Transformers

Transformer adaptation is typically distributed across model depth, even when the intended change is narrow. We investigate how adaptation site shapes what a model learns, how well that learning generalizes, and how selectively it is applied. We introduce a controlled benchmark spanning five objectives (lexical binding, factual association, behavioral policy learning, causal mapping, and procedural reasoning) and define each objective's "adaptation geometry" as its profile of acquisition, transfer, and boundedness under full-stack and early-, middle-, or late-layer LoRA. The objectives exhibit distinct geometries. Lexical binding favors early-layer adaptation for acquisition and boundedness but requires broader updates for transfer; factual association favors later layers among localized adapters; behavioral learning separates late-layer action acquisition from middle-layer policy gating; and causal and procedural transfer benefit most from middle- or full-stack adaptation. These patterns largely persist under parameter-matched controls, and most corresponding directional contrasts replicate across five model families. These findings establish adaptation site as a key design variable for controlling what models learn, generalize, and leave unchanged.

Transformer language models are usually analyzed through vector geometry, yet ordered context and rotary position encoding introduce explicit phase structure into query-key interactions. This paper develops a bounded spectral framework for examining rotary phase alignment, hidden-state continuity, and semantic drift without treating language models as literal physical wave systems. It first identifies ordered hidden-state sequences, rather than vocabulary indices, as valid domains for spectral decomposition. It then derives the Rotary Position Embedding (RoPE) attention score as a sum of magnitude-weighted cosine terms and proves a local stability lemma: uniformly bounded phase displacement limits degradation of the corresponding pre-softmax score. To extend phase analysis beyond native RoPE coordinates, the paper defines complex modal coordinates over fixed orthonormal direction pairs and introduces a weighted coherence functional for hidden-state trajectories. These constructions support a strict distinction between representational continuity and execution-boundary admissibility. Internal coherence may describe preservation of task-relevant relations, but it cannot authorize a consequential transition. Positioned against existing geometric, spectral, phase-modulation, representation-analysis, and mechanistic-interpretability accounts, the framework contributes a theoretical and methodological program for determining when spectral structure explains continuity and when governance must remain an external predicate over execution.

Fine-grained weight pruning and activation sparsification have emerged as effective approaches for reducing the compute and memory cost of inference for Transformer models. In the moderate-sparsity regime, Gustavson's dataflow provides a natural execution model for exploiting both activation and weight sparsity on vector processors through metadata-driven indexed accumulation. However, existing RVV architectures lack native support for this pattern, forcing kernels to rely on software index decoding and L1-backed indexed memory operations that keep sparse tensor contractions far below their roofline performance bound. We present Ventaglio, a runtime-configurable sparse execution unit coupled with RVV ISA extensions that drives sparse tensor contractions toward their roofline through indexed gather-accumulate-scatter support. Integrated into an open-source vector processing cluster and implemented in 12nm FinFET, Ventaglio accelerates sparse tensor contraction kernels by $6.9\text{--}7.4\times$ over optimized RVV baselines, with only $3.1\%$ area overhead for a cluster of tightly-L1 coupled vector processing elements. We build a performance-accurate instruction-level model of the Ventaglio extension, calibrate it against RTL implementation, and leverage it for scale-out performance analysis on a large $4\times4$ multi-cluster system. Using a DuoGPT-pruned LLaMA-3-8B model with practical $40\text{--}60\%$ dual sparsity, Ventaglio achieves $2.40\text{--}5.25\times$ and $2.06\text{--}3.16\times$ speedup over dense baselines during prefill and autoregressive decoding, respectively.

The quadratic cost of self-attention makes long-context inference prohibitively expensive, and proxy-based block-sparse attention has become a practical remedy. Existing methods typically rely on a proxy to predict a binary sparse mask and a kernel to consume this mask and perform sparse attention computation. Such an approach is effective under moderate budgets. However, as the budget tightens, the estimated proxy inevitably drops some salient blocks, while the kernel can only apply the sparse mask mechanically, leading to an evident drop in model accuracy. We propose CoSA, a two-stage training-free Sparse Attention under proxy-kernel CO-design, which couples a Kernel-Aware Proxy (KAP) with an Ordered-Skipping Kernel (OSK). In the first stage, the KAP selects blocks under a moderate budget and produces an ordered mask that prescribes the order in which KV pages are visited in the kernel inner loop. In the second stage, the OSK applies this mask and skips more blocks under a tightened budget given online-softmax statistics. Across mainstream LLM backbones and long-context benchmarks, CoSA attains higher accuracy at lower budgets. Impressively, CoSA achieves a 4.93$\times$ attention speedup and reduces end-to-end Time-to-First-Token by 2.53$\times$ under a context length of 128K with negligible performance degradation. Code is available at https://github.com/Tencent/AngelSlim.

A central challenge in interpreting learned decision-making systems is to determine whether their internal representations contain concepts that help explain their behavior. We report interpretability experiments for a tokenized autoregressive Transformer agent in the Game of Hidden Rules (GOHR). We focus on a compact two-rule task in which both hidden rules map object shapes to target buckets, but with different permutations. The policy is trained on episodes sampled from these two hidden rules and then evaluated with fixed weights. It is never given a rule label and does not use an explicit rule classifier; any rule information must be inferred implicitly from interaction history. In this setting, the correct rule is not identifiable before the agent tries an informative move and observes accept/reject feedback. Sparse autoencoders (SAEs) trained on the agent's decision-token embeddings recover this structure. When held-out decisions are labeled by simple concepts such as the chosen shape or bucket, SAE dimensions that are highly selective for a concept cover most decisions where that concept is present. Individual SAE dimensions also correspond to interpretable strategies such as probing one rule hypothesis and switching after negative feedback.

Reducing training precision is a key lever for improving the e ciency of large language model (LLM) training, but pushing beyond FP8 to 4-bit oating point (FP4) remains challenging due to instability during optimization. We identify a fundamental source of this instability in existing microscaling approaches: scale inconsistency induced by tensor transposition. In conventional 1D block quantization, forward and backward passes assign di erent scaling factors to the same values after transposition, leading to biased and unstable gradient updates. To address this issue, we propose a low-precision training framework based on 2D block FP4 quantization, which enforces transposition-invariant scaling and preserves consistency between forward and backward computations. We further combine this with truncation-free scaling and stochastic rounding to control quantization error and maintain unbiased gradients. To handle the sensitivity of attention mechanisms, we adopt MXFP8 quantization for query and key projections, yielding a practical mixed-precision design. We evaluate our method on dense LLMs up to 7B parameters and a 30B Mixture-of-Experts model, trained on up to 100B tokens. Across all settings, our approach achieves stable end-to-end FP4 training and closely matches BF16 performance, with less than 1.3% degradation in perplexity and downstream accuracy. These results demonstrate that enforcing forwardbackward scaling consistency is su cient to enable practical FP4 training at scale, providing a simple and e ective pathway toward more e cient LLM training.

Token-level sparse attention, as implemented by DeepSeek Sparse Attention (DSA) in production systems, makes the downstream attention efficient but shifts the bottleneck to the indexer that feeds it. To select the top-k tokens for each query, the indexer must still score every preceding token, incurring a cost of O(L^2) per layer for a sequence of length L. We observe that this per-query scan is largely redundant: nearby queries select highly overlapping top-k tokens, and the indexer scores are long-tailed along the key axis. We exploit these properties in PIVOT, Proxy Indexing Via One full-prefix Traversal, a training-free, drop-in replacement for the DSA indexer that shares one prefix scan across a group of nearby queries. PIVOT aggregates a group into a single proxy query, performs one shared full-prefix scan to obtain a candidate set, and then selects a top-k for each query from that set. Two variants trade speed for fidelity: PIVOT-Reuse shares the proxy top-k across the group for maximum speed, whereas PIVOT-Refine re-scores the candidate set with the indexer of each query and then selects an individual top-k, matching the dense indexer at a small additional cost. A single algorithm covers both inference phases, differing only in how groups are formed: fixed-size groups of consecutive queries in prefill, and the queries decoded together in one multi-token prediction (MTP) step in decode. On DeepSeek-V3.2 and GLM-5.1 across LongBench and RULER, PIVOT matches the accuracy of the dense DSA indexer while accelerating it by up to 4x and reducing end-to-end latency by up to 1.6x at long context.

Serving large language models at long context is bottlenecked by the key-value (KV) cache, which is read in full at every decode step. Attention keys are locally low-rank though globally high-rank: shared low-rank bases discard page-specific directions that a page's own compact basis retains. LOCKS gives every page its own spectral summary (resident, about a tenth the cache's size), reconstructs within-page logits, estimates each page's attention mass by log-sum-exp, and attends only the top pages; selection itself reads no candidate keys or values. Selecting on this summary alone stays within about a point of the full cache on long-document QA (LongBench-v1), tracks the read-every-key oracle on retrieval-dense RULER down to the smallest budgets, and shows its largest margins on long-form reasoning (AIME26, MATH-500), where baseline selectors collapse. At its shipped $2048$-token budget LOCKS matches FullKV aggregate quality at $100$K$+$ context while attending about $2\%$ of the tokens, and halves per-token decode latency ($2.0\times$ at $1$M tokens) against dense attention. LOCKS ships as a drop-in plugin for unmodified vLLM, with batched decode running in full CUDA graphs.

Transformer architectures have attracted increasing attention for solving partial differential equations (PDEs), owing to their flexibility in handling irregular discretizations and their ability to capture long-range physical dependencies. However, unlike discrete language tokens or fixed-resolution image patches, observed physical fields are finite samples of underlying infinite-dimensional functions. Consequently, effectively applying Transformers to PDEs requires a tokenizer that respects the functional nature of physical fields and constructs physically expressive tokens from arbitrary discretizations.To this end, we propose \methodname{Physics Transformer}, a function-projection-based Transformer architecture for physical field prediction. Physics Transformer treats a physical field as a continuous function and partitions its discretization into locality-preserving spatial patches. Within each patch, it dynamically learns a set of adaptive local basis functions and projects the sampled field onto these bases to obtain compact physics tokens. The resulting tokens capture diverse latent physical states while preserving fine-scale spatial structures, enabling efficient global interaction through factorized attention across space and physical states. The projected representation further supports efficient decoding at arbitrary query locations. Extensive experiments on diverse benchmarks, ranging from two-dimensional PDE dynamics to industrial-scale three-dimensional CFD simulations, demonstrate that Physics Transformer accurately captures fine-grained physical structures and achieves state-of-the-art predictive performance. These results establish function projection as a practical and effective foundation for designing Transformer architectures for PDE solving.

Rotary position embeddings (RoPE) modify attention scores through position-dependent rotations, but their effect on normalized token dynamics is not captured by the vanilla spherical self-attention model. We study the continuous-time dynamics obtained when queries and keys are rotated while values remain on the unit sphere. The resulting attention kernel is reversible and admits a sharp uniform softmax floor, yet the natural RoPE interaction energy has derivatives of both signs within one fixed nontrivial system. Every consensus state remains an equilibrium, and its transverse linearization is a reversible Markov operator whose kernel depends on the consensus point through its energy across RoPE planes. On a resonant single-frequency ring we derive an exact Bessel-aliasing spectrum, including non-coprime frequencies and the correct fixed-ring large-$β$ asymptotics. Globally, closed hemispheres are invariant, while pairwise non-obtuse configurations and strict open semicircles contract with explicit half-angle and single-point tail bounds. These regional estimates instantiate a kernel-generic positivity principle with the sharp RoPE softmax floor. RoPE also selects an explicit score-flattening twisted branch; the generic resonant family is non-hyperbolic and linearly unstable, whereas an odd antipodal family becomes a hyperbolic saddle after quotienting global rotation. In multiple dimensions, the local consensus gap can depend non-monotonically on the allocation of energy across frequency planes, so no universal ordering by frequency is valid. Independent matrix, finite-difference, and nonlinear-flow computations cross-check the theorem boundaries and the reported constants.

A central question in the in-context learning literature is whether transformers can organize episode-level adaptation around different inductive-bias families. We study this question in a controlled setting through latent algorithm routing: route-like behavior in which the solver-family preference changes with the latent data-generating regime while prompt form is held fixed, remains stable under nuisance perturbations, and is selectively influenced by targeted activation interventions without large losses in answer quality. We introduce ROUTEBENCH, a diagnostic benchmark whose regimes differentially favor global shrinkage, sparsity, robustness, and locality, operationalized by ridge-like, lasso-like, Huber-like, and kNN-like family representatives. Across dense decoder-only transformers trained from scratch at 44M-612M parameters, a 306M model closes 80.9 percent of the oracle-routing gap and achieves route F1 of 84.1. The effect remains substantial under natural-language renderings, shuffled supports, lexical paraphrases, and a unified four-way routing setting. Stronger adaptive alternatives, including an input-conditioned soft mixture and an unsupervised Gumbel router, narrow the gap but remain below the 306M and 612M models on route F1 and OOD performance. Probe controls and matched activation-patching controls further show that route-relevant internal directions are decodable and functionally involved in solver-family-consistent output behavior. These results provide controlled evidence that dense transformers trained on ROUTEBENCH can develop route-like internal variables, but they do not establish universal routing in pretrained language models or unrestricted natural-language reasoning.

As the inference phase of Large Language Models (LLMs) requires handling long context windows, the Key-Value (KV) cache initially appears to address this challenge but eventually becomes a significant bottleneck as the context window continues to grow. Low-rank compression has recently been studied as an effective approach to reduce KV cache memory while maintaining model performance. However, only a few existing methods treat the Key and Value caches differently, despite their distinct roles. Moreover, these methods typically employ fixed attention-head grouping, which may not fully exploit the structural similarity among attention heads. In this paper, we propose an improved low-rank KV cache compression framework. For the Key cache, we dynamically group attention heads based on Centered Kernel Alignment (CKA) similarity and allocate the rank budget adaptively under a parameter budget. For the Value cache, we adopt the same approach as ReCalKV, refining the low-rank decomposition through offline calibration to improve reconstruction quality. Experimental results on three instruction-tuned LLMs show that our method reduces the number of Key cache parameters while maintaining competitive accuracy. We further observe that the proposed strategy is particularly effective for Multi-Head Attention (MHA) models, whereas it should be applied more conservatively to Grouped-Query Attention (GQA) models, especially in long-context settings.

Diffusion transformers are essential for high-fidelity video generation, but long token sequences make attention a dominant inference bottleneck. Training-free dynamic sparse attention alleviates this bottleneck by computing only selected key-value blocks, yet existing methods struggle to sparsify attention both efficiently and accurately for two reasons: (1) Rigid, unpredictable, and costly routing: selecting a fixed fraction of top-ranked blocks by proxy score imposes fixed budgets, whereas retaining blocks to reach a target cumulative proxy probability mass yields dynamic but potentially imbalanced budgets; both incur non-negligible overhead from computing and materializing proxy scores. (2) Lossy keep-or-drop sparsification: unselected blocks are discarded entirely, degrading accuracy under aggressive sparsity. These limitations motivate cheaper dynamic-budget routing while limiting accuracy degradation. In this paper, we introduce training-free Sol-Attn (Sparsifying online attention), which unifies dynamic routing, sparse computation, and approximation correction in a single online-softmax pass, achieving a better accuracy-efficiency trade-off in sparse attention. The core of Sol-Attn is on-the-fly block thresholding with proxy-score reuse, which selects critical blocks by comparing block proxy scores against a threshold during online softmax. This design enables dynamic yet controllable block budgets without materializing the proxy map, while directly reusing the proxy scores of unselected blocks to approximate their contribution. Experiments across image and video generation tasks show that Sol-Attn advances the quality-efficiency frontier of training-free sparse attention, delivering 2.1 times and 2.3 times end-to-end speedups for video generation and editing, respectively, while preserving visual quality.

Transformer architectures have achieved remarkable success across diverse domains; however, directly applying their standard self-attention mechanism to recommendation often yields suboptimal performance, sometimes even trailing behind well-designed simple recommendation models. In this paper, we reveal that this performance bottleneck stems from severe embedding and attention collapse unique to recommendation scenarios. The heterogeneity and long-tail nature of recommendation data lead to a severe spectral collapse dominated by a few principal singular values. We further theoretically demonstrate that this triggers a vicious cycle in recommendation model's forward and backward propagation, which accelerates embedding and attention collapse and limits the model's scaling capability with increased depth. To address these issues, we propose SpecFormer, a novel Spectral-Aware Transformer designed for mitigating embedding and attention collapse in recommendation. Specifically, SpecFormer introduces 1) a Learnable Spectral Softening module to dynamically smooth the singular values distribution of the input token embeddings; 2) a Spectrum-softened Attention mechanism to model feature interaction under a more uniform spectral distribution space; 3) a Spectral Residual Position Encoding via Taylor expansion of singular values, explicitly providing a spectral inductive bias for feature interactions. Extensive experiments on one industrial and two public datasets demonstrate that SpecFormer significantly outperforms state-of-the-art baselines. Notably, SpecFormer has been successfully deployed in a real-world commercial recommender system and exhibits exceptional scaling capabilities: stacking SpecFormer layers actively improves the attention effective rank and recommendation performance.

Randomized features provide a scalable approximation to kernel machines, but their performance depends strongly on the choice of feature distribution. We propose a particle-based method that learns this distribution by optimizing kernel-target alignment while regularizing particles with a Riesz/Coulomb repulsive potential. The resulting Hamiltonian yields diverse, task-adaptive random features and admits a mean-field description through a McKean--Vlasov equation. We instantiate the method in linearized Transformer attention by learning positive random-feature maps in a first alignment phase, then freezing the kernel and training the remaining network parameters with cross-entropy. Experiments on synthetic classification and sentence-level benchmarks show that learned kernelized attention can improve accuracy, calibration, and robustness for several feature maps while preserving linear-attention inference complexity.

Siri Expressive Voices synthesize rich, configurable speech in real time and entirely on device, powered by AFM 3 Core Advanced, Apple's most powerful on-device foundation model. This work presents the memory-efficient audio synthesis architecture behind that capability: a detokenizer that converts the semantic audio tokens emitted by the foundation model into high-fidelity audio within the tight compute and memory budget of the Apple Matrix Coprocessor (AMX). We convert semantic audio tokens to a residual vector quantization (RVQ) representation with a three-component design, a streaming encoder, a temporal decoder, and a depth decoder, that systematically decouples temporal and depth processing. A single reusable depth decoder with Diffusion Transformer (DiT)-style stage conditioning generates all RVQ levels autoregressively, replacing the dedicated per-level decoders of prior multi-decoder architectures, while causal sliding window attention with fixed-window key-value caching yields constant memory complexity independent of sequence length. Deployed on the AMX, the detokenizer sustains roughly 10 ms per generation step, about 16x faster than real time, with a peak runtime memory of only 21 MB and 329 MB of on-device assets, enabling continuous streaming synthesis of 20-320 seconds of audio. This constant, small footprint replaces the linear and quadratic memory scaling of conventional transformer- and GAN-based approaches. Ablation studies validate the key architectural components, and audio quality assessment confirms that the architecture maintains synthesis fidelity while achieving efficiency gains over existing methods. Operating at a 1-billion-parameter activation size within AFM 3 Core Advanced, it improves Mean Opinion Score by +0.28 overall (4.15 vs. 3.87) and by +0.42 on conversational speech (4.24 vs. 3.82) over the prior on-device text-to-speech system.

LoRA fine-tuning can create intruder dimensions: new leading singular vectors of the updated weight matrix $W+BA$ that are nearly orthogonal to all pretrained singular vectors and that drive catastrophic forgetting. Since their discovery, no theory has predicted, layer by layer on measured spectra, when they appear. We derive a per-layer critical update strength $s^\ast=\barθ/(γσ_1(BA))$, computed from the measured spectrum of $W$ alone through the rectangular spiked-deformation transform, together with an exact secular-equation characterization of the updated spectrum, with no fitted parameters. In a pre-specified study spanning four dense Transformer families, a state-space model, a mixture-of-experts model, and an encoder-decoder (18 adapters, 9{,}840 layer scans), the law localizes the empirical threshold within a factor of two on $82\%$ of layers, separates intruder-bearing from intruder-free layers at deployment with a mean AUC of $0.89$, holds unchanged on six third-party adapters, and predicts where WikiText-2 perplexity begins to degrade; a combination of the two pre-specified edge evaluations reaches $98\%$ and is confirmed out-of-bag on the external adapters ($0.997$). Full fine-tuning disperses its update far below the threshold of every layer, which resolves the asymmetry between LoRA and full fine-tuning. Norm-matched interventions confirm that threshold-crossing layers, rather than update magnitude, carry the forgetting, and a spike-budget rule derived from the thresholds, requiring one SVD and no validation sweeps, reduces forgetting by $62\%$ on the most fragile model at no task cost.

Fully homomorphic encryption (FHE) provides strong cryptographic guarantees for private inference, but deploying transformer models under FHE remains prohibitively expensive. A key bottleneck is that non-linear operations such as softmax, normalization, and activation must be replaced with polynomial approximations compatible with the CKKS scheme, and the multiplicative depth consumed by these approximations dominates inference cost. Recent frameworks have advanced approximation techniques, yet all rely on manually configured approximation hyperparameters (e.g., number of iterations, polynomial degree), applied uniformly across all layers. While convenient, this uniform-configuration approach is overly rigid: different layers can tolerate different levels of approximation error without degrading predictive accuracy, and uniform configurations cannot exploit this variability to reduce latency. Allowing each layer to adopt its own configuration, however, causes the search space to explode with model depth, reaching roughly $10^{84}$ configurations for BERT/ViT (12 layers) and $10^{225}$ for LLaMA3 (32 layers), rendering manual exploration practically impossible. We present ATLAS, an automated framework that configures per-layer approximation settings by formulating the problem as a multi-objective optimization over latency and predictive accuracy. The resulting problem is inherently difficult: 1) competing objectives over a large decision space (120 or 320 variables for BERT/ViT or LLaMA3); 2) expensive evaluation, as each configuration takes 70-1,000 seconds even in cleartext; and 3) sparse optimization signals, as 35-50% of candidate configurations yield numerically invalid solutions. ATLAS addresses these challenges through a two-stage optimization strategy that progressively relaxes layer-wise constraints, combined with surrogate models to accelerate evaluation.

Large-scale calcium imaging has created an opportunity to build foundation-style models for neural population dynamics, but a central question remains unresolved: \textbf{whether a model pretrained on one collection of recordings can generalize to new datasets, experimental paradigms, and even species.} Existing approaches are often designed for specific tasks and evaluated on a single dataset, making it unclear whether their learned representations are reusable for new calcium trace datasets. To tackle this gap, we present \textbf{CAPT}, a \textbf{C}ontinuous \textbf{A}utoregressive \textbf{P}opulation \textbf{T}ransformer for calcium population dynamics. CAPT models continuous calcium traces directly through a continuous patch tokenization strategy and is trained autoregressively, enabling end-to-end pretraining and adaptation to diverse downstream tasks. We first pretrain CAPT on a large-scale mouse calcium imaging dataset and evaluate its transferability across independent mouse, larval zebrafish, and \textit{C. elegans} datasets collected by different laboratories. In these transfer settings, the pretrained backbone is frozen and only adaptation modules are updated. Across neural population forecasting and behavior decoding tasks, CAPT consistently outperforms specialized and general-purpose baselines. Alongside predictive performance, multimodal analyses using NeuroPAL annotations in \textit{C. elegans} datasets show that CAPT embeddings form a shared functional space across datasets and capture anatomical cell-identity-related structure. These results suggest that the continuous autoregressive modeling opens up possibilities for a simple route towards general-purpose neural foundation models for calcium imaging, which can generalize across datasets, experimental paradigms, and species. Code is available at https://github.com/TSuXinH/CAPT.

Modeling multi-entity temporal data requires capturing dependencies across entities, time, and their interactions. Transformer-based approaches perform well but often rely on deep stacks of layers to learn these heterogeneous dependencies implicitly, increasing computational cost. We revisit this problem from a structural perspective and decompose multi-entity temporal dynamics into three interaction types: spatial interactions among entities, temporal interactions across time, and cross interactions coupling the two domains. We propose a structured spatio-temporal transformer block that explicitly models all three within a single stage. It uses parallel spatial and temporal self-attention, followed by bidirectional cross-attention, and combines the outputs through learnable gated fusion. By directly encoding these complementary views, the model reduces the need for deep stacking. We evaluate the approach on video-based group activity recognition, skeleton-based human interaction analysis, and wearable sensor-based activity recognition. Despite its simplicity, the single structured Transformer block matches or outperforms deeper architectures with only 1.76M parameters. The results suggest that depth in prior models partly compensates for implicit and entangled interaction modeling, whereas explicit factorization offers a more efficient and transparent alternative. More broadly, this work supports a structure-first design principle: expressive multi-entity temporal reasoning can emerge by exposing interaction structure rather than relying on depth.

Multi-head Latent Attention (MLA), introduced in DeepSeek-V2, compresses key-value pairs through a shared low-rank bottleneck (cKV), achieving 81% KV-cache reduction during inference. Despite its adoption in massive production models, no prior work has studied what information this bottleneck preserves or discards, nor how it reshapes internal transformer circuits. We present the first comprehensive mechanistic interpretability study of MLA, training a 114M-parameter transformer (pretrained on a web/code/math mixture, fine-tuned on TinyStories) and analyzing its representations through SVD, attention head taxonomy, linear probing, and a disruption-attribution analysis. Our key findings are: (1) the cKV bottleneck learns a pure content representation, preserving entity identity (98% retention) while discarding positional information, validating MLA's separation of content from position via RoPE; (2) induction heads co-locate at a single layer (Layer 12), unlike their distributed formation in standard MHA; (3) a single "semantic hub" layer (Layer 15) simultaneously exhibits the highest SVD effective rank and strongest disruption-attribution score; and (4) the bottleneck is globally over-provisioned, using only 46% of its capacity on average. These findings suggest MLA does not merely compress attention passively, but reshapes how the model organizes content, position, and circuit structure. We view this as an initial data point and detail scope limitations in Section 5.

Neural scaling laws describe how loss decreases as models, data, and compute grow, but they do not answer a prior question: for a fixed task, what is the minimum model capacity required to solve it? We study this through the Entropic Bound, a spectral notion of task-intrinsic capacity for Transformers. We first prove that, in a linear attention surrogate, the intrinsic rank $r^*$ of the token-mixing operator is a tight lower bound: any rank-deficient model incurs unavoidable excess risk, and the bound is achievable at $r^*$. We further show that gradient descent recovers this rank under standard low-rank implicit-bias assumptions, confirm all three properties empirically, and show $r^*$ is recoverable from data before training. We then ask whether this transfers to real attention. A naive transfer fails, and a controlled interpolation ladder localizes the cause precisely: it is not softmax and not a rank constraint, but the input-conditioned nature of attention's mixing operator, which a static weight kernel cannot summarize. Motivated by this, we introduce an attention-native intrinsic rank -- the minimum query-key kernel rank realizing the task within the attention class -- and show that under this definition the full Entropic Bound structure (deficiency, achievability, recovery) is restored for both linear and softmax attention, with the energy effective rank as the estimator robust to softmax distortion. Finally, we map the boundary of data-only predictability: $r^*$ is exactly recoverable for linear QK attention, even without the value map at scale, while softmax attention admits only partial pre-training recovery due to nonlinear inversion and kernel-value identifiability effects. Our results reframe the Entropic Bound from a post-hoc descriptor into an attention-native capacity measure with a precisely characterized predictability frontier.

Automated driving systems (ADSs) are becoming ubiquitous. Future Software Defined Vehicles (SDVs) may be able to run multiple ADSs, both native and aftermarket such as Comma.ai's Openpilot. Monitoring systems to independently verify which automated driving system is active are important for safety monitoring, regulatory compliance, insurance assessment, and anomaly detection. In this paper, we first evaluate the effectiveness of three sequence-based classification models: Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM) networks, and a Transformer encoder model for identifying Level 2 automated driving systems using vehicle telematics data alone: Comma Openpilot, Tesla Autopilot, and Cadillac Super Cruise, along with manual driving. All three models achieve strong clean-data performance with macro F1-scores of 0.92 (GRU), 0.90 (LSTM), and 0.93 (Transformer encoder model) when trained on clean data; threat-matched training yields 0.904-0.916 macro F1 with only a modest clean-data penalty. Second, we introduce a modular robustness evaluation framework that simulates realistic telematics degradation through five corruption families at five severity levels (L1-L5). Continuous channels are perturbed using additive white Gaussian noise with cumulative drift, correlated cross-channel noise, and temporal jitter. Binary event signals are subjected to burst loss, delayed transitions, spurious toggles and cross-feature inconsistencies inspired by communication errors. Robustness is measured using macro-F1, which gives equal weight to each class and is suitable for imbalanced multiclass evaluation. Our evaluation reveals a sharp failure-mode split: event-level corruptions reduce macro-F1 only slightly (greater than equal to 0.87 at L5), while temporal jitter collapses macro-F1 to 0.44-0.50 across GRU, LSTM, and Transformer encoder model.

Weights and biases are normally optimized as separate parameter tensors, yet they do not represent separate functions when the input to an affine layer has nonzero mean. For an affine map $z=Wx+b$ with input mean $μ$, a weight update contains a sample-independent displacement $ΔWμ$ that is functionally indistinguishable from a bias update. We call this hidden contribution \emph{boundary motion} and decompose each update into a centered, sample-varying \emph{shape} component and a shared \emph{boundary} component. On a four-layer Transformer trained from scratch on IMDb, the bias-like term $g_bμ^\top$ has a median norm equal to 0.664 of the raw weight-gradient norm across affine layers and training checkpoints. More strikingly, the median ratio $\norm{ΔWμ}/\norm{Δb}$ is 134.7, while $\norm{ΔWμ}/\norm{Δb+ΔWμ}$ is 0.994. Thus, under AdamW, the observed boundary motion is almost entirely realized through the weight matrix rather than the explicit bias. We implement a diagnostic optimizer, Shape--Boundary Orthogonal AdamW (SBO-AdamW), that optimizes $g_W-g_bμ^\top$ and $g_b$ with independent Adam states and compensates the weight-induced boundary displacement. In a single-seed experiment, SBO-AdamW raises validation accuracy from 81.68% to 85.81% and validation-selected test accuracy from 78.73% to 82.73%, with the best validation checkpoint occurring at step 800 instead of step 3000. However, the moving-batch-center compensation produces severe bias-coordinate drift and strongly reduces boundary energy. The present evidence therefore supports hidden boundary motion as an important optimization mechanism, but it does not yet establish a final general-purpose optimizer. A stable centered-affine parameterization is identified as the required next step.

Quantum state preparation is a key component of many quantum algorithms. Performing this step efficiently is essential for realizing practical quantum advantage in quantum chemistry applications. Iterative algorithms like ADAPT-VQE can produce shallow ground-state preparation circuits, but become computationally prohibitive for the larger molecules relevant to materials science and pharmaceutical development. Here, we introduce ADAPT-GQE, a generative AI framework that learns to synthesize ground-state preparation circuits for electronic structure calculations. We first use ADAPT-VQE to generate high-quality reference circuits, which are then used as targets for training models for circuit generation. Once trained, the model can efficiently propose and score circuits, enabling reinforcement learning (RL) to drive circuit generation accuracy beyond the accuracy of the ADAPT-VQE training data. This pipeline achieves order-of-magnitude reductions in circuit generation time relative to ADAPT-VQE while maintaining comparable or improved state-preparation accuracy. We demonstrate ADAPT-GQE on imipramine, a well-established tricyclic antidepressant that serves as a representative, challenging target for computational modelling in drug stability protocols. We execute generated circuits on Quantinuum Helios-1, representing a milestone for AI-generated quantum chemistry circuits on state-of-the-art quantum hardware. These results establish a pathway toward automated quantum circuit synthesis for utility-scale quantum computational chemistry.

Accurate geological characterization of subsurface reservoirs from well log data is essential to support projects such as carbon capture and storage (CCS), geothermal development, and extraction of natural resources. Existing automated techniques for geological characterization primarily use sliding-window classification, which limits their ability to understand broader geological contexts, often leading to misaligned formation layers. To overcome these limitations, we introduce LithoFormer, a robust framework for stratigraphic inference using a Seq2Seq transformer model that ingests entire multivariate well logs in a single pass. The framework utilizes a channel-independent PatchTST backbone enhanced with rotary positional embeddings (RoPE) to capture long-range geological dependencies across entire multivariate well logs. A decoupled multi-task head is employed to jointly predict geological zonation and precise boundary probabilities, while a geology-informed loss function enforces physical constraints such as the Law of Superposition. Validated and deployed on three real-world datasets, LithoFormer demonstrates a 90% reduction in median boundary error and eliminates stratigraphic order violations compared to traditional sliding-window baselines. It also achieves a 80% reduction in manual expert labor and eliminates stratigraphic inconsistencies, providing a scalable and reliable solution for large-scale subsurface modeling.

With the rapid adoption of long-context large language models (LLMs), the continuously growing KV cache during decoding has become the critical memory bottleneck. To tackle this challenge, we propose HiKV, a novel algorithm-hardware co-design that exploits KV cache redundancy through hierarchical importance awareness. Algorithmically, HiKV compresses the KV cache at two granularities: Stage I evicts unimportant tokens within a fixed budget, and Stage II further loads only the significant elements of each retained token, reaching compression ratios unattainable at a single granularity. Architecturally, we develop a dedicated accelerator centered on a reconfigurable importance sorter that switches between the distinct sorting datapaths each stage requires, unifying the two-stage acceleration in one circuit with minimal overhead. Evaluated on representative LLMs, HiKV achieves up to 7.95x speedup and 90% energy reduction in the attention computation over the vanilla KV cache baseline within negligible 1% accuracy loss. Under iso-accuracy constraints, HiKV outperforms state-of-the-art importance-based methods by achieving an additional 1.82~4.87x reduction in external memory accesses. These benefits are enabled by specialized hardware components that add only 8% to the system area.

Feature-attribution methods assign scores relating input variables to a model's output, but do not by themselves characterize how explicitly defined interaction operators compose across its intermediate layers. We introduce \emph{interior interpretability}, a propagation-based perspective on internal model organization, and instantiate it for tabular Transformers using attention rollout. We interpret rollout as a row-stochastic operator encoding attention-mediated propagation between feature tokens. By applying classical Doeblin--Dobrushin contraction theory, we show that a rollout operator with a small Dobrushin coefficient is quantitatively close to a rank-one stochastic matrix whose common row is determined by its normalized column sums. This result gives a structural interpretation to the corresponding rollout propagation profile. In Transformers trained for metabolomic age prediction, the measured rollout contraction strengthens with depth. Trained and randomly initialized models also exhibit different propagation profiles, although the present experiments do not establish the predictive relevance of individual rollout-ranked variables. Exploratory comparisons with PCA and GradientExplainer approximations to SHAP reveal localized agreement among highly ranked variables but weak agreement across complete rankings. Attention rollout is therefore used here as a diagnostic of attention-mediated propagation, not as a causal explanation or faithful attribution of the complete Transformer.

In production processes for consumer products, assembly instructions are essential not only for planning but also for executing the production process. Likewise in robotics, it is crucial for an assembly robot to understand how components fit together and can be assembled. To facilitate these tasks, we contribute a method for constructing scene graphs to represent and characterize assembly relationships between components. Our approach does not rely on semantic data and is capable of handling a very small dataset. To realize this, the output of a Faster R-CNN model is used to create geometric representations, which are then processed by a transformer architecture to generate an adjacency matrix. This matrix serves as input to a Siamese network that uses message passing based on an attentional graph convolutional network (aGCN) architecture to characterize the connections between the components. We validate our method on a study dataset of toy model components which can be assembled into transportation vehicles.

Forecasting multiple time-series with high-dimensional covariates presents a core challenge: unifying common temporal patterns while retaining meaningful series-specific information. We introduce Hopformer (Homogeneity-Pursuit Transformer), a two-stage framework that addresses this challenge. In the first stage, we perform a Sparsity Pattern Aggregation (SPA) scheme extracting a common low-variance trend that incorporates the covariates. This acts as a homogenization layer. In the second stage, a LoRA-fine-tuned Transformer models the remaining complex dependencies in the residual. Our method is theoretically grounded. We prove that SPA achieves a near-optimal bias-variance trade-off via an oracle inequality. We also provide generalization bounds for the second stage under dependent time series data. Hopformer sets a new state of the art, improving MASE by an average of 6.56% across synthetic and real-world forecasting benchmarks.

In this work, we explore how the inference time of a Transformer Neural Network can be efficiently optimized with applications to real-time anomaly detection in financial time series. The financial time series are price series such as asset prices. Unfortunately, the data is often with errors or outliers that make the downstream data processing tasks useless, unstable or even harmful. Moreover, the amount of financial time-series data has been significantly increasing. Hence, there is a need for better data-cleaning methods in terms of accuracy and in terms of processing speed. Transformers as a neural network architecture have achieved superior performances in many tasks such as Natural Language Processing and Computer Vision. Time series modelling and especially anomaly detection tasks can benefit from the features of transformers architecture in multiple ways, including the capacity to capture long-range dependencies and interactions. Increasingly powerful hardware, such as field-programmable gate arrays (FPGAs), have seen increasing usage in recent years due to their reconfigurability and high performance. They can be efficiently utilized to speed up the computations of the Transformer architecture. We explore different Transformer architectures for time series modelling and how they can be efficiently implemented on an FPGA board (PYNQ-Z2). In particular, we examine the application of Transformers to detect anomalies in time series and we show how they can be efficiently implemented on an FPGA board to minimize latency. The code is available at https://github.com/thxi/icl_thesis

Hand gesture recognition via surface electromyography (sEMG) is fundamental to prosthetic control. In this field, deep learning approaches have become the gold standard. However, current architectures struggle to scale; model performance typically decreases as the number of hand movements increases. Performance degradation is tied to the increased statistical complexity of decoding expanded gesture sets and compounded by the limitations of state-of-the-art methods, which primarily rely on low-latency unimodal convolutional architectures. Convolutions operate locally, limiting model's ability to capture long-range sequential patterns. Unimodal setups cannot leverage complementary information from coordinated signals characterizing movement execution, such as inertial and eye-tracking data. These limitations motivate architectures that integrate local and global features across multimodal physiological sequences. To bridge this gap, this study introduces EMG-CrossFormer, an end-to-end hybrid convolutional-transformer for seamless multimodal integration. EMG-CrossFormer combines representations from an arbitrary number of unimodal encoders through cascaded cross-attention fusion layers, and decodes the fused representations using learnable gesture queries. EMG-CrossFormer was evaluated on four NinaPro datasets (DB2, DB3, DB7, and DB10) and benchmarked against six state-of-the-art models using an increasing number of modalities. Using only sEMG, EMG-CrossFormer achieved mean accuracies of 72.33%, 52.48%, 79.16%, and 73.49% on DB2, DB3, DB7, and DB10, respectively. Incorporating inertial signals improved performance to 90.66%, 80.40%, 92.79%, and 92.06%. These results show that joint local-global feature modeling improves sEMG-only decoding and that multimodal fusion substantially amplifies this benefit, underscoring the value of both design principles for complex hand gesture recognition.

Full self-attention in large language models scales as O(N^2), which limits long-context document analysis to 65,536 tokens and requires costly GPU clusters. The Reduced Interaction Sampling (RIS) inference engine addresses this constraint as a model-agnostic architecture. Without modifying weights, RIS reduces self-attention complexity to O(N log N) using sparse stochastic geometry that fits within commodity memory limits. We validate RIS on Qwen2-1.5B-Instruct across two regimes. In controlled evaluations at 32,768 tokens (where native dense attention serves as the upper bound), RIS-Stochastic at 1% density and 70 ensemble seeds achieves 75.00% accuracy, outperforming the native dense baseline (71.88%), while RIS-Stochastic at 5% density and 10 seeds matches it (71.88%). This demonstrates that sparse attention acts as a regularizer: low density (1%) over multiple seeds filters out sequence-level noise, whereas higher density (5%) reintroduces distractor noise. Under the tightest budget, RIS-Structural reaches 68.75% accuracy at 1% density with just 10 seeds, recovering 75% of the contextual gap relative to the zero-context floor (59.38%). At 65,536 tokens, where dense attention triggers out-of-memory faults, RIS yields retrieval gains of up to 14.06 percentage points over the zero-context floor (51.56%), which is confirmed as marginally significant under McNemar's paired test (p = 0.078 < 0.10). All evaluations run on commodity, unaccelerated CPU servers (16-128 GB of RAM), demonstrating that long-context LLM inference is feasible on standard academic hardware without GPU acceleration.

Data-adaptive sparse attention masks substantially outperform fixed patterns (e.g., BigBird and Longformer) and can even exceed dense attention on long sequences. Existing adaptive approaches---including SBM-Transformer, Dynamic Mask Attention, and NSA---typically require additional learnable parameters, custom gradient estimators, or specialized CUDA kernels. We show that classical data compression provides an effective masking signal with \textbf{no additional parameters}. By computing per-block gzip compression ratios, we identify non-redundant content blocks and route long-range attention selectively through them. Intuitively, blocks that gzip cannot compress contain information not predictable from local repetition, making them natural long-range attention targets. Because the compression profile is input-dependent, the resulting sparse mask adapts dynamically to content without learned parameters, auxiliary losses, or custom kernels. On PG-19 byte-level language modeling at 92M parameters with 8K context, our method achieves 1.71 bits-per-byte (BPB), outperforming dense attention (2.89), BigBird (2.34), Longformer (3.21), and a reimplemented SBM-Transformer (3.38)---the only learned-mask baseline---by up to 1.67 BPB while adding no parameters. The advantage grows with sequence length, with the gap over BigBird widening from 0.05 BPB at 4K context to 0.63 BPB at 8K, while convergence is 3.3$\times$ faster.

Transformers are widely used across many domains, including natural language processing, computer vision, web search, and DNA sequence analysis. Given their broad applicability, improving the performance of transformer models is critical. However, the high volume of data movement between processing units and memory during attention operations significantly limits their efficiency. Processing-In-Memory (PIM) mitigates this issue by performing computations directly inside memory. While prior work has proposed PIM-based transformer implementations, they suffer from costly inter-bank communication, and struggle to scale due to the limited capacity of memory banks. As a result, attention-related data must be split across banks, diminishing the potential benefits of PIM. In this work, we propose RED-PIM, an algorithm-architecture co-design that reduces attention latency by minimizing inter-bank data movement from O(N^2) to O(N) and shrinking intermediate attention matrices from N x N to d x d. By reorganizing matrix operations, performing computations locally, and employing an optimized data transfer strategy, RED-PIM significantly reduces computation cost and interconnect traffic. Compared to baseline PIM implementation, RED-PIM achieves inference time reductions ranging from 16.05% to 99.99% (geometric mean of 66.42%), with the largest gains on longer sequences. On real-world datasets, RED-PIM improves performance by 99.60% for long documents and 13.44% for shorter ones, while maintaining or improving accuracy. These results demonstrate RED-PIM's effectiveness for scalable and efficient transformer inference.

Large language models (LLMs) achieve strong generation and reasoning performance, but the Transformer architecture incurs high inference cost. Existing acceleration methods often rely on task-specific fine-tuning or training from scratch, increasing adaptation cost and limiting cross-task usability. We present an Adaptive Depth Sparse Framework (AdaDSF) that converts off-the-shelf pre-trained LLMs into depth-sparse models without full retraining. Our key insight is that layers contribute unequally to representation transformation, characterized by the cosine similarity between layer input and output hidden states. Based on this, AdaDSF assigns layer-wise token retention ratios from similarity statistics, uses a lightweight router to select informative tokens at each layer, and introduces a feature-preserving alignment objective to match intermediate and final representations between sparse and dense models. On GPT-NeoX and Qwen2.5 over language modeling and commonsense reasoning, AdaDSF substantially reduces inference FLOPs while preserving performance close to dense counterparts. Under comparable sparsity, AdaDSF consistently yields smaller accuracy degradation than strong baselines including MoD, D-LLM, and DLO.

Equivariant networks embed geometric symmetries as structural priors through weight sharing, achieving remarkable parameter efficiency across vision tasks. However, this parameter efficiency does not translate into compute efficiency: existing implementations unroll the structured weights into dense matrices and dispatch them to generic dense kernels, so the FLOPs of an equivariant layer are no smaller than those of a non-equivariant counterpart. In this paper, we observe that the equivariant linear (EQ-Linear) layer---the most fundamental and frequently used module in modern equivariant architectures---is essentially a circular convolution along the group dimension composed with a linear transform along the channel dimension. Building on this observation, we propose Flash EQ-Linear, an exact acceleration algorithm that reduces the complexity from $\mathcal{O}(NDC)$ to $\mathcal{O}(NDC/T)$ by combining the Fourier convolution theorem along the group dimension with the conjugate symmetry of the real DFT. We further provide dedicated CUDA kernels for Flash EQ-Linear, covering both forward and backward passes and both FP32 and FP16 precision. At the operator level, Flash EQ-Linear achieves up to ${2\times}$ forward speedup over PyTorch's F.linear; at the network level, Flash EQ-ViT and Flash EQ-Swin achieve up to ${1.7\times}$ end-to-end speedup over both equivariant and non-equivariant baselines. To our knowledge, this is the first time equivariant networks strictly dominate their non-equivariant counterparts along all three axes simultaneously: accuracy, parameter efficiency, and inference speed.Code is available at https://github.com/zhongchenzhao/FlashEQLinear.

Outsourced Transformer inference exposes clients to model substitution and incomplete execution, while direct replay removes the computational benefit of delegation. We present GKR-HND, a registered-model protocol for verifying the polynomial backbone of Homomorphic--Nonhomomorphic Decomposition Transformers. The retained verifier checks the GKR transcript and registered-weight openings, but delegates expensive public evaluations to an assigned computation worker. Assuming an honest retained verifier and prover--worker non-collusion, the verifier accepts only when the worker's signed, request-bound response agrees with the proof claims. Experiments with pretrained HND models validate the proof path and the delegated public computation without dense-matrix replay.

Language models are thought to exhibit the phenomenon of superposition, representing many more features than dimensions in their residual streams. Sparse autoencoders (SAEs) are designed to recover such features post-hoc, but training models that are interpretable by construction has remained impractical, as a per-layer over-complete bottleneck is prohibitively expensive in both memory and compute. To overcome this issue, we introduce the ParityTransformer, a GPT-2-scale architecture whose intermediate representations are efficient and wide / sparse by design. At each layer, a Deep Parity Bottleneck (DPB) replaces a learned over-complete basis with a parameter-free algebraic dictionary, providing a deterministic incoherence guarantee and eliminating the memory requirements that have prevented per-layer interpretable bottlenecks at scale. A DPB is a hierarchically structured sparse bottleneck which efficiently enforces sparsity using a multi-level mixture-of-experts approach: a hardware-aware implementation that closes the cost gap between activation sparse and dense training to a manageable interpretability tax. Empirically, ParityTransformers perform at least as well as post-hoc SAEs on sparse probing tasks, while out-performing on measures of feature absorption, steering effectiveness, and fine-grained causal interventions. Because subsequent computation acts only on features that survive the sparse bottleneck, the ParityTransformer's features are native to the model's forwards pass by construction, addressing the question of whether SAEs probe features the model actually uses during computation. We see this as a step toward training models whose internal representations are interpretable by design rather than recovered post hoc.

AI researchers describe state-of-the-art models as one thing repeated at scale: the Transformer, wired identically for text, pixels, or speech. Neuroscientists describe the cortex as a mosaic - dense Layer 4 in visual cortex for spatial encoding, thick Layers 5/6 in motion cortex for temporal integration - different jobs solved by different structures. This paper argues the gap is a structural error, not a stylistic one, and is measurable. A century of cytoarchitecture, from Brodmann to single-cell Patch-seq, shows distinct cognitive functions are implemented by qualitatively different structures, not by rescaling one template. The convolutional neural network is the field's own proof: local receptive fields and hierarchical depth encoded this prior directly, reaching strong image recognition on far less data than later architectures needed. The paper traces how this lesson was discarded: the "Hardware Lottery" made the Transformer the path of least resistance, not the principled choice, and Mixture-of-Experts, often cited as diversity, in fact partitions parameters among identical experts. A functionalist analysis shows the Transformer is best understood as a functional analog of the hippocampal formation, not a general-purpose cortex - the same mistake as treating cortex as one giant Broca's area, except the field has now standardized on a giant hippocampus, applied to tasks it was never built for: audition, executive gating, working memory. The paper closes with an alternative: a Heterogeneous Topological Network, a System of Systems in which distinct modules keep the inductive bias their computation demands and communicate through standardized interfaces. This is a design discipline for AI architects, not cognitive science: specify modularity before training, using structural evidence as a design input rather than reverse-engineering architecture from a trained model's behavior.

Structural generalization has been measured repeatedly by several benchmarks, yet it has never been formally defined. We give a definition that translates the two premises (compositional structure and unbounded generalization) into mathematical language. The definition itself is neutral: a compiler that hard-codes the rules satisfies it just as well. But structural generalization becomes a scientific question only insofar as the capacity can autonomously emerge from finite data. This question pits the computational lower bound $\mathrm{NC}^1$ against the learnable ceiling $\mathrm{TC}^0$ of pure Transformers. Under a Montagovian instantiation, each compositional rule splits into two projections: a syntactic face ($F_γ$) and a semantic face ($G_γ$). Tree evaluation on the $G_γ$ side is an instantiation of BFVP, which is $\mathrm{NC}^1$-complete (Buss, 1987). A pure Transformer must learn both faces at once, but Kraus et al. (2026) prove that its learnable class $\subseteq \mathrm{TC}^0$. Under the standard assumption $\mathrm{TC}^0 \neq \mathrm{NC}^1$, a pure Transformer cannot learn structural generalization. Neuro-symbolic systems achieve the best benchmark scores precisely because they inject $G_γ$, sidestepping the genuinely hard half. Benchmark scores cannot distinguish "learned" from "given." This is what this paper sets out to make clear.

Controllable image generation remains challenging for creative professionals, who often require precise regional control over materials, object identities, and spatial arrangements that cannot be reliably achieved through text prompting alone. Diffusion Transformers (DiTs) can natively ingest heterogeneous tokens stemming from texts and images, but they lack mechanisms for determining where and how these tokens should influence the output. We introduce appearance pointers, compact tokens that guide DiTs toward the correct appearance cues at the correct spatial locations by aligning text or image inputs with user-specified masks. Appearance pointers are produced by a region correspondence network and refined through a spatial aggregation mechanism, enabling the model to handle multiple regional descriptions without significantly increasing token load. Our approach introduces the first modality-agnostic interface for localized multimodal control in a DiT without retraining the base model from scratch. Across a range of metrics, our single model reaches or surpasses the performance of modality-specific state of the art methods, offering a simple and extensible path toward precise, region-aware, multimodal guidance in generative image synthesis.

We derive four memory-optimal inference artifacts for transformer attention using the Mathematics of Arrays (MoA), each following directly from the forward-pass Denotational Normal Form (DNF) of with the query-row index fixed to the current decode step. The artifacts are: (1)~a single-query decode DNF in which the $ψ$-reduction eliminates the $K^\top$ buffer algebraically, achieving $(d_k + nd_k+ nd_v+ d_v)\times4\,{B}$ Dynamic Random Access Memory (DRAM) traffic result numerically verified to $\|{err}\|_\leq2\times10^{-7}$; (2)~a C/OpenACC Graphics Processing Unit (GPU) kernel with Operational Normal Form (ONF) stride arithmetic and hardware-coalesced memory access, verified to $\|\mathrm{err}\|_\infty=0$ (exact IEEE-754 floating-point arithmetic); (3)~a multi-step KV-cache with $O(d_k+d_v)$ per-step append via MoA concatenation $\#$; and (4)~Grouped-Query Attention (GQA) and Multi-Query Attention (MQA) derived via $ψ$-selection, achieving a proven $\frac {h_q} { h_{kv} }$ reduction in KV traffic. All programs are verified against PyTorch scaled_dot_product_attention.

Sequential decision making in non-stationary and partially observable environments requires rapid adaptation to latent regime changes. However, existing Transformer decision models face a structural bottleneck in the retrieval mechanism: even when reward is used for training or exposed as an input token, attention retrieval remains primarily driven by observation-derived similarity. We formalize this limitation as feedback-blind retrieval, and formally show that, on feedback-informative tasks, observation-equivalent histories with different action-reward outcomes cannot be distinguished by any observation-only attention, resulting in suboptimal choice. To address this mismatch, we propose the Utility-Augmented Transformer (UAT), a new feedback-conditioned retrieval attention architecture in which a compact utility state modulates the query, key, and value projections, allowing action-reward history to directly alter context retrieval during the forward pass. UAT also enjoys an exact zero-gate degradation property that recovers the Vanilla Transformer when feedback is uninformative. Under finite-horizon compactness and Lipschitz assumptions, we prove that UAT strictly enlarges the observation-only Transformer class and can uniformly approximate feedback-dependent decision maps. Across four non-stationary benchmarks: synthetic navigation with hidden goal shifts, non-stationary sepsis treatment, cross-market portfolio allocation, and delayed-feedback recommendation, UAT consistently improves performance over observation-only, test-time adaptation, and input-level feedback baselines, with particularly large gains in noisier regimes that require stronger adaptation.

This paper explores Relative Positional Encoding (RPE) as an additive bias in Transformer architectures to solve the Team Orienteering Problem. By embedding in the attention mechanism pairwise spatial relationships among nodes of the graph that represents the routing problem, the transformer encoder can compute a richer spatial-aware graph embedding that allows the decoder to estimate better routes. Experimental results involving instances up to 100 nodes demonstrate consistent improvements in collected rewards and optimality gaps over vanilla Transformer architectures used by other state-of-the-art works. These findings highlight that explicit relational modeling significantly enhances scalability and generalization for complex combinatorial optimization.

In the \emph{latent posterior model} of transformer behavior, the next-token distribution arises from a posterior over latent predictive models conditioned on the context, mixed to generate continuations. We exploit this model in settings where it is exact, namely Bayes-filtered transformers (BFTs) meta-learned on sequences from a hierarchical prior, to introduce \textbf{Posterior Prefix Tuning (PPT)}, a new method for \emph{eliciting} behavior from a transformer: given a utility function on continuations, find a prompt under which the transformer generates continuations of high expected utility. For a BFT, the elicitation objective factors through the latent posterior, and the gradient of this objective can be estimated from samples of the prior alone. PPT optimizes the parameters of a distribution over hard prompts: it draws prior samples once from the BFT via predictive Monte Carlo (PMC), then estimates the gradient by importance sampling against them. The optimization performs no transformer forward passes and no backpropagation through the transformer, and the prior samples are utility-independent, so a single set of samples drives elicitation against any number of utilities at negligible marginal cost. We validate PPT on Beta--Bernoulli and reinforced urn BFTs across three utility families (reverse cross-entropy, frequency matching, Dyck validity).

Open-Vocabulary Object Goal Navigation (OVON) requires agents to operate under partial observability, making effective internal state updates critical for navigation performance. This update is implemented by the policy network, where recent approaches adopt Transformer-based backbones with self-attention over a context window to integrate temporal information. However, our controlled experiments show that performance does not scale with context length under Transformer-based policies, questioning the suitability of self-attention for state integration in navigation. To this end, we propose Linear Attention-based Navigation (LANav), which adopts linear attention (LA) as the policy backbone to maintain a structured state update rather than self-attention over the context window. Across multiple LA variants evaluated under identical settings, LANav consistently outperforms Transformer-based baselines. Performance improves as state update mechanisms become more structured and regulated, highlighting the importance of state update design. To improve state update effectiveness, we introduce Weighted State-Expansion Linear Attention (WSLA), which expands each attention head's state into multiple sub-states and uses learnable weighted readout to aggregate expanded sub-states. Equipped with WSLA, LANav achieves 36.4% average success rate (SR) on HM3D-OVON, outperforming Transformer-based counterparts by 6.3 percentage points in macro-averaged SR, while maintaining computational efficiency. Distance-stratified results show larger gains in long-distance episodes, while HSSD transfer and fine-tuning demonstrate robustness across scene distributions. Real-world deployment on a Unitree Go2 further achieves an 82% success rate over 50 trials, supporting the practical feasibility and sim-to-real transfer of LANav.

Vision Mamba models replace quadratic self-attention with linear complexity selective state space models (SSMs), emerging as efficient visual backbones. However, MambaOut demonstrates that a Gated CNN block can match or exceed VMamba on image classification, questioning the necessity of SSMs for vision. This raises a fundamental question: do VMamba and MambaOut encode visual information differently at the representation level? To investigate, we apply cross model centered kernel alignment (CKA) analysis and find that VMamba's final stage blocks form representations distinctly different from both MambaOut and its own preceding blocks. We therefore focus on the final block features, decomposing each spatial token into magnitude and direction. MambaOut concentrates class-discriminative information in high-norm foreground tokens that align with Grad-CAM attribution. VMamba, by contrast, produces high-norm tokens predominantly in background regions, misaligned with Grad-CAM, yet preserves discriminative signals primarily in token directions. These observations reveal that the two models rely on different encoding strategies. We connect this difference to high-resolution classification and semantic segmentation. VMamba distributes logit support broadly across object regions, whereas MambaOut relies on sparse dominant tokens, a strategy that becomes less stable as token counts grow. Under full fine-tuning for segmentation, VMamba consistently outperforms MambaOut. These results suggest that VMamba's advantage in dense prediction stems not merely from the SSM mechanism or sequence length, but from how semantic evidence is organized across token magnitude, direction. Ultimately, we conclude that token magnitude and directional structure serve as critical axes for improving visual backbones, particularly under dense supervision.

We study the inference-time behavior of deep linear encoder-only transformers through the lens of interacting particle systems. In this perspective, tokens are modeled as particles that interact dynamically through successive linear self-attention layers. We show that in embedding dimension two, for any key, query, and value matrices, the dynamics can be reformulated as a generalized Kuramoto-type model with pure second-harmonic coupling. This formulation is amenable to Watanabe--Strogatz theory which reveals the dynamics are intrinsically low-dimensional regardless of the parameter matrices. For a class of token initializations associated with the Ott--Antonsen (OA) manifold, we show that the parameter matrices induce a diverse variety of long-time behaviors in linear transformers, including clustering, oscillations, and bifurcations. The oscillations and bifurcations are characterized by uncovering a hidden Hamiltonian structure in the dynamics. By establishing a structural stability result, we further show that dynamics initialized near the OA manifold exhibit the same long-time behavior as those initialized exactly on the manifold. Motivated by our theory in dimension two, we conduct numerical experiments for analogous parameter regimes in higher-dimensional transformers. Our numerical experiments suggest that the long-time behaviors characterized in our theoretical results persist in higher dimensions.

Diffusion Transformers achieve strong image generation performance, but most operate in compressed latent spaces. Pixel-space diffusion avoids this information loss, yet existing approaches map each raw image patch to a single token, forcing one representation to handle both global communication and fine-grained details. We address this issue by proposing a new architecture, \textbf{DuSPiT}, a \textbf{Du}al-branch \textbf{S}ub\textbf{P}atch \textbf{Pi}xel \textbf{T}ransformer. This model separates global structural reasoning from local appearance modeling. DuSPiT uses a compact base branch for efficient global reasoning and a parallel, high-capacity pixel branch, organized into subpatch groups, to preserve detailed appearance, with the two branches interacting through cross-attention. Our results show that DuSPiT generates images with richer details and stronger fine-grained structures, while also achieving a better quality--efficiency trade-off than prior pixel-space diffusion transformers.

Electroencephalographic (EEG) abnormalities arise from dynamic changes in neural synchrony across spatial and temporal scales, yet many computational approaches reduce these dynamics to static features. We present a Spatial Multi-Expert Graph Transformer that models each EEG recording as a sequence of dynamic functional connectivity graphs. Time-resolved connectivity is estimated using the weighted Phase Lag Index (wPLI), and hierarchical graph encoding aggregates information from electrode to regional and global levels. A multi-expert transformer architecture enables subtype-aware reasoning, with a gating mechanism adaptively fusing expert outputs for global abnormality prediction. Experiments on the TUAB dataset show competitive abnormal EEG detection performance and demonstrate the potential of dynamic graph modeling with adaptive expert fusion for interpretable, subtype-aware spatial--temporal analysis.

Feed-forward networks hold two thirds of a transformer's non-embedding parameters, yet the architecture has not received a necessity test that controls parameters, compute, and depth at once. We pretrain attention-only decoder transformers (Simple Attention Networks, SANs) against standard transformers matched separately for parameter count, training FLOPs, and depth (2 to 48 layers), for up to 105B tokens at 6M to 87M parameters. Deleting feed-forward layers in place is costly: the standard transformer leads by 0.47 nats at matched depth and 0.26 nats at matched FLOPs. Reallocating the freed budget into attention depth closes the gap: at matched parameters the difference is 0.006 nats (0.27 percent of loss), reproducible to one part in ten thousand across seed pairs, shrinking across 5B, 30B, and 105B budgets, and holding near 0.02 nats across a 29x size range. Three measurements localize the remaining gap to parametric recall: attention-only models are better on context-grounded answers and worse where knowledge must come from weights. Weight spectra show why: routing matrices (Q/K) crystallize early, content matrices accumulate rank slowly, and removing feed-forward layers relocates this accumulation to the attention output projection. QK-normalization, not feed-forward layers or residual gating, keeps 48-layer attention-only stacks trainable. The deficit concentrates on low-context query prediction and localizes there entirely by the largest budget. A pre-registered test confirms the account: it predicts a 0.02 to 0.05 nat gap on knowledge-dense web text; a matched pair trained on fineweb-edu measures 0.040. Within the tested regime, attention does the rest.

Scaled dot product attention conflates directional alignment and vector magnitude, limiting its effectiveness as a similarity metric in Transformer models. We introduce L1 augmented attention, a simple and computationally parallelizable modification that subtracts a learned, head specific L1 distance between queries and keys from the dot product score. This hybrid similarity captures complementary geometric information. Dot product rewards directional alignment, while L1 penalizes coordinate deviations. To reduce the cost of L1 computation, we project queries and keys into low dimensional subspaces whose parameters specialize to preserve informative L1 structure. Evaluated on WikiText 2 using a compact transformer, L1 augmented attention achieves up to a 14.5% reduction in perplexity over the original transformer baseline and outperforms an RBF L2 kernel. Analysis of norm variance and learned L1 weights reveals distinct geometric roles across layers and strong head level specialization. These results demonstrate that enriching attention with L1 geometry provides a principled and effective improvement to similarity computation in modern language models, with practical benefits for both accuracy and parallel efficiency.

Transformer-based language models organize computation along an ordered depth axis, where shallow and deep blocks often develop distinct representational roles. We challenge the conventional view that these roles must remain tied to a block's position in the ordered sequence. We introduce Mobius Learning, a training architecture based on cyclic depth folding, in which different data streams follow cyclically shifted block orders. The same block group is therefore applied early in the block sequence for some data streams and late for others, so it is optimized in both shallow and deep roles, a phenomenon we call depth-role superposition. Surprisingly, in four-worker experiments with a modded GPT-2 small (124M) model trained on 2.5B FineWeb tokens using Muon, Mobius Learning achieves lower validation loss than a fixed-order looped Transformer at larger numbers of Transformer block-sequence passes. This counterintuitive result shows that a block group need not remain confined to one fixed shallow or deep role within the block sequence and opens a new design space based on cyclic depth folding. Crucially, this structure makes Mobius Learning particularly well suited to memory-constrained distributed training: raw training data remain local, while each worker stores one block group rather than the complete Transformer block stack.

In logistics automation, precise segmentation of unseen objects is crucial for efficient robotic manipulation in cluttered environments. Tasks such as bin-picking and shelf-picking require robust perception to handle occlusions, varying object shapes, and complex spatial arrangements. Traditional RGB-based methods tend to over-segment objects due to their reliance on texture, while depth-based methods often under-segment by focusing primarily on geometric features. To address these limitations, we propose DA-Fusion, a deformable attention-based RGB-D fusion Transformer designed for unseen object instance segmentation. DA-Fusion effectively combines the strengths of both RGB and depth data, enhancing segmentation accuracy in cluttered and multi-layered object environments. We also introduce the Object Clutter Bin Dataset (OCBD), a benchmark dataset specifically tailored for evaluating bin-picking scenarios in top-down views. Extensive evaluations demonstrate that DA-Fusion outperforms state-of-the-art methods across diverse environments, making it particularly suited for real-world logistics tasks.

Large Language Models (LLMs) have had a remarkable impact across many areas of machine learning. However, recent studies have shown that they struggle to reliably solve planning problems. At the same time, theoretical results have shown that transformers, the core architecture underlying modern LLMs, are Turing-complete. In this work, we investigate this apparent gap between the theoretical computational power of LLMs and their empirical planning performance. We propose Chain of Computation (COC), a computational architecture that places a transformer-based LM inside an iterative loop, leveraging its strength as a pattern-matching system. The COC uses a Structured Context Window (SCW) which provides a constant-sized context window with support for choosing which window is used at each planning step. Within this architecture, the LM is able to learn a planning policy, predicts the world model, and performs the arithmetic operations required during planning. We show that, when given an append-only SCW (resembling a Turing Machine tape), even relatively small LMs trained from scratch can learn planning policies and generalize from a small number of training instances within each planning domain, achieving success rates above 99.89% on BlocksWorld and the Pancake puzzle. Our analysis of failure cases in Tower of Hanoi (TOH) reveals that they arise from arithmetic operations or from encountering previously unseen tokens. We show that COC can solve TOH problem instances with up to 20 disks, requiring over 1 million actions, while requiring substantially less training data by either (1) planning with symbolical support for arithmetic or by (2) using a deterministic pushdown automaton (PDA) formulation for the SCW.

We develop an adjoint-sensitivity framework for positional influence in causal residual Transformers and separate unconditional analytic results from conditional boundary-shape conclusions. The principal unconditional theorem is the residual-to-depth-flow estimate for layer controls converging in $L^1$, complemented by a finite-token-to-Volterra attention estimate that explicitly controls the first cells near the causal endpoint. We define a normalized adjoint-energy influence density and derive its exact evolution along full-batch gradient flow. The adjoint admits an exact generator-term decomposition into residual transmission, nonlocal Volterra, and local channels, including all covariance cross terms. Causal masking can amplify early-position sensitivity and residual identity paths can transmit a right-localized terminal bias, but neither mechanism alone forces a U-shaped profile. We therefore state boundary advantages under independently checkable energy, correlation, and local-channel bounds; these conditions are sufficient rather than necessary. Finite-token influence balancing, positional reweighting, and task-aligned observability are presented as diagnostics or regularizers with explicit differentiation requirements, computational costs, and limitations. Controlled simulations illustrate that each intervention controls its designated surrogate, while observability balance or outer-loop reweighting need not monotonically reduce the influence-based Lost-in-the-Middle diagnostic.

We propose a transition-centred geometric analysis of transformer residual streams. Relative displacement measures how \emph{far} representations move between consecutive layers, and orthogonal Procrustes analysis separates each transition into a rigid rotation and a non-rigid residual. Across six instruction-tuned models, on code generation and cross-lingual translation, these measurements reveal reproducible depth regularities. Relative displacement is strongly layer-dependent; typically larger early and late, with a quieter middle third; and nearly invariant across conditions within each model. Rotation magnitude is nearly constant across depth, while Procrustes residual and angle concentration remain depth-modulated, with residual peaking at the final transition. During generation, non-English targets show larger final-layer displacement and residual than English targets. We present these as descriptive geometric regularities, not as measures of computational effort or causal explanations. The contribution is a measurement framework for residual-stream transitions and evidence that, in the settings studied here, depth curves are model-dependent and largely condition-stable.

Modern video transformers typically ignore principles from primate vision and are rarely evaluated against neural data, limiting their biological interpretability. We introduce a sparse winner-takes-all token selection module that replaces dense self-attention to improve efficiency and approximate competitive routing observed in biological visual circuits. We further propose a neuro-inspired split-and-fuse video transformer which uses two complementary pathways: a high-resolution, low-frame-rate "what" stream and a low-resolution, high-frame-rate "where" stream, fused before classification. On Kinetics-400 and Something-Something V2, our best variant operates on the Pareto frontier of accuracy versus inference time among models of comparable scale and pretraining, and showing improved robustness to spatial perturbations. Using representational similarity analysis between model embeddings and time-resolved EEG recordings for the same video stimuli, our model attains a peak brain-model correlation of 0.18 (about 78% of the noise ceiling) and consistently outperforms strong video transformer baselines, suggesting that pathway specialization and sparse competition are useful inductive biases for efficient, brain-aligned video understanding.

Transformers are remarkably versatile and their design is largely consistent across a variety of applications. But are they optimal for any given task or dataset? The answer may be key for pushing AI beyond merely scaling current designs. Method. We present a method to optimize a transformer architecture for a given dataset, which we use as a tool to study optimal task-specific inductive biases. This method replaces the most important non-linearities (GeLUs,;softmax) with functions learned on held-out data. We then train the resulting architectures on other datasets, as a way to evaluate the compatibility between pairs of tasks. Findings. On algorithmic toy tasks, we identify new architectures with dramatic improvements in learning speed, in- and out-of-distribution generalization, and stability across seeds. The new designs prove very task-specific however, and indicate that these tasks require inductive biases very different from those of standard transformers. On code and language modeling datasets, we also find architectures with consistent, yet smaller improvements. These designs transfer much better across datasets and domains (English & computer code). Implications. Our results show that standard transformers are rarely a local optimum in the space of architectures. Simple alternatives can perform much better but sacrifice universality. This suggests that there may be room for improved architectures that better support multiple capabilities simultaneously, such as fluency and robust reasoning.