Learned KV-cache eviction often faces a soft-to-hard mismatch: during training, differentiable gates typically attenuate token contributions, whereas inference saves memory only when KV entries are physically removed. We ask whether the attention substrate affects this soft-to-hard transition. Using GPT-2-scale Transformers trained on OpenWebText, we run a controlled $2\times2\times2$ comparison over attention type, learned gating, and positional encoding. Although sigmoid attention is worse as a dense language model, learned hard eviction changes the useful operating points: sigmoid-gated models delete KV entries with negligible PPL change relative to their own no-eviction references. Under a matched live-cache protocol on the same dense backbones, learned sigmoid gates obtain lower PPL than our H$_2$O and KeyDiff implementations, whereas softmax gates do not uniformly beat these post-hoc methods. The results suggest that attention normalization can substantially affect whether a training-time soft gate transfers cleanly to hard KV deletion.
Attention masks are relation-level controls: they specify which query--source pairs may interact. They do not provide a representation-carried token state that is non-participating at the attention boundary. We assign each token hidden carrier \(h_i\) an active-presence coefficient \(p_i=\lVert h_i\rVert^2/(τ+\lVert h_i\rVert^2)\). The same coefficient has two roles: it gates information emitted by token \(i\), and it determines the mass with which token \(i\) enters computations shared with other tokens. OAttention is the support-coupled attention realization of this rule. It gates the receiver output by \(p_i\) and weights source \(j\) by \(p_j\) in both the attention numerator and partition, while retaining the standard score, visibility relation, exponential competition, and value aggregation. This makes the zero-vector token a zero element and yields exact null-receiver, null-source insertion, self-attention insertion, and empty-support properties. The same token-level presence gives local O-components (OFFN, ONorm, and OInject), presence-weighted OStandardize, the O-Closure law \(M(H\oplus0)=M(H)\oplus0\), and an OTransformer by residual and compositional closure. The canonical operator is checked by contract tests and a GPU evaluation. In a zero-fine-tuning retrofit of a cloned pretrained TabPFN v3 regressor, calibrated hidden-carrier OAttention and Full-O variants change mean RMSE by $+0.088\%$ and $+0.177\%$, respectively, over 18 matched dataset--seed cases. A two-block ablation shows that OAttention alone does not preserve a NULL state through ordinary host components, whereas the OTransformer path does. These are scoped tests of exactness, active-path compatibility, and compositional necessity; they do not establish universal no-loss, arbitrary-host closure, learned attraction to the origin, or a general semantics for missing values.
Attention mechanisms have driven machine learning for a decade, from neural machine translation to language models that do general-purpose reasoning. This survey covers four connected threads: their formulation for sequence-to-sequence tasks, adaptation to computer vision, efficiency innovations that address the quadratic bottleneck, and advances in interpretability. We define three criteria: efficiency, expressiveness, and interpretability, and compare twenty-one methods using an EEI scoring framework. Scores come from a single rater with an assumed +/-1-point perturbation range. A deterministic Monte Carlo analysis with 200,000 samples shows that, under this perturbation model, rank changes of more than one position occur in 67-70% of samples on average. A rank-matched null model reproduces a similar stability profile, so the results support coarse tier-level comparisons rather than fine-grained rankings. The survey traces attention from Bahdanau-Luong alignment through the Transformer and into vision architectures. It reviews fixed and learned sparse attention, linear attention, IO-aware exact algorithms including FlashAttention, and state-space alternatives including Mamba. It also covers induction heads, superposition, and the attention-SSM duality. We further provide a structured narrative review, a benchmark synthesis with cross-study caveats, a five-problem research gap analysis, and a 2015-2026 evolution timeline. We conclude by framing attention research as an expansion of the efficiency-expressiveness-interpretability frontier and identifying future directions including unified efficiency benchmarks, learned routing for hybrid architectures, length generalization, and scalable mechanistic interpretability.
A standard self-attention layer consists of two interacting circuits: the query-key circuit that governs attention allocation, and the output-value circuit that maps attended representations to predictions. Collapsed and factorized parameterizations of the query-key and output-value circuits lead to qualitatively different attention patterns. In particular, some parameterizations give sharper attention to task-relevant tokens, at a similar training loss. We analyze how the parameterizations of these circuits shape the parameter trajectories in single-layer self-attention models trained for next-token prediction. Through gradient-flow analysis, we show that factorization induces implicit rescaling of the two circuits' learning rates. We derive closed-form dynamics showing that output-value and query-key parameters move along a line, with relative speeds determined by their learning rates. Faster query-key learning relative to output-value learning thus produces sharper attention, as the model compensates for slower output-value learning by increasing attention mass on relevant tokens. Experiments show that differences in the relative learning rates of the two circuits govern attention concentration. This improves attention interpretability proxies while maintaining comparable predictive performance.
Christopher Schröder, Lukas Gienapp, Ferdinand Schlatt +2cs.CL
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.
All Transformer-based large language models compute attention via the Euclidean inner product, an architectural choice that Dong et al. (2021) proved causes representational rank to decay doubly exponentially with depth in pure self-attention stacks. We develop a theoretical framework that targets this structural limitation at the mathematical level by replacing the flat Euclidean metric with learned per-token Riemannian metrics. Our contributions are threefold. (1) We prove that Riemannian attention scores with heterogeneous per-token metrics are non-Gram---they cannot be factorized as QK^T with factorization dimension O(d). We are explicit that this is a structural observation, not a proof of rank preservation. (2) We establish that low-rank metric factors render all geometric operations tractable: geodesic distance in O(d*r) per token and metric inversion in O(d*r^2) via the Woodbury identity---both far below the O(d^3) cost of a general matrix---making Riemannian attention feasible at billion-parameter scale with negligible overhead. (3) We present the Fiber Bundle Transformer, a complete architecture specification in which each token position carries its own Riemannian metric, attention is geodesic distance computation, feed-forward updates use metric-preconditioned steps, and the connection carries explicit curvature and torsion proxies. We derive formal predictions about correctly implemented geometric architectures and identify the central open problem: proving or disproving that heterogeneous Riemannian metrics prevent the rank collapse that row-stochastic attention matrices otherwise cause. This paper presents theoretical analysis and architectural design; empirical validation is the subject of future work.
Contrastive decoding methods such as DoLa improve the factuality of Large Language Models (LLMs) by contrasting the output distributions of mature and premature layers. However, DoLa's dynamic layer selection relies solely on divergences in output vocabulary distributions. In this work, we propose three attention-guided strategies: Attention-JSD, Attention-Entropy-Max, and Attention-Entropy-Min, which leverage structural information carried by internal self-attention mechanisms as a signal for layer selection. Experimental results on TruthfulQA demonstrate that our strategies, particularly Attention-JSD and Attention-Entropy-Min, consistently outperform the original DoLa. We observe significant gains on multi-answer metrics (MC2 and MC3), suggesting that attention distributions can provide a more sensitive signal for resolving factual knowledge than output vocabulary distributions.
The pre-softmax score of an attention head is a bilinear form $score(i,j) = x_i^T M x_j$ in a learned operator $M = W_q^T W_k$. 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 $p=0.029$), and zeroing the per-frequency RoPE phase $Im(M_t)$ 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, $q_BH <= 0.016$). 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.
A companion paper showed that a transformer's feed-forward layer can be rebuilt from explicit fuzzy set operations - intersection, set-difference, and a self-forgetting sequence quantifier - so its hidden units read as named logical operators at no cost to language-model quality. That left the other half of the transformer opaque. Here we carry the same idea into attention and join the two into one model. The mechanism is minimal: a head's value is passed through a sigmoid, so each value channel becomes a readable detector of whether a feature holds at a token. This adds no parameters and leaves the standard head otherwise untouched. A Boolean variant goes further, restructuring the value into an explicit within-token intersection and negation-capable set-difference. In both designs the output projection is left free, not tied to the vocabulary, which is the load-bearing decision: bounding what a head detects while leaving what it writes unconstrained yields selective detectors, whereas constraining the write does not. A bounded value is shaped into a readable detector by two selectivity pressures - one for sparse firing, one for decisive firing at the rails - and which a design wants is not universal. Across five specialized-attention designs at 125M parameters, 44 to 62 percent of value channels become crisp, contextually selective detectors, and their legibility rises with depth rather than crystallizing only on punctuation. Language-model quality is at parity with a conventional baseline. Finally, we couple the Boolean attention to the legible feed-forward layer and train an end-to-end legible-by-construction language model at benchmark parity: its feed-forward units are named set and quantifier operations throughout, and we can take a token it generates and read the named units that compose to produce it.
Lifelong continual learning remains an obstacle on the path to human-like intelligence. Modern transformers show sparks of intelligence with in-context learning. The quadratic nature of attention, however, prohibits transformers from performing this process on arbitrarily long sequences. In this work, we argue that extending in-context learning to lifelong settings is a practical solution for continual learning in AI agents. In particular, we argue that \emph{parametric forms of attention} are needed to understand a lifetime of context with transformers on a fixed hardware budget. These attention mechanisms learn the relationship between keys and their associated values at test-time with parametric regression. Our generalization of parametric approaches (linear attention, state-space models, fast weight programmers, and test-time training layers) contrasts with nonparametric counterparts like softmax attention. They replace the ever-growing key-value cache with an online-trainable neural network, maintaining a constant memory footprint. We highlight how parametric attention currently fall short of lifelong learning due to limited memory capacity or costly online updates. To address these issues, we pose a set of open questions with novel insights to guide the field toward long-horizon agents.
PaTH Attention showed that replacing RoPE's position-indexed rotations with accumulated data-dependent Householder reflections yields strong length extrapolation, though performance degrades at extreme context lengths. We ask whether this depends on Householder-specific structure or reflects a general property of accumulated transformations along source-to-query paths. We study a simpler variant keeping RoPE's block-diagonal SO(2) rotations but replacing position-indexed angles with accumulated token-dependent ones. It shows the same pattern: improved extrapolation then degradation at long contexts. We prove the result extends to accumulated orthogonal transformations satisfying certain regularity conditions: their products become incoherent after finitely many steps, suppressing attention to distant tokens. Accumulated rotations of queries and keys create a finite mixing window independent of context length; per-token suppression learned in training transfers unchanged to any evaluation length, and high-dimensional concentration produces a score gap suppressing far tokens while near-route transport preserves the target signal. Conversely, a lower bound shows accumulated rotations must eventually degrade: as the far set grows, no rotations preserve the near signal without explicit far-mass control. For SO(2) rotations, rotating values too makes residual far contributions combine incoherently, extending the range. Controlled experiments support these predictions: random accumulated rotations substantially improve extrapolation over RoPE, learned token-dependent rotations maintain near-training-length perplexity far beyond the training context, and rotating values helps over queries and keys alone. Rotation-only models still degrade at extreme lengths, while ALiBi stays length-stable, consistent with the need for far-mass control.
We show that the conventional gated MLP can be viewed as a rank-1 approximation to a bilinear attention mechanism with two distinct factors corresponding to the query and the key. We further show that moving the nonlinearity onto one factor breaks the exchange symmetry between the two factors and, for non-homogeneous activations, the inverse-scaling symmetry as well. This perspective may help explain why gated MLPs are effective in practice and inform the design of future architectures.
RoPE-trained transformers distinguish absolute position in their attention patterns, even though RoPE encodes only relative offsets in the inner product. We trace this leakage to two architectural components, The causal mask is responsible for the first: its per-query softmax denominator depends on the absolute query position by construction. The residual stream supplies the second. Under causal attention the activation at position $0$ attends only to itself and runs as a closed dynamical system from the embedding of the token at that position; downstream attention reads this trajectory through sink-reading heads. Both components appear in all three architectures we study, in architecturally specific balance: NTK scaling suppresses the residual-stream component, sliding-window attention allows it to accumulate with depth, and standard RoPE sits between. Replacing the \texttt{BOS} embedding before the forward pass removes $40\%$ of the residual-stream component at early queries. Attention sinks are token-anchored stabilizers that pass forward a deterministic fingerprint of the token at position $0$, constant across inputs when that token is the auto-prepended \texttt{BOS} and varying with it otherwise.
Ali Kayyam, Anusha Madan Gopal, M Anthony Lewiscs.LG cs.AI cs.CL cs.PF
Transformers have become the standard solution for various AI tasks, with the query, key, and value (QKV) attention formulation playing a central role. However, the individual contribution of these three projections and the impact of omitting some remain poorly understood. We systematically evaluate three projection sharing constraints: a) Q-K=V (shared key-value), b) Q=K-V (shared query-key), and c) Q=K=V (single projection). The last two variants produce symmetric attention maps; to address this, we also explore asymmetric attention via 2D positional encodings. Through experiments spanning synthetic tasks, vision (MNIST, CIFAR, TinyImageNet, anomaly), and language modeling (300M and 1.2B parameter models on 10B tokens), we discovered that our transformers perform on par or occasionally better than the QKV transformer. In language modeling, Q-K=V projection sharing achieves 50% KV cache reduction with only 3.1% perplexity degradation. Crucially, projection sharing is complementary to head sharing (GQA/MQA): combining Q-K=V with GQA-4 yields 87.5% cache reduction, while Q-K=V + MQA achieves 96.9%, enabling practical on-device inference. We show that Q-K=V preserves quality because keys and values can occupy similar representational spaces and attention operates in a low-rank regime, whereas Q=K-V breaks attention directionality. Our results systematically characterize projection sharing as an underexplored instance of weight tying in attention, with direct, quantifiable inference memory benefits, particularly valuable for edge deployment. The code is publicly available at https://github.com/Brainchip-Inc/Do-Transformers-Need-3-Projections