examines recent byte-level models and their failure to perfectly manipulate characters. State-of-the-art byte-level models use a hierarchical structure, starting at the byte level, downsampling to the word level, and then upsampling back to bytes. While this improves training and inference efficiency, we find that the hierarchical design itself limits character-level understanding, with pure byte-level models consistently outperforming hierarchical variants on character manipulation tasks. Ablating transformer layers into attention and feed-forward components further reveals that byte-level attention is the primary mechanism driving this behavior. Together, our results provide an explanation for the character-level failures of hierarchical byte models and establish a clear trade-off between computational efficiency and fine-grained character understanding.

Probing studies have established that syntactic information is decodable in early and middle transformer layers, but what happens to that information in later layers remains poorly understood. We apply a cross-layer generalisation analysis to three Greek-tuned large language models evaluated on tightly controlled minimal pairs: object-relative constructions in Modern Greek, where canonical (Subject-Verb-Object; SVO) and non-canonical (Verb-Subject-Object; VSO) orders differ only in within-clause word order, while preserving propositional meaning. When a probe trained on late layers (20-31) is tested on each early layer individually, it produces below-chance transfer (cluster-corrected, p<0.01), classifying 99.3% of non-canonical sentences as canonical. Probe coefficients reverse sign around layer 22, indicating a directional recoding toward the canonical form rather than simple information loss. These findings characterise a representational format change in late transformer layers that goes beyond the well-established decline in syntactic decodability, and they generate a directly testable prediction for human EEG and MEG decoding studies using the same stimuli. Code and stimuli are publicly available on OSF.

Memory is widely viewed as an important unsolved problem for LLMs and VLMs, and current benchmarks typically evaluate it by testing accuracy over long text or video. However, accuracy alone misses properties that matter for real long-horizon tasks. We introduce ECCBench, a benchmark and evaluation protocol that measures memory beyond a system's capacity--its raw accuracy at a specific budget--via three axes we call ECC: efficiency--the computation, in FLOPs, needed to answer from memory; compression--whether compressible inputs are remembered more accurately or efficiently; and calibration--whether the system abstains in response to its own uncertainty and the cost of an error. We find that pretrained VLMs compress their memory over text but not video and are poorly calibrated on both. Among a broader set of memory backbones, several non-Transformer architectures achieve better compression-calibration tradeoffs than RoPE Transformers, suggesting they may be useful components for agents operating over long horizons.

Learning generalizable algorithmic computations remains a challenge for neural networks, as reflected in persistent failures on compositional and length generalization benchmarks. We present a provably correct, transformer parameterization (with only 280 learnable parameters for Boolean algebra tasks) capable of learning and evaluating problems of any depth or length. We assume inputs are fully parenthesized, well-formed expressions. Our approach conceptualizes algorithmic tasks as circuit models embedded in transformers, enabling depth-1 circuit reduction in a single forward pass. To achieve depth generalization, we introduce a positional encoding that tracks each gate's depth within the circuit, enabling the model to identify evaluable subexpressions at each iteration via masked hard attention, with $O(n)$ per-iteration complexity via linear attention. Combined with an autonomous halting criterion, the model terminates after $d$ iterations for problems of depth $d$, yielding $O(n \cdot d)$ total complexity. We show that training on shallow problem instances (depth 1 and depth 2) effectively recovers interpretable parameters that {\em snap} into place, resulting in exact length generalization. Though we establish that our construction provably evaluates Boolean expressions -- a universal symbolic computation -- of arbitrary length perfectly, in other experiments we also demonstrate that our transformer variant can learn and generalize perfectly (100% accuracy) on other common length generalization benchmarks, including modular arithmetic and ListOps.

The development of long-context Large Language Models (LLMs) is constrained by the memory bandwidth bottleneck and quadratic complexity of the attention mechanism during decoding. To overcome the inherent trade-offs between the memory overhead of metadata-based metrics and the computational inefficiency of adaptive selection strategies, we present Faster Flash Decoding (FFD), a novel hardware-algorithm co-design framework designed to break the memory wall in long-context decoding. FFD integrates the selector and computer into a fully fused kernel, replacing external metadata indices with content-aware scanning via low-bit quantization. Furthermore, we introduce the top-delta strategy, which dynamically filters blocks to achieve distribution-adaptive sparsity without global synchronization. Offering a training-free and plug-and-play solution, FFD also enables the reuse of scanning results for computation, achieving up to 11.6x kernel-level speedup and scaling to 256K context length, with 2.37x end-to-end throughput improvement. Empirical validation on RULER and LongBench confirms that FFD maintains model accuracy while delivering high-ratio sparsity, with code available at https://github.com/qluoluo/faster-flash-decoding

Whisper exposes speech through a fixed 1500-token encoder interface, now a default representation for ASR decoders and Whisper-based speech language models (SpeechLMs), yet its redundancy remains largely unexamined. We propose stride-k subsampling, a deterministic indexing operation that retains every k-th token after the convolutional stem or encoder transformer. Across five Whisper scales, k=2 preserves baseline WER at both positions, with CKA attributing this stability to acoustic overlap at the stem and attention-induced redistribution at the encoder output. Applying stride-2 at both positions cuts audio tokens by 75% and total GFLOPs by 52-58%, with small WER costs on most ASR benchmarks and larger costs on harder ones. The same configuration extends to three Whisper-based SpeechLMs, yielding modest accuracy drops on stronger baselines and larger drops on weaker ones, while reducing end-to-end latency by 19.6-27.4%. Requiring no training or auxiliary computation, stride-k subsampling exploits Whisper's preprocessing redundancy, indicating that its audio-token interface carries more capacity than downstream tasks require.

LLM training is increasingly vulnerable to silent data corruption (SDC), yet existing protection methods largely treat Transformer computations uniformly because their vulnerability remains poorly understood. We present the first systematic characterization of SDC vulnerability across major computation interfaces in both the forward and backward passes of Transformer training. Our analysis reveals two distinct error propagation mechanisms: forward-pass vulnerability is highly location dependent, with faults on the Q/K path producing persistent training deviations, whereas backward-pass vulnerability is largely governed by gradient exponent distributions rather than computation locations. Motivated by these observations, we propose TrainSDC, a characterization-guided protection framework consisting of Q/K-path recomputation, residual-gain monitoring, and exponent-aware gradient scaling. Experiments on Llama 3.2-1B and Qwen3-0.6B show that TrainSDC maintains training behavior close to fault-free execution under both sparse and dense fault injection while introducing only 1.65%-6.76% runtime overhead.

Representational similarity is foundational to analyses of deep networks, yet distances between point-valued representations are not intrinsically tied to downstream function: nearby states may produce different behaviors, while distant states may behave similarly. We instead give representations volume, turning similarity into statistical distinguishability. Overlapping stochastic representations necessarily induce overlapping downstream distributions, grounding latent comparison in model function and bringing it under information-theoretic tools such as the data-processing inequality. We realize this idea in pretrained transformers through a light-touch modification to LayerNorm: at each residual-stream read, we normalize the state, add isotropic Gaussian noise, and renormalize. During distillation fine-tuning, one learned allocation parameter per residual-stream read distributes a fixed global rate budget across the processing stack. The resulting model can be viewed as transformer blocks reading the residual stream with learned finite precision under a shared global rate budget. Using the Bhattacharyya coefficient, we trace which counterfactual distinctions are preserved through MLP blocks or selectively exposed to the query, key, and value computations of individual attention heads. Experiments on ViT-S and GPT-2 small reveal the depthwise propagation of continuous visual perturbations and head-specific sensitivity to token distinctions aligned with known attention motifs. These results establish distinguishability as a functionally grounded lens on transformer computation that complements existing interpretability approaches.

Transformer-based decoders for 3D instance segmentation typically commit to a fixed number of queries and positional modeling calibrated on the training distribution rather than on the scene at hand. Indoor scans vary widely in spatial extent and object count, so a fixed query set over-initializes small scenes and under-initializes large ones, while learned absolute and relative encodings are bound to the training scenes' extents and can saturate. We present AQ3D, which is designed to handle scenes of various sizes during training and inference. Queries are instantiated at a fixed ratio of the scene's superpoints, forming an overcomplete set whose background rejection is entirely left to the decoder. Positional information is encoded using 3D RoPE over quantized metric coordinates, replacing learned bounded lookup tables of prior decoders. Further, we improve the decoder itself by using attribution-based superpoint pooling, a mask refinement branch, and a cosine classifier for background rejection. Experiments show our method sets a new state-of-the-art on validation and hidden test splits across the datasets ScanNetV2, ScanNet200, and ScanNet++V2 among decoder methods trained without additional data augmentation. Code is available at \href{https://github.com/kenomo/aq3d}{github.com/kenomo/aq3d}.

We give a complete characterization of equivariant multi-head self-attention (MHSA): if an MHSA layer is equivariant to a symmetry group $G$, then $G$ can only act by permuting head-clusters, with QK and OV matrices satisfying an equivariance constraint tied to the group action. As a consequence, we prove that any fixed MHSA architecture that achieves exact equivariance by polynomially parameterizing unconstrained MHSA parameters inevitably leads to expressivity loss within the class of equivariant maps: the equivariance locus of unconstrained MHSA forms a union of extremely many Zariski-irreducible components in a reduced parameter space, and any single architecture covers at most one. For $G=D_4$ acting on $C$ copies of the regular representation as the token feature space, we show that there are $Ω(C^{64})$ components for eight attention heads.

Hybrid linear-attention architectures have recently scaled to large open-weight models, offering quality competitive with full attention while substantially reducing key/value (KV) cache growth. However, their in-place recurrent-state updates complicate cache management: prefix reuse requires state checkpoints alongside full-attention KV, while storing state checkpoints in full increases memory pressure, leading to more evictions and repeated prefill. By analyzing the decay structure of Gated DeltaNet (GDN) and Kimi Delta Attention (KDA), we find that different heads and channels retain prefix information over markedly different timescales, which we term \emph{retention horizons}. This variation suggests substantial compression potential in persistent state checkpoints. Building on this observation, we introduce \emph{Decay-Aware State Compression} (DASC), which derives retention horizons from model weights, selects long-horizon state units, and packs them into a ragged state checkpoint layout. To integrate efficiently with tensor-parallel inference engines, DASC furtherly balances compressed state checkpoints across TP ranks. On reuse, DASC either zero-fills omitted units or refreshes them from a bounded suffix with additional compute cost. Across retrieval and end-to-end reasoning benchmarks on Kimi-Linear, conservative DASC configurations remain close to full caching while compressing KDA recurrent state checkpoints by $2.63\times$. Under fixed state checkpoint memory budgets, the resulting capacity gains reduce mean Time to First Token (TTFT) by 42.6% and improve input throughput by 68.4%. At larger compression ratio, suffix refresh recovers much of the accuracy lost to more aggressive omission, at the cost of additional replay computation. Qwen with GDN exhibits a similar quality--efficiency trend, showing that DASC extends from channel-wise KDA to head-wise GDN.

Transformer-based electrocardiogram (ECG) models commonly tokenize waveforms into fixed temporal patches. Though convenient, fixed patching can split heartbeat structures across token boundaries. We study beat-synchronous tokenization as a physiologically grounded alternative, comparing fixed patches with three beat-aligned strategies: resampled beats, adaptive pooled beats, and resampled beats augmented with R--R interval information. Experiments span two settings: 10-second 12-lead diagnostic classification on PTB-XL after MIMIC-IV-ECG masked pretraining, and 60-second single-lead rhythm classification on Icentia11k after patient-level contrastive pretraining. On PTB-XL, resampled beat tokens achieve the highest mean macro Area Under the ROC Curve (AUROC; 0.8945) and nearly match the best fixed-patch macro Area Under the Precision-Recall Curve (AUPRC; 0.7414), reducing average sequence length from 100 to 11.2 tokens. On Icentia11k, beat-synchronous tokenizers obtain comparable AUPRC to fixed patching with better stability across runs. These results suggest morphology-preserving beat tokenization is a compact, competitive alternative to fixed temporal patching.

We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and additional 51B parameters of n-gram embedding tables held off the accelerator. On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs. Token mixing uses a layer-wise hybrid of Gated DeltaNet (GDN) and global attention, with one full-attention layer in every four; at continued-pretraining time those full-attention layers are replaced by Qwen Sparse Attention (QSA), which scores context at micro-block granularity with a compressed lightweight indexer. The residual stream is widened to four branches and read through an elementwise gate, a design we call the Gated Residual (GR). Capacity is added outside the backbone by a single n-gram embedding layer whose tables are prefetched from host memory. We evaluate every candidate change along three axes: loss together with downstream benchmarks; the cost of the change in training, prefill and decode; and its effect on the optimal hyperparameters and training stability. Loss and downstream accuracy do not always move together: enlarging the n-gram vocabulary lowers loss monotonically while downstream accuracy saturates. The architecture and the Muon optimizer together shift the optimal learning rate and batch size upwards, render batch-size warmup unnecessary, and substantially improve stability under stress tests. Loss, benchmarks, efficiency and stability form one design problem. Solved jointly, they yield a recipe that is simultaneously more efficient, more capable and more stable.

Hybrid large language models interleave full-attention layers with linear-attention layers to reduce the cost of long-context inference. This structure complicates prefix caching: full-attention key-value caches are token-addressable, whereas linear-attention layers maintain recurrent states that cannot be rolled back to arbitrary prefix boundaries. Existing hybrid prefix caching methods address this mismatch by storing recurrent-state checkpoints. As a result, token-level matches are directly usable only at positions aligned with stored checkpoints, constraining prefix reuse to a discrete set of boundaries. We present Tail-Replay, a prefix caching mechanism that enables unconstrained token-level prefix reuse in hybrid large language models. The key insight is that linear-attention mechanisms such as Gated DeltaNet can be viewed as a structured, lossy compression of the input prefix: gated recurrent updates progressively attenuate the contributions of earlier inputs. Consequently, the recurrent state of a matched prefix can be well approximated by replaying only a short, recent suffix of that prefix. Tail-Replay exploits this property by caching the exact full-attention key-value cache while omitting recurrent-state checkpoints. On a cache hit, it reconstructs the linear-attention states by replaying a short, recent suffix of the matched prefix. As a result, the reuse boundary is determined by the shared tokens rather than by recurrent-state checkpoints. We evaluate Tail-Replay on three Gated DeltaNet-based hybrid models using the LongBench and RULER benchmarks. With only a 5--10% replay budget, it retains 92.8--99.9% of full-prefill quality on LongBench and RULER. For serving efficiency, we evaluate time-to-first-token speedups across multiple matched-prefix lengths---8K, 16K, and 32K. The speedup grows with prefix length, reaching $9.1$--$14.3\times$ over full prefill at 32K.

Large language models (LLMs) have demonstrated strong capabilities in handling long-context tasks, but processing such long contexts remains challenging due to the substantial memory requirements and inference latency. In this work, we discover that certain attention heads exhibit sequential consistency in their attention patterns, which can be persistently identified using a coefficient-of-variation-based algorithm. Inspired by this observation, we propose CateKV, a hybrid KV cache method that retains only critical token information for consistent heads, thereby reducing KV cache size and computational overhead, while preserving the majority of KV pairs in adaptive heads to ensure high accuracy. We show the unique characteristics of our algorithm and its extension with existing acceleration methods. Comprehensive evaluations on long-context benchmarks show that, while maintaining accuracy comparable to full attention, CateKV reduces memory usage by up to $2.72\times$ and accelerates decoding by $2.18\times$ in single-sample inputs, and boosts throughput by $3.96\times$ in batch scenarios.

Electronic Health Record (EHR) prediction models in the intensive care unit must learn from sparse and irregular measurements while preserving the clinical meaning of time and supporting transparent decision-making. We present INTERVenE, a family of Transformer architectures whose input is an interval-based, knowledge-based temporal abstraction (KBTA), a token stream of named clinical concepts (states, trends, events, contexts) drawn from a curated medical ontology, rather than an unnamed bin index or a raw measurement triplet. This naming layer is what we ask KBTA to do: it makes the model's per-token attributions resolve to clinical concepts by construction. INTERVenE offers two complementary variants: an auto-regressive decoder that generates future abstraction trajectories with a per-step risk readout (localizing \emph{when} and \emph{after which events} risk rises), and a bidirectional encoder for single-pass joint risk and time-to-event prediction. Evaluated on 57,078 MIMIC-IV admissions against GRU-D, STraTS, and KarmaLego, INTERVenE-Enc reaches a support-weighted AUPRC$_w$ of 0.672, improving by 0.041 over the strongest neural baseline with non-overlapping 95% bootstrap CIs, while also taking the best AUROC$_w$ (0.901) and length-of-stay MAE (44.4,h). INTERVenE-Ar (AUROC$_w$ $0.854$, AUPRC$_w$ $0.587$ under the same evaluation contract - a strictly harder generative readout) provides a complementary token-level risk trajectory. An input-representation ablation confirms the lift transfers across structured discretizations, positioning KBTA-based intervals as the interpretable substrate that makes per-token attributions resolve to meaningful clinical concepts within the deployed model.

Transformers rely on position embedding mechanisms in long context modeling in most cases. Rotary Position Embedding (RoPE) embeds positional information with independent 2D rotations, forming relative position terms in self-attention. However, its pairwise, block-based, and decoupled structure limits deep mixing and robustness across channels. We propose HD-RoPE, which extends RoPE from independent 2D rotations to higher-dimensional rotations and introduces a Paley-I orthogonal basis to obtain balanced, isotropic, and dense phase mixing within each rotation subspace. This significantly enhances channel coupling and rotational degrees of freedom while maintaining orthogonal stability and the relative position closure property. Furthermore, HD-RoPE is easily optimized for engineering efficiency without introducing additional trainable parameters. We have conducted extensive evaluation results demonstrating that HD-RoPE achieves significant performance improvements over standard RoPE across various popular benchmarks and in both long and short contexts.

Quantization has been widely adopted in LLM training and inference to reduce cost and improve efficiency. However, low-bit quantization of the \emph{attention} module often introduces large errors at very low bit-widths, causing performance degradation. Existing methods mainly rely on smoothing techniques to handle outliers, while we propose a hybrid quantization design to better balance accuracy and efficiency. Specifically, we propose \textbf{HyQuant}, an efficient hybrid quantization framework for LLM attention. HyQuant quantizes most attention states into low-bit formats while retaining a small set of vertical-line tokens and local-window states in high precision. These accuracy-critical regions are selected using lightweight vertical-line-aware attention-pattern signals, reducing quantization error with limited overhead. In the Prefill stage, HyQuant uses a hybrid-precision quantized attention operator that preserves vertical-line tokens and a local sliding window in full precision while quantizing the remaining context. In the Decode stage, HyQuant applies the same principle to KV-cache compression and fuses KV dequantization with attention computation to improve memory and hardware efficiency. Across diverse tasks, models, and datasets, HyQuant maintains nearly lossless accuracy with an extremely simple design, demonstrating the efficiency and practical feasibility of hybrid quantization for LLM attention. Code is available at: https://github.com/jerrysfls/HyQuant .

Long-context inference is bottlenecked by the memory footprint of the key-value (KV) cache, especially for small models under tight resource budgets. Existing KV cache eviction methods score tokens using the model's attention distribution or, in attention-free variants, each key's distance from a global reference point. Using a controlled leave-one-out probe, we find that attention magnitude is unrelated to a token's causal contribution to the answer (Spearman $ρ=-0.004$), challenging the premise behind dominant eviction methods. We introduce TwinKV, a training-free, attention-free redundancy signal that detects whether a token's key has a near-duplicate elsewhere in context. Rather than replacing existing policies, TwinKV acts as a composable repair pass: given a policy's fixed retained set, it identifies evicted tokens with no surviving duplicate (\emph{orphans}) and retained tokens whose information is duplicated elsewhere (\emph{redundant donors}), then swaps them while preserving the original budget and scoring rule. We compose TwinKV with four recent eviction policies across LongBench, LooGLE, RULER, and a short-context MMLU-Pro no-harm control at compression ratios ${0.3,0.5,0.7}$. On Qwen3-4B, TwinKV improves a majority of configurations for two policies, is near-even for a third, and helps only a minority for a fourth adaptive baseline already near a performance ceiling; gains across the three non-ceiling policies are smallest at the loosest ratio. On RULER with Llama-3.2-1B, however, that fourth policy improves in every evaluated cell because its Alone score leaves substantial room to improve. More broadly, Llama-3.2-1B shows a smaller average LongBench gain but a higher fraction of improved cells on LongBench and LooGLE than Qwen3-4B, plus a clean RULER win. We also identify few-shot classification exemplars as a task structure where TwinKV does not help on either model.

KV-cache eviction can do more than compress. In long-context LLMs, keeping only some cached tokens sometimes matches or exceeds full-cache accuracy, because many redundant prefill tokens otherwise dilute attention away from the tokens that carry the answer. This benefit is not uniform, and evicting the wrong tokens can drop accuracy to zero on tasks that require precise retrieval, so the useful question is not only which tokens to keep but also whether to evict this input at all. We show that one label-free number computed from the prefill attention, the drop between early and late layers in how much attention heads agree on which tokens to read, predicts per input, before any decoding, which of the two cases an input falls under. We build this into PAGE (Partition-Aware Gated Eviction), a wrapper that runs any SnapKV-style evictor when the drop is large and keeps the full cache when it is small, with no training, labels, or fine-tuning. PAGE is a safety mechanism rather than a compressor, so we measure it by the failures it prevents. It cuts the harm rate on capacity-bound inputs from 0.75 to 0.026, and on multi-key retrieval with Mistral-7B plain SnapKV falls from 99% to 0% as the budget shrinks, while PAGE holds it at 89%. Elsewhere, it passes the base evictor through unchanged, which is the intended behaviour and is what we observe in 8 of 16 cells. Code is available at https://anonymous.4open.science/r/PAGE-018239.

Joint text-to-video-audio generation produces synchronized visual and acoustic content, but the long sampling trajectories and heterogeneous multimodal computation of large models make inference prohibitively expensive. We present TurboT2VA, a distillation and inference framework for accelerating a 19B-parameter joint video-audio model. Large-scale T2VA distillation is challenged by modality-imbalanced optimization, the difficulty of continuous-time consistency training at scale, and the quality--diversity trade-off. TurboT2VA addresses these issues with per-modality normalization and a progressive curriculum comprising discrete consistency warm-up, continuous consistency refinement, and joint consistency--distribution matching. The curriculum first establishes a stable, diverse generation trajectory and only then introduces distribution-level refinement. On LTX-2, four-step distillation reduces generator latency from 50.52s to 2.51s at the standard evaluation resolution of 512$\times$768, achieving a 20.1$\times$ speedup while maintaining strong visual quality, audio fidelity, diversity, and video-audio synchronization. We further develop an architecture-aware inference stack that combines guarded W8A8 and fused operators, padded-text compaction, and modality-aware sparse attention while preserving dense cross-modal and text-conditioning paths. Under the high-resolution deployment setting at 1024$\times$1792, the complete stack reduces generator latency from 318.74s to 5.83s on one NVIDIA H20, achieving a 54.67$\times$ generator-only speedup. Inference code and generation demos are available at https://github.com/thu-ml/TurboDiffusion/tree/main/turbot2va.

Muon motivates designing optimizer geometry around the function of each parameter block and uses the spectral norm for hidden linear layers. For the language-model head, the spectral norm is not a faithful measure of functional change. Softmax removes shared logit shifts, whereas the spectral norm can assign arbitrarily large size to updates that change no output probability. We therefore treat the LM head and softmax as one module and derive an update geometry for their composition. Hilbert's projective distance respects this invariance as it measures the largest change in pairwise log odds. For an update $S$ with token rows $s_i^\top$, we show that the largest Hilbert distance over $\left\lVert h\right\rVert_2\leq H$ is exactly $H D(S)$, where $D(S)=\max_{i<j}\left\lVert s_i - s_j\right\rVert_2$ is the Euclidean row diameter. This diameter replaces the spectral norm in the resulting Muon-style steepest descent problem. An exact solution is possible, but its direct formulation contains one $d$-dimensional vector variable for every token pair. For a vocabulary size of approximately $50$k, this means more than one billion token pairs, making the calculation impractical at every training step. We instead impose a stronger common-ball constraint and derive projected RowNorm as an $O(Vd)$ solution. For the exact RowNorm oracle, we prove that its first-order decrease is at least $1/\sqrt{2}$ of the exact diameter-constrained optimum. With Muon on the backbone, experiments across three seeds at 190M, 380M, and 640M parameters show that RowNorm reduces mean final step diameters and empirical Hilbert RMS perturbations by factors of $45$--$60$ and $12$--$15$, respectively, with only a $0.0057$--$0.0153$ increase in mean final validation loss.

Attention dominates token mixing, but it collapses relation formation and flow allocation into a single score-to-flow step. We introduce Relation, which separates them by first organizing pairwise evidence into explicit Self and Exchange relations and deriving information flow afterward. Relation first decides whether a token should rely on itself or draw from its history, and if it draws from history, where to look. This relational organization gives rise to Full Relation, FlashRelation, Linear Relation, and Hybrid Relation. Across matched decoder-only models, Full Relation achieves lower mean final-validation NLL than MHA and reaches the paired MHA final training loss with 4.5-7.3% fewer tokens. Structural diagnostics further show that Relation learns a distinct depth organization: the first layer acts as a current-token anchor and a high-rank router, while later layers shift strongly toward history. In a fixed-context reference benchmark, FlashRelation is 4.17-5.28x faster than the materialized Full Relation implementation. Across scale-matched production workloads, it reaches 89.7-92.9% of PyTorch FlashAttention throughput while executing the exact Full Relation operator. Hybrid Relation demonstrates that Full and Linear Relation layers can be composed within a single decoder. These results support a relation-first view of token mixing: ask Self, ask Others, then let Flow follow Relation.

Model compression is critical for deploying networks on resource-constrained edge devices. While pruning-based methods can significantly reduce model size, they often suffer from abrupt performance collapse beyond a sparsity thresh-old, making it difficult to identify the feasible compression limit of the model. To address this challenge, we propose a boundary-Learning reverse regrowth framework, BRIDGE, that reformulates compression as a constructive boundary-search problem. Unlike forward pruning, our method first drives the model to an extremely sparse state to expose the collapse region, and then selectively regenerates the critical structure to restore performance. The proposed framework employs a hierarchical regeneration strategy, including coarse-grained layer selection and fine-grained regeneration parameter selection, to accurately identify which parameters require recovery. Experiments show that our method can recover models from the brink of collapse on both CNNs and Transformer architectures, demonstrating its architecture in-dependence. BRIDGE achieves a performance improvement of up to 1.49% in unstructured pruning and up to 4.77% in structured pruning. These results demonstrate that reverse regeneration can effectively extend the compression limit while maintaining stable performance. The source code is available at https://github.com/EnumaCaliber/BRIDGE.

We present a training-free method for multi-modal trajectory prediction that achieves comparable accuracy to a 57M-parameter transformer while requiring no GPU and zero learned parameters. The method builds a transition table of historical state-to-next position pairs and retrieves neighbors using a product kernel over spatial proximity, bearing, speed, and temporal context. Two inference modes operate over this shared representation: diversity-penalized sampling produces trajectories covering distinct plausible routes, while beam search finds the highest-likelihood path. On the TrAISformer benchmark (Danish Maritime AIS), our method achieves competitive accuracy at full data availability and dramatically outperforms the transformer in data-scarce regimes---remaining stable down to 10% of training data where TrAISformer degrades catastrophically. This enables deployment in new geographic regions from an order of magnitude less historical data.

Transformer-based language models are known to sometimes generalize to sequences longer than seen during training, but we lack a precise characterization of which tasks admit length generalization. It is not even known which regular languages transformers length-generalize on -- and this is a foundational class of languages. Our contributions are to establish the first complete characterization of which regular languages transformers length-generalize on and provide a decision algorithm running in polynomial time in the size of the language's syntactic monoid. These results rely on an effective characterization of the regular languages in C-RASP, a recently-established formalism that expresses which languages transformers length-generalize on. This characterization is challenging because classical tools like Krohn-Rhodes decomposition theory for finite semigroups are insufficient for C-RASP. Firstly, the basic building blocks of Krohn-Rhodes theory -- flip-flop and simple groups -- are not expressible in C-RASP. Secondly, the basic building block of C-RASP (unbounded counting) is not expressible by the finite semigroups of Krohn-Rhodes theory. Thus, length generalization on regular languages is controlled by an algebraic property that is invisible to classical finite decomposition theory. We generalize classical decomposition theory from finite semigroups to the infinite additive group on the integers, allowing us to characterize C-RASP in terms of iterated wreath products of the integers and derive a provable polynomial-time decision algorithm for regular language membership. Experiments across a broad test suite of regular languages confirm that our theory captures transformers' length-generalization behavior more accurately than existing classifications.

Rotation-based post-training quantisation commonly applies an orthogonal transform across an entire attention head to reduce outlier-induced error. RoPE instead partitions each head into two-dimensional frequency pairs, raising the question of whether a transform respecting this decomposition can improve on full-head mixing. Prior work has established the per-pair rotations that commute with RoPE. We state the converse result that, for distinct frequencies, no other single-head orthogonal map commutes with RoPE. For the head-shared parameterisation used in our experiments, we then derive the rotation angle that minimises the larger channel variance under a pooled-covariance, position-averaged surrogate and verify that the implementation attains its analytic minimum. The evaluated head-shared pairwise configuration does not improve accuracy in the tested dynamic W4A4KV4 setting. Across four checkpoints, replacing the full-head Hadamard with this configuration increases perplexity at both short and long context lengths. Composing the pairwise rotation with the Hadamard satisfies the selected $\pm0.05$-PPL interval criterion under the default estimator. Estimating the shared angle from K alone improves pairwise-only on every checkpoint but does not close its gap to full-head mixing. The analytic objective controls a position-averaged second moment of a pooled calibration covariance, whereas the dynamic quantiser sets its step from a tokenwise group range. The pairwise transform also has only two-channel mixing support. Along a controlled interpolation from two-channel to full-head mixing, K range, relative quantisation error, and perplexity degradation decrease as support increases. These results show that optimality for a structured surrogate need not reduce quantisation error when the surrogate and mixing support are misaligned with the quantiser's scale-setting statistic.

We ask whether language-model pre-training can be decomposed into smaller, independently trainable jobs that can later be recomposed into a coherent larger model. We introduce Mixture of Training (MoT), a scaffolded modular pre-training procedure that partitions a target Transformer into contiguous layer blocks, trains each block inside a frozen pretrained aligner scaffold, and then recomposes the trained blocks with an optional short end-to-end adaptation pass. On a 1.3B-parameter Gemma-style model trained on C4, MoT provides a small-scale proof of mechanism: independently trained depth slices can be recomposed into a usable language model, and a quality-parity schedule reaches the same reported perplexity as the monolithic baseline. This parity setting processes more aggregate tokens and has a shorter idealized layer-equivalent critical path after aligner preparation; its effective compute advantage depends on reusing the aligner across runs. We therefore present MoT not as a general replacement for monolithic pre-training, but as a small-scale framework for studying whether scaffolded sub-runs can act as reusable training units.