Overview

Multi-head attention studies how Transformers use parallel query-key-value computations to select and combine information across tokens or features. Researchers probe head specialization with activation analysis, ablation, and circuit discovery, while theoretical work examines expressivity, learning dynamics, and the number of heads needed for particular computations. Architectural approaches include sparse and grouped attention, cross-head communication, and compressed projections, evaluated in settings such as long-context retrieval, language modeling, vision, and reasoning benchmarks. A central tension is that specialized heads can support useful capabilities, but their roles vary by task and model, and reducing attention cost does not always preserve accuracy.

Latest papers

77 papers6 papers in the last four weeks, down 25% on the four weeks before. 0.0% of all new papers.

Nov 2, 2026

Exact Fusion of Softmax Attention Heads Can Require Exponentially Many Atoms

Fusing independently normalized attention heads can increase the number of explicit key-value entries, or atoms, needed to preserve their output. We prove a sharp separation for Wd(q)=∑j=1dσ(qj)W_d(q)=\sum_{j=1}^dσ(q_j): its dd two-atom heads use 2d2d head-atom terms, whereas its minimum single-head representation has exactly 2d2^d atoms, even with arbitrary fixed replacement keys, scalar values, and positive masses. The same lower bound holds for every fixed uniform error below 1/21/2 on Rd\mathbb R^d, and the threshold is sharp. The proof uses exposed-key normal cones; exact equality on any nonempty open query domain inherits the bound. The resulting atom bound for a fixed head count is tight when that count divides dd. A normalized family retains its exponential exact minimum while approaching a constant on fixed-radius query balls. For pretrained-model caches, a classical affine-rank bound and a certified instance exclude exact vector preservation at a specified atom count. Learned-cache comparisons on two disjoint banks of 64 contexts show budget-dependent reversals and calibration effects.
Oct 30, 2026

Hard Sample-Conditioned Decorrelated Attention for Robust Face Recognition

Multi-head attention in Vision Transformers (ViTs) allows independent attention patterns across heads, but without explicit regularization, heads may learn redundant patterns rather than complementary roles. Correlation-based penalties encourage head disentanglement, yet apply uniform regularization across all training samples. In face recognition (FR), hard samples whose embeddings lie close to competing identities are critical for optimization and generalization. Motivated by this, we propose ViT-HarDis, which conditions attention head disentanglement on sample difficulty: samples with high cosine similarity to negative class centers receive stronger head decorrelation, enforcing specialization where identity ambiguity is highest. Experiments show improvements over both the baseline and uniform disentanglement on most benchmarks. Our analysis shows that ViT-HarDis reduces the similarity of hard samples to competing class centers while decreasing inter-head attention similarity, and that applying disentanglement only to the final transformer layers outperforms regularizing the entire network. Trained on WebFace4M, ViT-HarDis achieves SOTA results among ViT-based methods on IJB-B, IJB-C, and TinyFace Rank-1, while achieving the best IJB-B and TinyFace results when trained on MS1MV3, improving FR without architectural modifications or additional inference cost. Code: https://github.com/TaharChettaoui/Hard-Sample-Conditioned-Attention-for-FR.
Oct 23, 2026

Head Dimension Matters in Language Models for Long-Context Retrieval

Multi-head attention retrieves and combines context representations through query--key interactions and value aggregation, with head dimension constraining the per-head representation space. In long-context processing, it remains unclear whether allocating a fixed parameter budget to fewer, wider heads improves information retrieval. In this paper, we investigate how head dimension affects the performance of long-context retrieval tasks. Our empirical studies reveal that pretraining loss and short-context non-retrieval task performance are not affected much by head dimension, while long-context retrieval performance improves with wider heads at fixed model sizes. We find that models with wider heads have a higher proportion of retrieval heads, associating more parameters with retrieval. Our ablation studies also show that the retrieval performance improves with wider heads even when the projection weights in attention modules are frozen during long-context continual pretraining, suggesting that wider heads not only benefit updates in attention modules but lead to more effective adaptation of components such as embeddings and MLPs.
Oct 18, 2026

Formally Verified Autoresearch for Analyzing the Expressivity of Multi-Head Attention

Multi-head attention is a central component of the transformer architecture, yet how its expressivity depends on the number of heads remains poorly understood. We study a one-layer, attention-only transformer with additive positional embeddings and define the head complexity H∗(f)H^\ast(f) of a Boolean function ff as the minimum number of attention heads required to represent it. We derive an exact linear-fractional normal form for head complexity, prove a general threshold-degree lower bound, and give an explicit ten-bit function for which this bound is strict. For symmetric Boolean functions, we independently show that head complexity equals the number of times the function changes value between consecutive Hamming-weight levels, even when positional embeddings are allowed. This agrees with the concurrent characterization of Hsu and Xu (arXiv:2608.04243) in a model without positional embeddings. The symmetric lower bound combines our general threshold-degree bound with the classical result of Aspnes et al. (1994). These results were discovered and formalized by a frontier model through conjecture generation, theorem proving, and machine verification, while human researchers asked the high-level questions and reviewed the generated theorems. We make the Lean formalization and research traces publicly available to support agent-assisted theorem proving and formal verification in mechanistic interpretability. This work provides a proof of concept for using the capabilities of frontier models, grounded by formal verification, to develop an understanding of transformer architectures.
Oct 15, 2026

Can Capability Enhancement be Localized?

Transformer-based language models have shown strong capabilities across many tasks and domains. However, prior work shows that, for a given task, not all attention heads matter equally. Specifically, they suggest that many redundant heads can be safely removed with little-to-no effect, while a small set of important heads, when ablated, severely damages the model's performance on the given task. In this work, we investigate a complementary aspect, called Capability Enhancement Localization (CEL): given a model and a task, does there exist a small set of heads whose removal improves the task performance? We first investigate this question empirically across models from different families. Using a simple two-stage routine that screens candidate heads and evaluates their joint ablations, we identify sparse head combinations whose removal improves task-specific performance without further fine-tuning. Motivated by these observations, we ask theoretically whether such a phenomenon can naturally arise from training dynamics, and under what conditions. To this end, we construct a simplified multi-head softmax attention model and two tasks with conflicting features, and show that gradient descent from small random initialization provably concentrates the conflicting features into particular heads. We further show that ablating such heads improves performance on one task while harming the other, even though training converges to a global optimum of the joint objective. This establishes a possible training-induced theoretical explanation for CEL.
Oct 14, 2026

Choosing the Number of Attention Heads from Intrinsic Dimension: A Small-Scale Study

The number of attention heads in a transformer is usually set by convention or grid search. We test a simple heuristic for setting it in advance: estimate the intrinsic dimension of token representations without labels, and use roughly that many heads. We estimate intrinsic dimension with a nested-dropout VAE trained on DistilBERT embeddings of WikiText-2, choosing the KL weight at the knee of the VAE's rate-distortion curve rather than by hand. This gives k∗=9.2k^*=9.2 (95% CI [8.6, 9.8]). In a small 4-layer transformer trained from scratch, 10 heads and 32 heads are statistically equivalent on SST-2 and on MRPC (paired TOST over 10 seeds, ±1\pm1 point margin), with 69% fewer attention parameters for the 10-head model. Training 10 heads from the start also slightly outperforms training 32 heads and pruning to 10 (0.6 points, p=0.008p=0.008). The evidence is narrow: two classification tasks, one small architecture, no pretrained models, and no formal link between intrinsic dimension and head count. We present the result as a practical heuristic worth testing at larger scale, not as a general principle.
Oct 9, 2026

Asymptotic Analysis of Discrete-Time Self-Attention Dynamics in Transformers

In machine learning, a self-attention dynamics describes the attention mechanism of transformers, and is obtained treating the layer index of a transformer as time index. Even though naturally discrete-time, this dynamical system is most often studied in continuous-time. Scope of this paper is to provide a thorough investigation of the discrete-time version of the self-attention dynamics: a complete classification of its equilibria is given, and their stability properties are studied, in both the single-head and the multihead case. We show that in both cases the main attractor of the dynamics is a consensus equilibrium aligned with the principal eigenvector of the value matrix (in the single-head case, a corresponding matrix of weights in the multihead case), even though other locally asymptotically stable bipartite equilibria can be show to exist at least in the single-head situation. The behavior is confirmed by looking at the learned weight matrices of a real transformer.
Oct 9, 2026

Data-Adaptive Mahalanobis Metric Learning for Cross-Head Attention in Transformers

In this paper, we propose Data-Adaptive Mahalanobis Cross-Head Attention (DAMCHA), a simple and effective attention mechanism that generalizes Multi-Head Attention (MHA) through the lens of metric learning. Standard MHA implicitly learns a static, head-wise Euclidean similarity corresponding to the diagonal blocks of a unified cross-head metric matrix, discarding all off-diagonal interactions across heads. DAMCHA addresses this limitation via two key innovations. First, it learns the full cross-head metric matrix, inducing a Mahalanobis-like similarity that explicitly captures inter-head interactions beyond the restricted block-diagonal structure of MHA. Second, it parameterizes the metric matrix as an input-dependent function, enabling the attention geometry to dynamically adapt to the intrinsic manifold structure of the data rather than remaining fixed after training. We further introduce principled regularization strategies, most notably stack-wise parameter sharing, to ensure computational efficiency and stable optimization. DAMCHA serves as a drop-in replacement for MHA and its variants in various Transformer architectures, delivering superior expressive power and learning performance with comparable or reduced computational overhead. Extensive experiments demonstrate that DAMCHA-based Transformers consistently outperform competing baselines across a range of benchmark tasks. Code is available at https://github.com/erwinmsmith/DAMCHA.
Oct 7, 2026

Gaussian Equivalence for Multi-Head Self-Attention

A theoretical understanding of multi-head self-attention is fundamental to the study of modern neural networks. Using random matrix theory, we establish Gaussian equivalence for multi-head self-attention: replacing softmax attention with rescaled scores plus Gaussian noise preserves the limiting spectral law of the centered output. This equivalence also covers value and output projections that depend on the keys. The resulting laws separate the effects of head allocation and projection widths, and distinguish spectrum-preserving across-head sharing from within-head key--value dependence.
Oct 1, 2026

Let the Heads Talk: Beyond Diagonal Graph Attention

Sheaf Neural Networks generalize scalar-weighted message passing by replacing scalar edge weights with linear transport maps between local feature spaces. Yet the role of this matrix-valued transport is entangled with the broader sheaf-diffusion construction. We isolate the transport primitive through quiver representations and establish a direct connection with multi-head attention. Treating attention heads as coordinates of a local transport space reveals that standard multi-head attention implements diagonal edge maps: along each directed interaction, a source head can contribute only to the corresponding receiver head. Allowing off-diagonal entries instead enables edge-conditioned communication across heads before neighborhood aggregation. We show that this operation cannot, in general, be absorbed into a single shared linear map applied after aggregation. Building on this characterization, we introduce Topological Attention (Top-A), a multi-head attention that learns edge-dependent off-diagonal routes while preserving the original same-head paths and exactly recovering vanilla attention when the additional routing vanishes. We evaluate Top-A on relational reasoning, heterogeneous graph learning, and algorithmic reasoning, including out-of-distribution generalization, with heterophilic node classification as a contrast setting. The results show that cross-head transport is most useful when the task benefits from interaction-dependent transformations, while heterophily alone provides no systematic advantage. These findings identify edge-conditioned cross-head communication as a distinct computational primitive of matrix-valued transport.
Sep 29, 2026

Learned Queries and Keys Are All You Need: Replacing the Value Projection with Structured Transforms

To reduce the number of parameters and cache memory requirements of transformers we introduce dual-headed transformers instead of three heads. We studied Walsh-Hadamard Transform (WHT), Discrete Cosine Transform (DCT), Discrete Fourier Transform, filterbank based Shearlet Transform, and Multiplication-Avoiding (MA) operators to construct dual heads. We combine spatial patches and their orthogonal transforms (or Shearlet and MA operators) in a structure similar to the attention block. We obtained better results than triple headed transformers in ImageNet. Extensive simulation examples are presented.
Sep 27, 2026

MorphAtt: A Neuromorphic Accelerator for Efficient Multi-Head Attention Processing in Spiking Vision Transformers

Spiking Vision Transformers (SViTs) are developed as an energy-efficient alternative to conventional ViTs for computer vision tasks at the edge. However, huge parameter counts and complex multi-head self-attention (MHSA) operations make it challenging to achieve high energy efficiency in SViT inference, especially in tightly constrained applications. To maximize efficiency gains of SViT processing, we propose MorphAtt, a novel digital accelerator that expedites SViT inference through streamlined processing. Specifically, it processes MHSA operations using cascaded hardware modules: a Spiking Query-Key-Value generator (SpikeQKV), a low-complexity Spiking Multi-Head Self-Attention engine (SpikeAtten), and Reparameterization Convolution (RepConv) modules. To mitigate traffic congestion in on-chip memory accesses and data reuse, specialized inter-module buffers are integrated within the dataflow. Under synthesis using 32nm CMOS technology, MorphAtt achieves 792-1605 GOPS of throughput, while incurring ~39-55 mW of power consumption and 1.5 mm^2 of area, which lead to 20.3-29.1 TOPS/W of energy efficiency. These results also demonstrate that our MorphAtt offers better performance and efficiency trade-offs than state-of-the-art, thereby enabling highly energy-efficient vision-based AI systems at the edge.
Sep 26, 2026

CoWindow Attention: Full Causal Coverage Is a Collective Property

FullAttn repeatedly exposes the complete causal history to every attention head, creating substantial redundant computation and memory traffic even with IO-efficient dense kernels. We introduce CoWA, a structured attention architecture that distributes access to the causal history across KV heads. All heads share near-diagonal and prefix-sink windows, while complementary long-range windows partition the remaining history. Their union provides full causal coverage although each head attends sparsely to distant tokens. This position-defined attention pattern requires no learned router or indexer, is used consistently during training and inference, and aligns with KV-head tensor parallelism. A window-matched ablation at 8K isolates the effect of complementary long-range allocation: CoWA with 100% collective coverage reaches 89.73% accuracy, compared with 89.97% for FullAttn, while duplicated long-range windows perform substantially worse. Across a broader controlled associative-recall comparison with matched token budgets, CoWA closely tracks FullAttn as the context grows, whereas other sparse patterns lose a substantial fraction of the associations. In an attention-operator benchmark at 128K tokens with tensor parallelism, CoWA reduces forward and backward latency during training by 7.4x and 8.6x and decoding latency during inference by 3.0x over FullAttn. Its per-rank peak operator memory matches FullAttn during training and is 7.6x lower during decoding. Across scaling-law training from 0.6B to 14B parameters, CoWA closely tracks FullAttn in perplexity while reducing total training FLOPs. The resulting 14B models and 32B models from separate continued training achieve comparable knowledge, reasoning, and long-context retrieval scores to FullAttn. These results show that full causal coverage can be a collective property of the head ensemble rather than a duplicated property of every head.
Sep 22, 2026

Transformer Heads Looking for Order

In this note, we show that the problem of checking, whether a sequence of bits is ordered, is not doable by 1-head 1-layer transformers but is doable by a 2-head 1-layer transformer. Unlike similar previous results, our results assume the model where transformers have an output MLP.
Sep 3, 2026

The Head Complexity of Boolean Functions in Single-Layer Attention

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: kk heads compute kk-bit parity but cannot compute (k+1)(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+1k+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: 2n2^n heads suffice to compute every nn-bit binary function, with one head per monomial in its multilinear expansion, while a counting argument shows almost all such functions require Ω(2n/n2)Ω(2^n/n^2) heads. This lower bound matches the upper bound to within a poly⁡(n)\operatorname{poly}(n) factor, even when dimension and precision are unbounded. Together, these results characterize head requirements for Boolean computation in this model.
Sep 1, 2026

Multi-Head Self Attention is a Parameter Identification Mechanism

We prove that a multi-head scaled dot product attention can be viewed as a parameter identification strategy. The ratio of unidentified parameters to the total number of parameters scales like the reciprocal of the number of heads (1/2→1/(2H)1/2 \to 1/(2H)), meaning models with more heads are structurally more identified. A subtle side effect of the mathematics observation that attention can never be fully identified. Similarly we also show that some bias terms can have no effect on softmax-based attention layers in both the single- and multiple-head settings, though this is mostly a curiosity that should have a marginal effect on model size and model training/prediction efficiency. We also touch on modern improvements to transformers including RoPE and GQA from this perspective, illustrating how those as well can improve the ratio of meaningful'' parameters to all parameters. Simple numerical examples demonstrate that training can indeed involve updates that overlap model-invariant subspaces that arise from a lack of identification. As part of our experiments we use a rebalancing'' approach that can ``fix'' updates that overlap unindentified subspaces but do not try to present evidence this should actually be adopted. Instead we simply view our numerical results as exploring and confirming the theoretical results. As a whole we discuss a purely mathematical/statistical explanation, identification, for why specific architectural choices in transformers may have improved performance.
Sep 1, 2026

Scaled Idempotence in Transformer Attention: Paired OV Geometry and Shared-Value Algebras

We identify a recurrent algebraic regularity in Transformer attention: a sparse subset of effective OV operators T=OV⊤T=OV^\top nearly closes under composition, T2≈αTT^2\approxαT. Across six pretrained endpoints spanning 2.8B--235B parameters, 3.98--8.00% of heads reach squared closure alignment P≥0.9\mathcal{P}\geq0.9, while no matched within-layer O/V mismatch does. An exact principal-coordinate factorization, T=QOKQV⊤T=Q_OKQ_V^\top and T2=QO(KDK)QV⊤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 KK while preserving singular values, norms, factor spans, and principal angles reduces median closure from 0.336 to 1.04×10−41.04\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, TiTj=αjTiT_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.
Aug 31, 2026

Inherent Trade-offs when Enforcing Correct Data Symmetries in Multi-head Self-attention

We give a complete characterization of equivariant multi-head self-attention (MHSA): if an MHSA layer is equivariant to a symmetry group GG, then GG 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=D4G=D_4 acting on CC copies of the regular representation as the token feature space, we show that there are Ω(C64)Ω(C^{64}) components for eight attention heads. We illustrate this with experiments on artificial tasks and show that the predicted head-cluster structures can arise naturally when the attention layer is pressured to become equivariant.
Aug 25, 2026

Mahalanobis-Based Multi-Head Attention for Complex State Propagation

In this paper, we propose Mahalanobis-Based Multi-Head Attention (MHA-CSP), a novel attention mechanism that replaces the standard dot-product with a Mahalanobis distance-based RBF kernel, which effectively computes attention in an infinite-dimensional feature space without increasing the parameter count. Crucially, the positive definiteness of the Mahalanobis distance enables a direct construction of Tree Attention: attention scores are built directly from accumulated distances, with a LogSumExp correction that rectifies the raw distance by subtracting the log-sum of edge exponentials. Moreover, the multi-head Mahalanobis distance matrices are themselves repurposed to construct an attention meshing mechanism, enabling cross-head kernel collaboration that simultaneously boosts accuracy and training efficiency. Extensive experiments demonstrate that MHA-CSP, with only 119K parameters and teacher forcing applied exclusively at the final hidden state, consistently outperforms Transformer and GCN baselines trained from scratch under identical conditions on long-sequence state tracking tasks. While these baselines rely on dense attention or graph propagation, MHA-CSP achieves robust structured reasoning via synthetic distance rectification---powered by Mahalanobis-based attention---and efficient information bypass inherited from the CSP backbone. This result highlights the effectiveness of complex-valued state propagation with collaborative multi-head rectification in capturing symbolic structures, establishing a new efficiency-performance trade-off for structured reasoning.
Aug 5, 2026

Matrix Zonotopic Attention: A Context-Adaptive Value Projection for Set Transformers

Multi-head attention combines an input-dependent softmax routing with an input-independent linear value projection, so the per-sample operator mapping aggregated values to outputs is the same for every input set. We study the consequences of this asymmetry for permutation-invariant set targets. We introduce the Transformation Degrees of Freedom (TDOF) of a target operator, a complexity measure counting the input-dependent directions an exact representation requires, and present a depth-separation analysis showing that context-rigid attention needs depth proportional to the target's TDOF, whereas a single layer with a context-adaptive value family can represent the same target. Building on this analysis, we propose Matrix Zonotopic Attention (MZAttn), which replaces the fixed value projection with a context-adaptive matrix-zonotope family: a centre matrix plus a sum of generator matrices weighted by input-dependent gates. The construction reduces to standard multi-head attention at initialisation, preserves permutation equivariance, and admits a data-driven reachability interpretation. Experiments on a range of set-prediction tasks are consistent with the TDOF prediction that the architectural advantage is selective: it appears on targets that depend on the input set in a high-rank, sparsely combinatorial way, and is small on aggregate-statistic targets where parameter-matched standard attention is already competitive.
Aug 4, 2026

Attention-based representations for multi-task computation

Multi-head attention layers produce vector representations that support multiple downstream tasks. We establish bounds on the number of heads required in two simple and concrete multi-task scenarios. In the first scenario, a vector representation is sought so that linear predictors can compute both the smallest and largest numbers in a given list. In this case, it is known two attention heads with small embedding dimension and bit precision level suffice. We prove that a single attention head requires exponentially higher embedding dimension or precision level. In the second scenario, a vector representation is sought so that a polynomial threshold function can compute the XOR of a given string of nn bits. This scenario is analogous to the first one for n=2n=2, since XOR is readily computed by a linear function using a vector representation that encodes both the AND and the OR of the two bits. We observe that nn-bit XOR requires the product of the number of heads and the polynomial degree to be at least nn, and we construct multi-head attention layers that match this lower bound. These results generalize to arbitrary (symmetric) Boolean functions, where the bound is given in terms of the threshold degree.
Aug 4, 2026

When Attention Goes Blind: Numerical Failure in ALiBi Positional Encodings

We identify a previously overlooked failure mode of ALiBi positional encoding: its linear bias scaling underflows floating-point precision, which zeroes out a large fraction of attention weights and renders the affected attention heads partially blind. We analyze this failure mode, characterize its impact, and examine four mitigation strategies. We further demonstrate its occurrence in state-of-the-art pretrained models based on ALiBi. Comprehensive pretraining experiments with 148M-parameter decoder models help us to disentangle its effects from out-of-context degradation. We find that ALiBi's failure mode can substantially impair token retrieval while having only a minor effect on standard decoder benchmarks. We propose four training-time mitigation strategies and evaluate them individually and in combinations, finding that log-scaled distances yield the most consistent improvements in passkey retrieval. Despite this problem, default ALiBi slopes remain a surprisingly strong baseline, particularly for needle-in-a-haystack retrieval. Based on these findings we provide concrete recommendations on how to train models with ALiBi.
Aug 4, 2026

Provably Learning Multi-Head Attention with Queries

We study the problem of learning multi-head softmax attention from black-box input-output access. The learner may query arbitrary real-valued token sequences and observe only the scalar output at the final token. Recent work gives an algorithm using O(d2)O(d^2) value queries to recover the single-head parameters (W,v)(W,v). For multiple heads, the same work establishes identifiability under the assumption that the heads occupy pairwise orthogonal subspaces. Applying the single-head recovery algorithm separately to the heads additionally requires bases for these subspaces to be known. We recover a canonical representation by merging heads with the same WhW_h, summing their corresponding vhv_h, and discarding a merged head when this sum is zero, without these subspace assumptions. By varying the number of copies of a token, our algorithm obtains samples of a rational function whose interpolation separates the canonical heads. Additional queries formed by adding selected token vectors then match the same head across different queries. When the oracle outputs and all subsequent computations are exact, the learner chooses its query vectors at random and recovers the canonical pairs {(Wh,vh):h∈[H]}\{(W_h,v_h):h\in[H]\} up to permutation with probability one. When HH is known, it uses exactly 4Hd2−2H+14Hd^2-2H+1 value queries of maximum length 2H+12H+1. If only a known upper bound H0H_0 is available, the algorithm uses 4H0d2−2H0+14H_0d^2-2H_0+1 value queries of maximum length 2H0+12H_0+1. For approximate oracle outputs, we give conditions under which the parameter error is at most a model- and query-dependent constant multiple of the output error. Finally, we extend our result to a one-layer Transformer with multi-head attention followed by a bias-free ReLU feed-forward network. Under additional conditions, we recover a functionally equivalent Transformer without relying on a separate algorithm for learning the feed-forward network.
Jul 25, 2026

Through the Bottleneck: How Multi-head Latent Attention Separates Content from Position in Language Models

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.
Jul 22, 2026

Multi-Head Attention Residuals

Transformers propagate information across depth through a single additive residual stream: every sublayer reads only the most recent state. Attention residuals relax this by letting each sublayer attend, through a learned softmax. However, that read uses a single query shared across the entire width, so every feature subspace must read the depth history through one distribution. The cost of this forced compromise grows with how much the subspaces disagree about which layers to read, and disagreement grows with model width. We introduce Multi-Head Attention Residuals (MHAR): the routing query is reshaped into H per-subspace heads, each with its own softmax over the depth history. The read becomes block-diagonal, the reshape adds zero parameters and negligible compute, and H = 1 recovers attention residuals exactly. Trained from scratch on a deduplicated Nemotron-based anneal corpus that is quality-filtered and STEM- and code-heavy, MHAR improves validation loss over a standard Transformer at 100M, 350M, and 1B (-0.061, -0.149, and -0.140). It achieves the best result among four methods in every setting, with the gain increasing from 100M to the larger scales. The head count is a real design axis rather than a free knob: validation loss is U-shaped with respect to H, with a flat optimum at H = 4 or H = 8 across scales. We adopt H = 8 for large-scale models; over-splitting beyond this point (H = 16) consistently gives back part of the gain. A direct probe of the trained queries confirms that learned subspace disagreement is the underlying driver. Fused Triton routing kernels increase attention-residual training throughput from 0.2-0.5x to 0.55-0.88x of the baseline while maintaining near-baseline peak memory. An identity-preserving conversion using delta attention residuals supports 8B mid-training, yielding improvements of +3.2 on GSM8K and +3.1 on GPQA.
Jul 12, 2026

From Self-Attention to Connection Laplacian: A Unified Operator View of Transformers

Self-attention is a ubiquitous primitive in modern sequence models, yet its operator-level geometry is only partially understood. We view a token sequence as a vector field over the token-position graph and identify attention as a connection walk: messages are aggregated by a nonnegative walk matrix while being transported along each edge by a learned linear map. Within this framework, we prove that single-head attention (SHA) is exactly a connection propagation step with constant transport, and that multi-head attention (MHA) is exactly a single edge-dependent connection walk whose effective transport is an attention-gated mixture of headwise transports. We further clarify the conditions under which the corresponding generator reduces to a random-walk connection Laplacian, highlighting the roles of stochasticity, reversibility, and metric-compatible transports. Empirically, we find that trained Transformers across scales (from 124M to 8B) and structures (encoder/decoder) exhibit geometric structure consistent with our theory: effective attention graphs converge to stable geometric operators in deeper layers, learned transports self-organize into approximate scaled isometries, and both phenomena strengthen consistently with scale. Overall, the paper provides a precise connection-walk formalism that links self-attention to classical geometric operators, along with a set of operator-level tools for analyzing transformer models from a geometric perspective.
Jul 7, 2026

Fingerprint, Not Blueprint: How Positional Schemes Set the Default Spectral Algebra of Attention

The pre-softmax score of an attention head is a bilinear form score(i,j)=xiTMxjscore(i,j) = x_i^T M x_j in a learned operator M=WqTWkM = W_q^T W_k. Because M is generally non-symmetric, hence non-normal, it has a complex eigenspectrum and non-orthogonal eigenvectors, the regime where non-Hermitian and random-matrix tools apply. We ask what this spectrum encodes, at three levels for previous-token and induction circuits. Statically, across seven pretrained models spanning three positional schemes, the strongest previous-token heads are spectrally rotational under RoPE and non-rotational, or content-like, where position enters outside QK (learned-absolute and ALiBi); the model-level separation is perfect at every top-k examined (exact permutation p=0.029p=0.029), and zeroing the per-frequency RoPE phase Im(Mt)Im(M_t) eliminates induction on a pre-identified previous-token head in all three RoPE models. Dynamically, over public Pythia checkpoints every head originates at the random-matrix (Ginibre) null; the rotational signature emerges with the behavior, not before it, and the population-median suppression that yields the final profile follows circuit formation, so the profile is a consolidated fingerprint, not a precursor. Causally, and at toy scale, no spectral channel is necessary: constrained two-layer training reroutes around every ban with capability intact, albeit at a significant formation delay (four pre-registered contrasts, qBH<=0.016q_BH <= 0.016). The cost structure exposes each scheme's default: imposing symmetry slows learned-absolute models by a factor of 2.9, whereas a RoPE head with a fully symmetric static M still routes directionally via the phase channel, impossible under absolute positions. Within the settings examined, the positional scheme sets the default spectral algebra of an attention head's solution: a fingerprint sculpted after function, not a hard constraint upon it.
Jul 1, 2026

Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads

In long-context use, large language models frequently synthesize answers from the meaning of a relevant context span rather than literally copy-pasting them. Identifying which attention heads perform this synthesis matters for interpreting long-context model behavior. Yet existing detectors miss these heads by construction: they reward heads whose attended token matches the generated token, a literal-copy criterion that captures where a head reads but not what it writes through its output-value (OV) circuit, the very mechanism that carries non-literal retrieval. We introduce Logit-Contribution Scoring (LOCOS), a write-aware detector that scores each head by the projection of its OV-circuit output onto the answer-token unembedding direction, contrasting needle and off-needle source positions in a single forward pass. Across three model families (Qwen3, Gemma-3, OLMo-3.1), mean-ablating the top LOCOS heads on the NoLiMa non-literal retrieval benchmark collapses ROUGE-L at lower head counts than prior attention-based detections; on Qwen3-8B, ablating 50 heads drives ROUGE-L from 0.401 to 0.000 while the strongest baseline still retains 0.292. The selected heads are retrieval-specific: parametric recall and arithmetic reasoning stay at baseline under the same ablation. On Qwen3-8B, the same ablation also drops MuSiQue from 0.55 to 0.08 and BABI-Long from 0.62 to 0.20, while a random-heads control stays within 0.05 of baseline.
Jun 25, 2026

Prism Transformer: Progressive Head Schedules for Hierarchical Attention Processing

Multi-head attention conventionally partitions the hidden dimension equally across all heads at every layer, enforcing an identical representational subspace dimension (dh = dmodel/h) throughout the models depth. In this work, we identify this uniform allocation as a fundamental structural bottleneck: due to their restricted dimensional space, early-layer heads are unable to faithfully capture complex, high-dimensional contextual patterns. To resolve this, we introduce the Prism Transformer, a novel architectural paradigm that replaces the static, uniform head configuration with a progressive head schedule. By monotonically increasing the head count across layers, the Prism Transformer naturally establishes a local-to-global representational hierarchy: early layers leverage fewer, exceptionally wide heads to capture complex, local compositional patterns, while deep layers deploy many, narrow heads to decompose these patterns into specialized linguistic features. Crucially, this structural shift is parameter-neutral, compute-neutral, and introduces zero training or inference overhead, preserving identical weight matrices and FLOP budgets as the standard Transformer. Across three model scales (124M, 354M, and 757M), the Prism Transformer consistently outperforms uniform baselines, achieving consistent reductions in validation loss alongside consistent gains on downstream zero-shot benchmarks (including PIQA, HellaSwag, ARC-Easy, and WinoGrande). Our findings demonstrate that non-uniform subspace allocation unlocks latent capacity within the standard Transformer budget, enabling more effective use of model capacity.
Jun 25, 2026

DnA: Denoising Attention for Visual Tasks

The softmax activation in multihead attention (MHA) is the de facto standard for attention-based models in visual perception tasks. However, standard softmax can produce noisy attention patterns that dilute relevant features and degrade its performance. In this paper, we propose Denoising Attention or DnA, in which, first, a positive query identifies which image features belong to the correct class, and a negative query identifies closely associated but irrelevant image features. DnA then projects these interactions into two distinct subspaces with larger principal angles, promoting subspace separation and improved discriminability. Using a ViT-B backbone, our proposed DnA achieves an absolute gain of 0.8% on ImageNet-1K compared to the baseline. We further show improvements across multiple visual understanding tasks, including video understanding with video transformers (1.8%) and video LLMs (0.5%). Our extensive empirical analyses justify the design choices involving two interacting subspaces and the denoising effect of DnA.