Transformer Layers

Transformer architectures exhibit cross-layer redundancies, yet post-training compression pipelines typically optimize layers in isolation or rely on heuristic grouping strategies that disregard layer-specific activation geometries. We introduce a principled, training-free framework that sequentially optimizes cross-layer weight pairings and shared-dictionary factorizations. Rather than forcing weights of adjacent layers to share a basis or heuristically merging activation statistics, our approach identifies structurally compatible projections and learns a shared representation that better preserves each layer's distinct calibration geometry. Coupled with structured sparsity, this yields highly efficient weight decompositions without sacrificing functional fidelity. Across diverse architectures, scales, and modalities, our method achieves state-of-the-art results, consistently outperforming independent structured weight decompositions and alternative pairwise weight factorizations, which operate under heuristic grouping strategies. By replacing heuristic engineering strategies with a convergent, optimization-driven pipeline, we establish a theoretically grounded foundation for scalable, transformer compression across different modalities.

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.

Depth-recurrent language models iteratively apply a small layer stack, decoupling per-token compute from distinct parameter count. To determine whether such a model genuinely utilizes its depth, both recurrence and layer-pruning literatures rely on a shared evaluation: truncating depth at inference time, plotting quality against retained depth fraction, and reading off the slope. While cheap and training-free, this metric suffers from an unexamined flaw: it extracts a single scalar from an intervention that alters multiple model properties simultaneously. Depth truncation concurrently reduces the number of block applications, decreases the volume of distinct computation performed, and pushes the readout head onto an out-of-distribution residual stream. The observed slope conflates all three factors, yet is conventionally interpreted as reflecting solely the second. We propose the Depth Control Protocol (DCP), a diagnostic suite that disentangles these three quantities. DCP comprises three positive controls that isolate each factor while varying the others, a negative control applying the identical interventions to dense transformers to ensure the effect is not an artifact of the measurement protocol, and a controlled training intervention to verify causality. The linchpin control, running the full budget of block applications while executing only a single distinct iteration, is strictly realizable only in depth-wise weight-sharing architectures, since in a dense network repeating a layer yields an entirely different model rather than the same model in an alternative configuration.

Scaling laws predict how loss decreases with increases in computation. We show, contrary to conventional wisdom, that architectural interventions can modify scaling exponents in pre-training, leading to power-law improvements in performance as computation increases. As an anchoring point, we consider the architectural formulation of looped transformers. Although not typically used in this way, looping, also known as recursive depth, provides a mechanism for model growth, by increasing the number of loops during training. Model growth, with and without shared weights, provides the biggest changes to the scaling exponents. In particular, a 7.4B model growth architecture matches GPT-3 13B on CORE with roughly $20\times$ less compute, and has compute efficiency gains that increase with scale. Moreover, simply using a boundary operator in a vanilla transformer, which normalizes and injects an earlier block, also provides an exponent increase, although to a lesser extent. In the data-constrained, multi-epoch setting, standard looping has a useful regularizing effect, where we find it is compute-optimal to increase the number of loops with scale. These results can be understood through the lens of computational depth: for a given computational budget, we wish to increase the usable depth of the transformer, which can lead to efficiency gains that increase with scale.

We analyze the learned input-output behavior of GLU-based neurons in large language models (LLMs). We propose a simple analysis method: For each neuron, we compute the cosine similarities between its input (reading) and output (writing) weight vectors. In this scheme, a strong negative cosine similarity indicates the neuron weakens the direction it detects in the residual stream, so we call this a weakening neuron. This allows us to gain a number of novel insights. First, we show that nine different LLMs have similar patterns: weakening neurons appear mostly in late layers whereas their counterparts, (conditional) strengthening neurons, are frequent in early-middle layers. Second, we find that weakening neurons display surprising behavior: even though there are few, they activate often and have a large influence on model behavior. Third, weakening neurons have a strong effect on model output when gate values are negative -- which is surprising since negative gate values are not expected to encode functionality.

Mixture-of-Experts (MoE) architectures decouple model capacity from computational cost, yet incur high memory footprints as parameters grow linearly with the number of experts. Recurrent Transformers achieve parameter efficiency by reusing layer weights, but typically lack the capacity for competitive language modeling. We propose Mixture of Reused Experts (MoRE), a hybrid that shares expert pools across groups of adjacent layers. Each layer retains its own router but selects from a larger shared pool, expanding the diversity of routing combinations without additional parameters. To enable shared experts to distinguish between layers, we introduce lightweight learnable depth embeddings that condition each layer's input before routing. Experiments across three model scales (114M-1.15B parameters) show that MoRE consistently achieves lower perplexity and stronger downstream performance than standard MoEs and state-of-the-art weight-sharing architectures at matched compute and parameter budgets, with only minimal modifications to existing MoE implementations.

We introduce a simple architectural modification to decoder-only transformers: a persistent recurrent state that observes hidden representations via cross-attention, updates itself through a GRU, and modulates subsequent processing via gated addition. Inserted between the lower and upper halves of a 6-layer transformer, this module adds only 3.7% additional parameters while reducing evaluation loss from $2.438 \pm 0.004$ to $1.743 \pm 0.018$, corresponding to a 28.5% reduction on held-out language modeling data. The improvement is statistically significant across 5 random seeds ($p < 0.01$) and corresponds to reduced overfitting (generalization gap 0.12 vs 0.26). Through controlled ablations, we demonstrate that the improvement stems entirely from the persistent memory topology, not from auxiliary self-prediction objectives. A model with identical topology but no auxiliary loss performs equivalently, while a random auxiliary loss provides no benefit. Representation probing reveals that the persistent state encodes narrative position (52% vs 33% chance level)---information that standard attention maintains less efficiently. Our results suggest that bridging transformer layers with a lightweight recurrent memory is a simple, effective approach to improving generalization in small-scale language models.

We propose a pre-fine-tuning probing method for Parameter-Efficient Fine-Tuning (PEFT) layer selection, aiming to obtain more stable and higher gains with fewer trainable parameters when adapting large vision--language models (VLMs). Unlike the common practice of applying LoRA and other adapters to all layers at once---where layer selection often relies on heuristic rules---we focus on the vision encoder and directly evaluate the "adaptability'' of each Transformer layer. Specifically, we characterize each layer from two perspectives: (i) the statistical properties of its Q/K/V projection weights (e.g., norms and condition numbers); (ii) robustness under controlled parameter perturbations. We then systematically compare these indicators with the downstream performance gains brought by applying PEFT to a single layer. Across experiments covering seven benchmarks and five PEFT variants, we observe a consistent correlation: layers (or matrices) with larger weight norms and higher condition numbers are usually more robust to perturbations and are more likely to yield larger fine-tuning gains. These results show that distribution-statistics analysis and perturbation tests before fine-tuning can provide practical signals for adaptation-layer selection, thereby maintaining or improving performance while reducing trainable parameters.

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

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

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

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

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

The strong performance of foundation models for tabular tasks comes at substantial inference costs. Distilling models into task-specific architectures reduces model size and computational demands but also sacrifices in-context adaptability. Here we introduce TACTICL, an automated task-aware compression framework for tabular in-context learning models that jointly prunes transformer layers and replaces them with lightweight adapters trained on downstream tasks, thus blending in-context with in-weight learning. We study TACTICL on 47 benchmark datasets and show that we can substitute up to 85% of layers without substantial performance drop on a given downstream task. We further show that TACTICL maintains robustness to data shifts, leaving its in-context ability intact. Overall, TACTICL provides a robust framework for exploiting the depth-wise redundancy of tabular foundation models by combining task-specific adaptation and structured compression. We provide the code at: https://github.com/Hebog/tfm_compression

Activation steering edits the behaviour of a frozen language model by adding a learned vector to its residual stream, and current practice fixes the injection layers globally per task. We argue that the best layers are an instance-level decision, and we make per-instance, multi-layer selection both well understood and deployable. On two open-weight 8B models and six binary persona traits, a per-instance oracle over layer subsets shows that the best layers vary from one input to the next: on most trait-model pairs, no fixed global layer set recovers the per-instance benefit. A greedy rule that ranks layers by single-layer marginal effect recovers nearly all of the oracle's benefit, but both must score candidate layers against the gold answer, so neither can run at deployment; the rule instead becomes the target a prompt-only predictor is trained to reproduce. Our deployable recipe needs no label at inference: a per-instance layer ranker read off the prompt embedding, a classifier that infers the steering direction, and an adaptive gate that scores short steered passes against that inferred direction and steers no more layers than necessary. The recipe recovers most of the oracle's lift (the bulk on the stronger model, a clear majority on the harder one), never drives any trait-model pair below its unsteered alignment baseline on average, and largely avoids the fluency collapse that strong global selection incurs at higher layer counts. A mechanistic account, "direction over magnitude", explains the behavioural flip under a mis-directed global set, the output collapse from steering too many layers, and the ceiling of unsteerable inputs.

Standard Large Language Models (LLMs) execute layers sequentially. Dynamic layer routing, i.e. search for a different execution path through layers involving layer repetitions, skips and other moves, can improve performance. Existing routing approaches often require updating model weights, running expensive search loops per test instance, or demand ground-truth labels during inference. In this work, we propose Markov Chain Routing of Transformer Layers (MACRO), a framework that learns task-specific routes over LLM architectures without modifying underlying parameters. MACRO models layer routing as a context-dependent Markov policy conditioned on layer indices, computation budget phases, directional displacements, and operator context, supporting skip, repeat, and residual hidden-state addition operations. The Markov route distribution is updated via feedback on training data and decoded using a top-k Viterbi algorithm to isolate high-probability candidate programs. We evaluate MACRO across diverse reasoning and knowledge benchmarks on multiple open-weight LLMs. MACRO achieves a +5.0% average accuracy improvement over the unrouted baselines, with largest gains on small models. We outperform the best dynamic routing approach Dr. LLM by +7.2%, while reducing route-search time 9.4x (from 14.8 to 1.6 hours). Our code is publicly available at https://github.com/Batorskq/MACRO.

Depth-routing residual architectures allow Transformer layers to retrieve earlier representations instead of inheriting only the immediately preceding state. Existing Block Attention Residuals, however, use a single content-dependent depth mixture to construct the inputs to queries, keys, and values. This design couples two functionally different decisions: queries and keys determine where attention matches, whereas values determine what content is retrieved. We therefore ask whether matching and content retrieval should be forced to read from the same depth. We introduce Role-Decoupled Attention Residuals (RD-AttnRes), a minimal extension that shares one depth route between queries and keys while learning an independent value route over the same residual sources. Tying the two routing queries exactly recovers the parent architecture, while decoupling them adds only one model-width vector per layer and introduces no additional token-to-token attention operation. We evaluate RD-AttnRes using a frozen, paired pretraining protocol on FineWeb-Edu with five matched seeds for both 120M- and 343M-parameter models and a 2.0B-token training budget. RD-AttnRes improves validation negative log-likelihood in all 10 matched comparisons. The mean reductions are 0.0301 and 0.0247, corresponding to perplexity reductions of 2.97 percent and 2.43 percent at 120M and 343M parameters, respectively. Early-budget controls indicate that neither the additional parameter count, duplicated routing execution, nor a fixed value route reproduces the improvement. Routing diagnostics further reveal persistent divergence between the query-key and value depth distributions. These results suggest that, within the evaluated training regime, attention matching and content retrieval benefit from distinct reads over the residual hierarchy.

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.

While large language models (LLMs) can solve advanced reasoning problems in seconds, we show that even frontier models fail to perform a much simpler operation: exactly copying an input string that lies well within their context windows. We attribute this failure to positional encodings in Transformer architectures, whose inductive bias favors copying through a shortcut based on matching local contexts rather than carefully locating the corresponding input positions. To address this issue, we introduce 2D-RoPE, which organizes text into a 2D grid rather than a 1D sequence and assigns each token a row ID and a column ID. Under this view, copying becomes simply retrieving input tokens at a fixed column offset, which makes the task easy to learn. In synthetic copy experiments, shallow Transformers with 2D-RoPE achieve perfect copying at input lengths hundreds of times longer than those seen during training, whereas standard positional encodings fall far behind. We further show that the advantage of 2D-RoPE language models on copy tasks consistently holds in large-scale pretraining on DCLM with model sizes up to 1.4B parameters. Overall, our results suggest that viewing text in 2D can benefit language modeling, and we hope this encourages future work to further explore the potential of 2D positional encodings.

In pre-LayerNorm looped transformers, LayerNorm inside the recurrent block acts as an implicit gain controller: by coupling the block's local Lipschitz constant inversely to the activation scale, it renders the recurrence Jacobian non-normal -- asymptotically contractive at every verified fixed point even where its operator norm exceeds 1 -- so the true stability budget is the spectral margin, not an operator-norm bound. That margin depletes as the carry $ρ\to 1$, and a minority of initializations never converge to a fixed point at all, so the diagonal carry constraint $ρ(\bar{A}) < 1$ is necessary but not sufficient for convergence of the full recurrence. Training experiments across six tasks, including a controlled ablation, reveal that the linear carry is not the depth-memory mechanism: gradient descent routes memory through the block's more expressive nonlinear recurrence and leaves the stability-constrained carry at rest -- the carry's role is stabilization, not memory. We characterize the boundary of this claim: on tasks with axis-aligned per-channel structure, gradient descent does recruit the carry. All results are derived analytically and verified in a from-scratch, CPU-scale implementation; verification at larger scale is needed.

Fully homomorphic encryption (FHE) enables computation on encrypted data, but practical encrypted Transformer inference is bottlenecked by the sequential composition of many nonlinear blocks. We study whether Structured Newton Layer Parallelism (SNLP) can make this inter-layer composition more FHE-friendly: each Transformer block still requires polynomial approximations for operations such as softmax and RMSNorm, but SNLP reduces the layerwise sequential nonlinear depth from L stages to a small number of solver iterations plus linear structured corrections. Using a simulation framework based on Chebyshev polynomial approximations, we measure error accumulation under sequential versus SNLP inference across 8 models and 4 architecture families. On a 0.5B IDN-trained model, SNLP reduces symbolic bootstraps from 53 to 20 (2.65x) with only +1.2% perplexity degradation, while lowering error amplification (1.36x vs. 1.42x). Across all tested models, SNLP has lower amplification than sequential inference. Ablations show that softmax approximation dominates the error budget and CKKS arithmetic noise is negligible in our setting, suggesting that SNLP is complementary to block-level FHE-friendly operator design rather than a replacement for it.

Reinforcement learning (RL) has become a central component of post-training large language models (LLMs), yet little is understood about how RL adaptation is distributed across transformer layers. Existing approaches typically update all model parameters uniformly, implicitly assuming that every layer contributes similarly to the gains obtained during RL post-training. In this work, we challenge this assumption through a systematic layer-wise study of RL training. Surprisingly, we find that training a single transformer layer can recover most of the gains achieved by full-parameter RL training, and in some cases even surpass it. To quantify this phenomenon, we introduce the quantity layer contribution, which measures the fraction of full RL improvement recovered by training a layer in isolation. Across seven models spanning two model families (Qwen3, Qwen2.5), three RL algorithms (GRPO, GiGPO, Dr. GRPO), and multiple task domains including mathematical reasoning, code generation, and agentic decision-making, we observe a remarkably stable pattern: RL gains are highly concentrated in a small subset of, and in many cases even a single, transformer layers. More strikingly, the same structural pattern consistently emerges: high-contribution layers concentrate in the middle of the transformer stack, while layers near the input and output ends contribute substantially less. The resulting layer rankings remain strongly correlated across datasets, tasks, model families, and RL algorithms.

Deep neural networks with repeated architectural blocks, such as transformers, often exhibit structured relationships across layers that emerge during training. Motivated by this observation, we introduce \emph{Depth-wise Gradient Augmentation}, a general optimization paradigm in which the update applied to each layer is obtained by transforming the collection of block-wise optimizer updates along the depth dimension. Within this framework, we study \emph{Gradient Smoothing}, a family of depth-wise smoothing methods, and instantiate it with a simple local \emph{Window Smoothing} operator. The resulting method operates directly on block-wise updates produced by arbitrary base optimizers (e.g., SGD, Adam, Muon), incurs minimal computational overhead, and is compatible with existing optimization pipelines. We evaluate Gradient Smoothing across a diverse set of architectures and training regimes, including language model pretraining, RL post-training of LLMs for reasoning, diffusion modeling, and image classification with Vision Transformers. Across these settings, Gradient Smoothing consistently improves optimization and generalization performance without modifying model architectures or training objectives. We further show that it promotes more structured representation evolution across depth, consistent with its interpretation as a structured depth-wise preconditioning method. Together, these results establish Depth-wise Gradient Augmentation as a promising framework for exploiting cross-depth structure in optimization and demonstrate Gradient Smoothing as a simple and broadly applicable instantiation.

Audio deepfake detection systems often fail to generalize across domains because they rely on features tied to specific attacks or recording conditions. Self-supervised speech models offer rich multi-layer representations, yet existing approaches either use a single layer or fuse all layers indiscriminately, and only reveal layer importance after training. We propose a model-agnostic, two-stage methodology that identifies informative depth zones before any task-specific model is trained. In the first stage, lightweight XGBoost probes evaluate each transformer layer's cross-domain discriminative power, producing a layer ranking. In the second stage, a compact neural classifier fuses only the selected layers through per-layer attention pooling and a shared bottleneck projection, while the backbone remains frozen. Applied across three backbones, the probing reveals two key findings. First, informative layers cluster in depth zones rather than at uniquely optimal positions: within-zone substitutions fall within multi-seed noise, while zone violations degrade performance by up to 5x. Second, the probing produces backbone-specific selections rather than a fixed layer recipe. On XLS-R-300M, four probing-selected layers with 1.34M trainable parameters achieve 4.94 +/- 0.32% equal error rate on In-The-Wild and 5.07% cross-domain average over four shared datasets, a 28% relative improvement over the best prior frozen-backbone result (Xiao and Vu, 2025) using all 25 layers with identical training data.

We ask a simple question about decoder-only transformers: between which two layers is the probability of a predicted token actually produced? Existing layer-wise readout tools answer only approximately. The logit lens and its trained variant report a per-layer level of probability but give no additive decomposition; their estimates are biased and non-monotone across depth. Direct Logit Attribution and related residual-stream methods are additive, but only in logit space, the softmax nonlinearity breaks additivity in probability space, precisely the quantity one usually cares about. Layer Conductance integrates gradients per layer, but attributes each to its own baseline and so does not sum to the total change in prediction. We introduce IG-Lens, a telescoping application of Integrated Gradients along a single path through the hidden states from a baseline to the final layer. Crediting each segment to the layer it terminates at yields a layer-wise attribution whose sum is exactly the change in target probability, with the softmax inside the integration path rather than linearized away. Our default estimator credits each integration step its observed change in target probability (a prediction-aware reweighting in the spirit of IDGI) rather than its raw gradient. Because the readout is a one-dimensional probability, this collapses each segment to a telescoping sum of endpoint values, so completeness holds exactly (to floating point) at any step count, removing Riemann discretization error while suppressing steps that show gradient sensitivity without a change in output. We give the telescoping identity and its proof, verify completeness to floating point, and describe a single-pass batched implementation computing the full token-by-layer map without any backward call. Code: https://github.com/anhnda/IGLens.

We introduce the context-ready transformer, a new recurrent neural network architecture built from a D-layer transformer block that pre-contextualizes each token before it enters the block. During left-to-right generation, a correction network combines the previous position's block output -- a cached summary of past context -- with the current token embedding, so the tokenenters the block already contextualized rather than as a raw embedding. At sequential inference, the correction chain makes the architecture a recurrent neural network. For training, we unroll the correction process K times over the full sequence, processing all positions in parallel at each step. A pretrained transformer can also be converted to a context-ready model by adding a zero-initialized correction FFN and fine-tuning. We evaluate across widths, depths, block sizes, and two datasets, with all comparisons against standard transformers, variants, and ablations. A D=5 model beats a 12-layer transformer while generating 1.7x faster on an A100. With K=10, a single-layermodel (D=1) beats a 6-layer transformer with a 2.6x inference speedup, and sequential inference matches parallel K=10 to within 0.01 PPL. The architecture benefits most from wide representations and long contexts. On a pointer-chasing task, D=1 trained with BPTT solves all 10 composition levels, while standard transformers exhibit staircase-like depth dependence.

Transformers achieve strong performance, but their internal computations remain opaque. We view each Transformer layer as a dynamic graph whose nodes are token representations and per-head attention outputs, with Multi-Head Attention (ATT) and MLP as module boundaries. On this graph we use LIG (Layer-wise Integrated Gradients), which applies set-to-set Integrated Gradients (IG) at nonlinear module boundaries. Set-to-set IG applies IG to a map from a set of input token representations to a set of output representations, evaluating token-to-token contributions, which is not standard in prior IG applications. This extends IG from the usual scalar-objective setting to set-to-set maps via an L2 scalarization, and composes within-layer contributions in the spirit of Layer-wise Relevance Propagation (LRP), with IG completeness playing the role of LRP-style conservation at each boundary. We use LIG to analyze (i) the agreement between module-wise composition and layer-whole attribution under an L2 criterion, and (ii) within-layer information flow by tracing separated ATT and MLP contributions. On BERT-base and PTB, configurations that best preserved within-layer consistency used the target token's embedding as the ATT baseline and either the ATT output at a=0 or Zero as the MLP baseline. We therefore present LIG as a diagnostic XAI tool at module-boundary granularity, without model-specific retraining or per-operation interpreter design. Code is available at https://github.com/eightsuzuki/layer-wise-integrated-gradients.

Standard Transformers use a single self-attention pathway to model both global dependencies and local patterns, creating tension between long-range structural reasoning and fine-grained local representation learning. We propose a FiLM-coordinated dual-branch Transformer for language modeling, where each layer explicitly contains a global branch and a local branch, and feature-wise linear modulation (FiLM) is used for dynamic cross-branch coordination instead of simple concatenation or static addition. The key idea is that the two branches represent different dependency views of the same input, making channel-wise calibration more suitable than heavy token-level interaction. We therefore design a bidirectional FiLM module in which each branch generates per-channel scaling and shifting parameters to condition the other. Experiments on multiple small-scale language modeling settings show that the proposed structure consistently outperforms same-width single-branch baselines and weakened dual-branch variants under a fixed lightweight configuration. On TinyShakespeare and a 1M-character subset of WikiText-2, the full dual-branch FiLM model achieves the best results among same-width structural baselines. Multi-seed results support the stability of the gains, while mechanistic analyses show that FiLM learns input-dependent, layer-dependent, and channel-selective modulation patterns rather than static scaling. Parameter-matched widened single-branch baselines also indicate that the current design still leaves room for improvement in parameter efficiency.

We study internalization processes, by which neural-network-based systems absorb an explicit computational procedure into their own weights, and how they facilitate learning. We investigate how transformers internalize the simulation of semiautomata by internalizing chain-of-thought (CoT) tokens, which classes of semiautomata are harder to internalize, and expose the flip side of internalization, that is, a progressive degradation of out-of-distribution performance. We then provide the first provable analysis of successful internalization: for the task of learning parities, we show that a simplified one-layer transformer provably first learns the target with explicit CoT supervision and then internalizes the autoregressive generation as CoT tokens are progressively removed, learning to directly compute the parity. This task is computationally hard to learn from data without CoT supervision. Finally, we discuss how learning through internalization relates to the \textit{Positive Distribution Shift} phenomenon recently introduced by~\citet{Med+26}.

Deep neural networks are widely believed to derive their expressive power from their ability to form \textbf{hierarchical representations}, capturing progressively more abstract and compositional features across layers. In language modeling, \textbf{transformers} have emerged as the dominant architecture, with early layers capturing local syntactic patterns and later layers encoding more complex clause-level dependencies. While this intuition has shaped model design, there remains a lack of rigorous theoretical work demonstrating \textbf{how} deep transformers represent such hierarchical structures. In this work, we analyze the expressiveness of deep transformer models through the formal lens of bounded-depth, non-recursive context-free grammars. For this class of grammars, we explicitly construct transformers with positional attention whose depth grows linearly with grammar depth, while the neuron count scales with the number of derivation-tree shapes and quadratically with the number of production rules. Our theoretical results support the linear representation hypothesis by demonstrating that these architectures possess the structural capacity to encode abstract grammatical states into low-dimensional, linearly separable subspaces within the residual stream.

The considerable layer-wise redundancy in large language models (LLMs) has established non-uniform sparsity allocation across layers as the standard pruning approach for efficient compression. Existing layer-wise allocation methods that estimate allocation strategy from local signals such as activation outliers or weight spectra mainly derive from local layer importance, whereas the final post-pruning performance is also influenced by the network's subsequent compensatory capacity. In this paper, we directly characterize this property through controlled perturbation experiments. We make the following empirical findings. First, layers exhibit highly heterogeneous responses to pruning-scale perturbations. In most cases, early layers amplify perturbations, while middle and late layers actively absorb them, with relative L2 drift decreasing monotonically across depth and direction realigning toward the unperturbed hidden-state trajectory. Second, absorption is a large-perturbation phenomenon. Under small perturbations the network exhibits amplification across all layers, and the transition to absorption occurs smoothly as perturbation magnitude grows to pruning scale. This enriches the linearized accumulation theory underlying related works. Building on these findings, we define an absorption coefficient per layer and propose absorption-aware correction, an orthogonal augmentation that improves OWL and AlphaPruning by reducing perplexity by 7.13% and boosting zero-shot accuracy by 1.02% across multiple model families at 70% sparsity.

KV cache eviction methods typically use a single retention-rule family throughout a model, making eviction-method identity a model-level design choice. Yet Transformer layers differ substantially in their attention behavior, representations, and sensitivity to compression, suggesting that a uniform rule may overlook useful layer-wise structure. This raises a basic question: should eviction methods themselves vary across layers? We investigate this question by composing existing eviction methods across Transformer layers and systematically exploring the resulting layer-wise design space. Using simple offline profiles, we construct fixed routes and study how their quality varies with method placement and cache budget. On LongBench, heterogeneous routing improves performance on a majority of tasks over homogeneous policies at the same cache budget. Even when method counts are held fixed, the profile-guided placement ranks second among 100 evaluated assignments, demonstrating that routing quality depends strongly on where methods are placed. Moreover, the same fixed route outperforms the best of nine standalone baselines across all five tested cache budgets. Together, these results establish layer-wise method composition as an exploitable, placement-sensitive design dimension for KV cache compression.

Mixture-of-experts (MoE) layers enable the scaling of transformer models while keeping the inference compute fixed. While task-expert specialization has been observed in empirical studies of frontier MoE transformer models, existing theoretical work analyzes this using continuous mixture models that cannot be used to model natural language effectively. An important open question is to \textit{theoretically explain task-expert specialization in transformer MoE models using discrete models of language}. To address this, we represent structured knowledge via syntactic templates and finite key-value dictionaries, and prove formally that a single-layer MoE transformer can encode knowledge by using experts that specialize in the corresponding tasks. Our construction shows how queries are routed to unique, task-specific experts whose size depends solely on the intrinsic complexity of the given task (i.e. the combined size of its syntactic templates and factual dictionary). Our construction provides a theoretical support for empirical results on localized knowledge circuits in MoE models. We support our theoretical findings with experiments evaluating model performance under varying MoE loss functions.

Sparsity allows scaling model parameters without proportionally increasing computational cost. While mixture of experts (MoE) models are made increasingly sparse, individual experts typically remain large and dense. Here, we demonstrate that further increasing sparsity by shrinking each expert to consist of a single neuron and selecting a tiny fraction of many available neurons can improve compute efficiency and interpretability. Counterintuitively, the key to achieving both is removing the nonlinearity typically applied to the experts, resulting in a network of sparsely gated linear neurons (sgatlin). In an isoflop comparison, we find that replacing all transformer feedforward layers with sgatlin improves perplexity in language models across different compute budgets. At the same time, the sparsity and linearity of the resulting feedforward circuits present new opportunities for model interpretability. In a small-scale case study, we demonstrate that feedforward circuits in sgatlin can be interpreted without having to train additional replacement models. We find that they form semantically structured clusters and are causally implicated in factual recall. Our findings paint a possible path towards compute-efficient and interpretable transformer feedforward layers.

This paper uses geometry to explain how a machine learning model can be stolen using an already existing well-known method. The author has shown the exact conditions required to perfectly copy the final layer of a transformer network. When looking deeper into the hidden layers the author has explained clear limits. The author has also demonstrated that a hidden network cannot be fully reverse engineered just by looking at the final results. The research clearly maps out what can and cannot be stolen from a model.

Zeroth-order (ZO) optimization enables memory-efficient fine-tuning of large language models (LLMs) using only forward passes, but it remains unclear how useful adaptation is distributed across layers. In this work, we reveal a surprising phenomenon: ZO fine-tuning is sharply dominated by a single decoding layer. Across multiple LLM families and downstream tasks, fine-tuning this dominant layer alone consistently matches or even exceeds full-model ZO fine-tuning. We further show that the dominant layer is task-agnostic but model-specific, and can be identified before training through a simple inference-only analysis of activation outliers. Specifically, the dominant layer consistently aligns with the first activation-outlier layer in the pre-trained model. To explain this phenomenon, we analyze how perturbation effects propagate under ZO optimization. We find that the dominant layer combines two key properties: high perturbation sensitivity and early placement in the residual stream, allowing perturbation-induced effects to propagate and accumulate through remaining subsequent decoding layers. As a result, this layer produces disproportionately strong and stable optimization signals under forward-only updates. Extensive experiments on LLaMA2-7B and Qwen3-8B across nine benchmarks show that dominant-layer ZO fine-tuning improves average performance over full-model MeZO and LoRA-based ZO fine-tuning while achieving up to 4.52$\times$ training speedup.

Low-Rank Adaptation (LoRA) is a popular parameter-efficient fine-tuning (PEFT) method that restricts weight updates to low-rank adapters, introducing a fixed low-rank inductive bias by optimizing in a low-dimensional subspace. In this work, we question whether a fixed-rank constraint is the most effective inductive bias for parameter-efficient fine-tuning. We introduce Learnable Rank LoRA (LR-LoRA), a PEFT method in which the adapter rank is learned during the training process. Instead of prescribing a uniform rank for all adapter layers, LR-LoRA allows the optimizer to determine the appropriate rank for each layer. Using this approach, we find substantial layer-wise variation in the learned ranks, with the attention and MLP layers in the transformer models exhibiting systematically different rank preferences. Across a range of language understanding and commonsense reasoning benchmarks, LR-LoRA achieves state-of-the-art performance in most settings and consistently outperforms strong PEFT baselines, demonstrating that a learnable rank provides a more flexible and effective inductive bias than fixed-rank adaptations.

Agentic language model systems alternate between two structurally distinct step types: structured tool calls (short, deterministic, low perplexity) and open-ended planning/reasoning steps (long, complex, high perplexity). Despite this heterogeneity, current inference systems apply identical compute to every step. We introduce LayerRoute, a lightweight adapter that learns to selectively skip transformer blocks on a per-input basis. LayerRoute augments each of the 24 transformer blocks in Qwen2.5-0.5B-Instruct with: (1) a per-layer router (~897 parameters, Linear(896,1)) that outputs a hard binary gate via the straight-through estimator, and (2) LoRA adapters (rank 8, ~1.08M parameters) on the Q/K/V/O attention projections. The backbone weights remain frozen. A single end-to-end training pass on agentic data (Hermes, Glaive, GSM8K, Turing) with a gate regularisation term forces the system to discover which blocks are skippable per input type. After 3,000 steps (6.4 minutes on an A100 40GB), LayerRoute achieves a 12.91% skip differential: tool calls skip 15.25% of FLOPs while planning steps skip only 2.34%, using only 1.10M trainable parameters (0.22% of the 494M backbone). Quality improves over the base model due to LoRA adaptation, with perplexity delta of -1.29 on tool calls and -1.30 on planning.

Probes are routinely paired with an intervention: ablate the direction the probe found, run the model, and read the change in task accuracy, taking a large drop as evidence that the computation depends on what the probe read and a near-zero drop as evidence that it does not. Either inference requires that the ablation have removed the target from the layer. We find that the ablation does not remove what it targets. A probe refitted on the ablated activations recovers its original accuracy in every cell we test, and keeps recovering when the probe's entire row space is deleted rather than a single axis, because the quantity survives in the orthogonal complement. Because a refitted probe recovers, neither a large task drop nor a near-zero one establishes whether the model needed the target, and one probe fit detects this. Replacing the ablation with iterative nullspace projection, scored against random subspaces of matched dimension, reverses the conclusion: representations that looked causally inert carry most of the task. The correction also separates where a variable is most readable from where deleting it does most damage, and those are not the same layer in any pretrained model we study. The erasure is defined by a linear probe family, so removing a nonlinearly encoded quantity remains open.

Large Language Models (LLMs) have achieved strong performance across diverse natural language tasks, yet their outputs often suffer from hallucinations -- content that is misaligned with factual information. In this work, we conduct a comprehensive layer-wise analysis of the decoding process and reveal that hallucinations tend to originate from deeper decoder layers. To address this issue, we introduce \textbf{DeLask} (\textbf{De}coder \textbf{La}yer \textbf{Sk}ipping), a novel decoding framework that dynamically skips layers prone to producing hallucinations. DeLask leverages the theoretical insight that the forward computation of an $L$-layer Transformer is conditionally equivalent to $L$ steps of gradient descent. We define a \emph{driftance value} by computing the cosine similarity between gradients derived from consecutive decoder steps, identifying problematic layers when the descent direction reverses. Rather than discarding such layers entirely, DeLask partially aggregates their hidden states with preceding layers, thereby preserving consistency while suppressing erroneous signals. Extensive experiments across diverse LLMs and benchmarks demonstrate that DeLask consistently mitigates hallucinations and enhances overall reliability, providing a lightweight and generalizable decoding framework for improving the robustness of large-scale language models.

Transformers have achieved remarkable success across a wide range of applications, and a growing body of work suggests that part of their strength comes from their ability to learn and execute algorithmic procedures. However, our understanding of how transformers learn such algorithms remains limited, especially in the presence of layer normalization (LN). In this work, we study principal component prediction as a concrete testbed for understanding the training dynamics of transformers with LN. We prove that a looped linear transformer with LN, trained by gradient descent, converges to a solution that implements the power method, with each self-attention layer performing one power iteration. Notably, the model is trained only for principal component prediction, rather than being explicitly supervised to implement the power method. Our finding thus reveals an "algorithmic implicit bias" of looped transformers with LN: principal-component prediction can in principle be achieved by many mechanisms, yet gradient descent selects one that realizes the power method. We further provide a concrete comparison between transformers with and without LN: even with layerwise guidance from power iterations, a transformer without LN cannot exactly learn the power method, whereas the corresponding transformer with LN can, leading to a provable performance gap in principal component prediction. Our results provide, to our knowledge, the first theoretical analysis of the training dynamics of looped and single-layer transformers with LN, and shed light on the role of LN in transformer models.

Recent SVD based compression methods for large language models like SVD LLM and Basis Sharing can be unified under one optimization problem. While mathematical proofs and tests on Pythia models show this unified approach improves weight reconstruction error by up to 46% percent it fails in practical tasks. Downstream metrics like perplexity and accuracy severely degrade compared to standard per layer SVD LLM. The authors explain this failure mechanistically. Although the bundle method mathematically couples adjacent layers the transformer residual stream actually decouples them during forward passes. Thus per layer optimality matters more than joint cross layer optimization. The paper concludes that weight space reconstruction is a flawed objective for cross layer compression and future methods must focus on per layer activation reconstruction instead.

Looped transformers apply a shared block multiple times and have emerged as a parameter-efficient route to scaling compute in language models. However, at fixed FLOPs a looped model has strictly less capacity than a baseline transformer. We propose a novel dual-path block that can flexibly scale compute, the number of sequential operations applied to a hidden state, and capacity, the parameters available at a single step. For this we expose both axes as parallel pathways within a single layer: a deep sublayer re-applied K times with shared parameters, and a wide sublayer with an enlarged feed-forward network applied once. Independent per-token gates combine both axes and allow detailed per-token routing analyses. We show that across two FLOP budgets, our dual-path model surpasses iso-FLOP matched models on language modeling and downstream evaluations, while using fewer parameters than the baseline at matched FLOPs. The learned gates are directly interpretable and show systematic per-token allocation with function words and lexical content trend wide, while punctuation, symbols, and arithmetic tokens trend deep.

Chain-of-Thought (CoT) prompting substantially improves the sample efficiency of transformers, reducing the complexity of tasks like parity learning from exponential to polynomial in the input length. However, generating explicit reasoning steps at inference is computationally expensive. Implicit Chain-of-Thought (ICoT) has emerged as a promising empirical remedy that trains models to internalize intermediate steps within their hidden states, but its theoretical foundations remain poorly understood. We give the first theoretical analysis of ICoT, proving that an $L$-layer transformer trained under our proposed Log-ICoT curriculum learns $k$-parity with $\mathsf{poly}(n)$ samples and $L = \log_2 k$ training stages. This matches the sample efficiency of explicit CoT while eliminating its inference overhead, and extends prior one-layer parity guarantees to multi-layer architectures. Compared to standard ICoT, which removes thinking tokens one at a time, Log-ICoT removes them in geometric chunks, reducing the number of stages from linear in $k$ to logarithmic. Experiments on multi-layer transformers confirm the theory and visualize how reasoning is progressively absorbed into deeper layers.

Sparse autoencoders are usually trained one layer at a time, even though transformer residual stream activations are strongly coupled across depth. This creates a practical problem for multi-layer interventions: different layerwise dictionaries can spend capacity representing the same carried-forward information, and replacing several layers at once can produce interactions that are not predicted by single-layer behavior. We introduce Residualized Sparse Autoencoders (ReSAEs), which fit an affine map between selected layers and train each later-layer SAE on the unexplained residual rather than on the full activation. Reconstructions are mapped back into the original activation space through the fitted affine chain, so ReSAEs can be evaluated with the same intervention protocols as ordinary SAEs. On Pythia-1.4B and Gemma-2-9B, residualization reduces decoder redundancy and improves sparse probing and targeted perturbation in most tested settings. Despite reconstructing less of the raw activation variance, ReSAEs recover more transformer cross entropy under multi-layer replacement. This gain is clearest under teacher-forcing and at sufficient sparsity online, indicating that ReSAEs preserve the components of the activation most relevant to the model's downstream computation. These results suggest that removing linearly predictable cross-layer structure is a useful default for multi-layer SAE interventions.

Feedforward network (FFN) layers account for a large fraction of parameters and nonlinear expressivity in Transformer-based large language models (LLMs). Despite the evolution from ReLU and GELU to gated variants such as SwiGLU, most FFN designs still use a single fixed activation function, applying the same nonlinear transformation to all tokens. In this work, we propose Mixture of Activations (MoA), a token-adaptive FFN design that mixes a dictionary of activation functions using lightweight input-dependent gates while sharing the same linear projections. As an input-independent counterpart, we also introduce learnable activations (LA), which form linear combinations of activation functions for both ReLU-type and SwiGLU-type FFNs. Theoretically, we establish strict finite-width expressive separations among fixed-activation FFNs, LA, and MoA: LA strictly contains fixed-activation FFNs, while MoA strictly contains LA, with the additional expressivity arising from input-dependent nonlinear hybridization. Empirically, we evaluate MoA through extensive pre-training experiments on dense and MoE language models ranging from 0.12B to 2B parameters under different token budgets, optimizers, and learning rate schedules. MoA consistently achieves lower terminal loss and exhibits more favorable scaling behavior than well-tuned baselines, with minimal parameter and computational overhead. These results suggest that token-adaptive activation mixing is a simple and effective mechanism for improving FFN expressivity in LLMs.

In the paper we show that there is an analogy between the operations occurring in a layer of a transformer (projections and layer normalizations, disregarding the feedforward neural network) and a step in the power method. Coherently with this analogy, we show that passing through a layer the tokens tend to be tilted towards the principal eigenvector of a matrix which is the product of the output and value weight matrices of that layer. In the special case of a transformer with shared weights (i.e., in which all layers have identical weights) then the alignment with this principal eigenvector is particularly evident empirically, and can also be shown analytically. The analogy also suggests a method to steer the output of the transformer towards an arbitrary desired direction in token space.

The observation that activation sparsity emerges in MLP blocks of standardly trained Transformers offers an opportunity to drastically reduce computation costs without sacrificing performance. To theoretically explain this phenomenon, existing works have shown that activation sparsity does not result from the data properties or data fitting but from the implicit bias of the training process. However, these connections are obtained with strong assumptions, which cannot be applied to deep models standardly trained with a large number of steps. Different from these works, we find that the flatness of loss landscapes is also closely related to the MLP activation sparsity and can serve as a weaker and naturally emerging assumption standard deep networks. Specifically, we find that 1) the MLP activation sparsity equals a ratio between "augmented flatness" (a weighted sum of flatness measures) and the product of the input norm and activation gradient of the MLP. We empirically find that this ratio decreases during training, leading to sparse activations. 2) We also propose the notion of derivative sparsity, which reduces to activation sparsity under ReLU, but further enables pruning in the backward propagation and is more stable than activation sparsity. With the theoretical findings, we can further encourage activation sparsity by decreasing the numerator and increasing the denominator of the ratio using three methods. These plug-and-play modifications can effectively reduce the ratio and produce sparser activations. Experiments on ImageNet-1K and C4 demonstrate relative improvements of at least 36% on inference sparsity and at least 50% on training sparsity over vanilla Transformers, indicating further potential cost reduction in both inference and training

Independently trained transformers compute the same function in residual-stream bases that differ by a uniform random rotation on $\mathrm{SO}(d_{\mathrm{model}})$. We call this phenomenon polymorphism: same function, mutually unintelligible interior coordinates. One matrix multiplication per model pair removes it: an orthogonal Procrustes fit on a single batch of activations transfers sparse-autoencoder feature dictionaries and steering vectors between independently trained models, with no retraining. The phenomenon is invisible to the standard SAE universality metric. Decoder-column cosine similarity matches across seeds at 98%, the SAE-universality headline number, while an SAE trained on one seed reconstructs another seed's activations at negative explained variance, worse than predicting the constant mean. The decoder columns align; the encoder reads from a rotated frame. A single Procrustes rotation $R$ restores reconstruction to within 0.025 EV of the within-seed ceiling at every internal site. $R$ is Haar-distributed: $\|R - I\|_F$ matches the random-orthogonal prediction $\sqrt{2 d_{\mathrm{model}}}$ to 0.1% at $d_{\mathrm{model}} = 512$, and a Kolmogorov-Smirnov test of $R$'s eigenvalue spectrum against Haar $\mathrm{SO}(d_{\mathrm{model}})$ returns $p \approx 1.000$ pooled and per-pair. Diff-of-means steering vectors transfer in three regimes by alignment with $R$'s invariant subspace: clean when pinned by shared output weights, partial when overlapping the rotated subspace, inverted otherwise. With no shared I/O (Pythia), all three collapse to universally inverted. The same rotation account holds across training checkpoints within a single run. Validated on a 104k-parameter Dyck-3 transformer and nine independently-trained Pythia-70m seeds on The Pile, via a pre-registered four-bar operational framework. Frontier-scale (10B+) replication remains open.

We introduce the Preisach Attention Layer (PAL), a novel sequence modelling architecture grounded in the classical Preisach hysteresis operator from mathematical physics. PAL replaces the softmax attention mechanism with a binary relay operator parameterised by learned activation and deactivation thresholds, maintaining a stack of local extrema as its internal state. A single-layer PAL-Transformer with O(1) depth is Turing-complete under arbitrary precision arithmetic, achievable through simulation of a two-stack pushdown automaton -- in contrast to the O(log n) depth required by standard hard-attention transformers. Second, we prove that the function classes computable by PAL and by the transformer are incomparable: PAL computes historical range statistics in O(1) layers that require O(log n) layers for transformers, while transformers support random-access retrieval that PAL cannot perform without auxiliary state. The separating property is rate-independence -- PAL responds only to the sequence of local extrema, not to absolute token positions or temporal spacing. Third, we show that the extremum stack constitutes a minimal sufficient statistic of the input history for all rate-independent functionals, providing a formal analogue of the wiping property in classical hysteresis theory. PAL is thus an efficient architecture for tasks with long episodic memory and weak positional dependence, with O(n log n) total inference cost versus O(n^2) for standard attention.

Foundational Models pretrained on huge amount of data learn representations that evolve across depth, forming a hierarchy of embeddings with distinct semantic content and geometric structure. Contrary to the widespread practice of using only the final layer or shallow mixtures, we show that task-relevant information is distributed non-monotonically across layers and cannot be recovered by naïve aggregation. Through a geometric and empirical study across multiple modalities, we show that effective transfer depends on identifying which layers encode task-discriminative structure and how their embeddings are geometrically organized. We introduce Layer-wise Optimal Embedding Selection (LOES), a constructive spectral method that identifies task-discriminative subspaces by minimizing residual error under orthogonality and isotropy constraints. To align fine-tuning with this selection principle, we further propose Geometric Regularization Loss (GeoReg), which enforces a simplicial structure on class manifolds and stabilizes representation geometry during fine-tuning. Across a wide range of architectures, depths, modalities, and data regimes, LOES consistently outperforms standard baselines, with gains that grow as model depth increases. Beyond accuracy, our method reveals how semantic factors are distributed across layers, thereby enabling cross-lingual and cross-modal interpretability analyses. Together, our results provide strong evidence that layerwise embedding geometry is not incidental but central to how deep models represent and transfer knowledge.

Learning rate configuration is a fundamental aspect of modern deep learning. The prevailing practice of applying a uniform learning rate across all layers overlooks the structural heterogeneity of Transformers, potentially limiting their effectiveness as the backbone of Large Language Models (LLMs). In this paper, we introduce Layerwise Learning Rate (LLR), an adaptive scheme that assigns distinct learning rates to individual Transformer layers. Our method is grounded in Heavy-Tailed Self-Regularization (HT-SR) theory, which characterizes the empirical spectral density (ESD) of weight correlation matrices to quantify heavy-tailedness. Layers with weaker heavy-tailedness are assigned larger learning rates to accelerate training, while layers with stronger heavy-tailedness receive smaller learning rates. By tailoring learning rates in this manner, LLR promotes more balanced training across layers, leading to faster convergence and improved generalization. Extensive experiments across architectures ranging from LLaMA to GPT-nano, optimizers including AdamW and Muon, and model scales from 60M to 3B parameters with up to 100B training tokens demonstrate the effectiveness of LLR. LLR achieves up to 1.5x training speedup and consistently outperforms uniform-learning-rate baselines. In particular, it improves the average zero-shot accuracy of 1B models from 47.09% to 49.02%, and that of 3B models from 48.58% to 50.61%. A key advantage of LLR is its low tuning overhead: it can transfer nearly optimal learning-rate settings directly from the uniform baseline. Code is available at https://github.com/hed-ucas/Layer-wise-Learning-Rate.

Existing expressivity results for transformers typically rely on hardmax attention, high precision, and other architectural modifications that disconnect them from the models used in practice. We bridge this gap by analyzing standard transformer decoders with softmax attention and rounding of activations and attention weights, while allowing depth and width to grow logarithmically with the context length. As an intermediate step, we construct hardmax transformers with ternary activations and well-separated attention scores that simulate Turing machines using Chain-of-Thought (CoT). This lets us convert the constructions to equivalent softmax transformers without the unrealistic parameter magnitudes or activation precision that prior approaches would require. Using the same technique, we analyze a recently proposed summarized CoT paradigm and show that it simulates Turing machines more efficiently, with model size scaling logarithmically in a space bound rather than a time bound. We empirically test predictions made by our results on a Sudoku reasoning task and find better alignment with learnability than for prior high-precision results. Our code is available at https://github.com/moritzbroe/transformer-expressivity.

While the approximation properties of single-layer Transformer architectures have been studied in recent works, a rigorous theoretical understanding of the multi-layer setting remains limited. In this work, we establish that multi-layer Transformers possess fundamentally different approximation capabilities from single-layer ones: for certain retrieval tasks, any single-layer Transformer requires least $Ω(\varepsilon^{-k})$ parameters to achieve precision $\varepsilon$, where $k$ grows linearly with sequence length $T$, whereas a two-layer Transformer with a single head per layer achieves the same approximation precision with at most $O (\varepsilon^{-1})$ parameters. To understand this separation, we identify two structural mechanisms underlying multi-layer approximation. Specifically, softmax attention can only efficiently retrieve the token attaining the maximum attention score, incurring exponential-in-length parameter cost for $k$-th largest retrieval with $k \geq 2$. Moreover, the parameter cost of decoding coupled information scales with the size of the retrieved token set. Motivated by these findings, we propose InfoFlow, a framework for multi-layer Transformers. The framework tracks an information set of accessible input positions at each token and layer, assigning an explicit approximation rate to each mode of information propagation. This abstraction recovers known approximation bounds, remains consistent with experimental observations on trained networks, and yields concrete predictions in settings where direct theoretical analysis is currently intractable. Our results provide a principled framework for reasoning about the approximation efficiency of multi-layer Transformers.

Autoregressive language models execute Transformer layers sequentially, creating a latency bottleneck that is not removed by conventional tensor or pipeline parallelism. We study whether this layerwise dependency can be relaxed by treating the hidden-state trace across layers as the solution of a nonlinear residual equation and solving it with parallel Newton-style updates. While this view is principled, exact Newton corrections require expensive Jacobian-vector products and naive fixed-point iterations are unstable on trained Transformers. We introduce Structured Newton Layer Parallelism (SNLP), a training and inference framework that replaces exact layer Jacobians with cheap architecture-induced surrogate dynamics. In residual Transformers, this yields Identity Newton (IDN), where the correction reduces to a prefix-sum-like update; in mHC-style architectures, HC Newton (HCN) uses the model's residual mixing matrix. We also study SNLP-aware training, including pretraining regularization and direct SNLP-forward SFT. Experiments on Nanochat-scale Transformers show that SNLP exposes a practical speed-quality frontier: on 0.5B models, it reaches up to 2.58x wall-clock speedup, and a less aggressive configuration reaches 1.40x speedup without increasing PPL. The useful tradeoff comes from the biased finite-iteration computation induced by IDN/HCN rather than exact recovery of the sequential trace. We further show that SNLP-forward SFT can preserve downstream task accuracy, and that SNLP can serve as a drafter for self-speculative decoding while a sequential verifier preserves output correctness.

Transformers have become the dominant architecture for sequence modeling by using self-attention to enable expressive and highly parallel processing. However, the resulting quadratic time and memory costs limit efficiency in long-context settings. Recurrent models such as LSTMs provide explicit nonlinear state updates and strong state-tracking capabilities, yet their strictly sequential computation limits parallelism. We introduce the Parallel Recursive LSTM (PR-LSTM), a hierarchical recurrent architecture that replaces left-to-right recurrence with recursive nonlinear state composition over a balanced computation tree. Tokens are first mapped independently to latent states, which are then recursively merged by a learned gated composition block. This structure uses the reduction pattern underlying parallel scans as a fixed execution schedule, rather than assuming an associative recurrence. As a result, PR-LSTM retains nonlinear gated state representations while reducing recurrent parallel depth from linear to logarithmic. Empirically, PR-LSTM achieves strong sequence-length generalization on formal-language benchmarks, solving more tasks than standard RNN, LSTM, and Transformer baselines, while avoiding the quadratic scaling of attention. These results suggest that recurrent computation can be reorganized hierarchically to expose parallelism without restricting the transition dynamics to linear or associative forms.

When researchers ask whether two transformer layers are "equivalent" for compression, they often conflate distinct tests. Replacement asks whether one layer's map can substitute for another's in place; interchange asks whether two layers approximately commute when their positions are swapped. Both are output-grounded swap-KL probes, but they need not agree: on pretrained transformers the protocol gap can change which layers look safe to prune by several-fold under the same evaluator, especially when replacement distances are high. We measure both protocols across checkpoints and architectures. On a Pythia training trajectory (410M and 1.4B), the replacement-interchange gap grows from initialization to convergence. Under one matched WikiText-2 contract at 8B scale, Qwen3-8B enters a divergent regime: interchange-guided removal is several-fold safer than replacement-guided at the same layer budgets, while Llama-3.1-8B ties the two protocols for pruning cost even though interchange KL is lower, showing metric gaps need not map one-to-one to removal. Before layer removal or merging, score both swap-KLs on the target checkpoint; the diagnostic requires only unlabeled forward passes.

Layer pruning removes entire Transformer decoder blocks from large language models, but introduces a mismatch between the hidden state received by the next surviving layer and the distribution it was trained to process, leading to significant performance degradation. We propose Ghosted Layers, a training-free recovery module that addresses this issue by solving a boundary activation alignment problem. Our method derives a closed-form optimal linear operator from a small calibration set to reconstruct the activation discrepancy introduced by the pruned layers. We show that this solution corresponds to the unconstrained optimum of the alignment objective, whereas existing methods are restricted to constrained solutions over limited operator subspaces. Experiments across multiple LLM backbones and pruning strategies demonstrate that our method consistently improves accuracy and perplexity over prior training-free baselines, while preserving the efficiency gains of layer pruning. Official code repository: https://github.com/daniel-eai/ghosted_layers_official_repository/.

Do transformers, when trained on sequential reasoning traces, build internal models of the underlying task? And if so, does the structure of those internal representations mirror the structure of the domain? We train an 8-layer transformer on Sudoku solving traces and perform a mechanistic analysis of its internal computation. We establish two results. First, the model builds a substructure world model: it does not represent the board state cell by cell, as a human analyst would expect, but organizes information around the rows, columns, and boxes that Sudoku's constraints act on. Second, we identify a naked-single circuit: a small set of dedicated neurons in the final MLP layer, each individually detecting when exactly one digit remains possible for a specific cell, and reliably promoting that digit. These findings show that the geometry of an emergent world model is shaped by the constraint algebra of the domain, not its surface presentation, and that the resulting decision circuit is sparse, monosemantic, and fully interpretable. More broadly, they demonstrate that mechanistic interpretability tools can recover an end-to-end algorithmic account of how a transformer solves a combinatorial reasoning task.

This note shows that no self-attention layer post-processed by a rational function can sign-represent the parity function unless the product of the number of heads and the degree of the post-processing function grows linearly with the input length. Combining this lower bound with rational approximation of ReLU networks yields a margin-dependent extension for self-attention layers post-processed by ReLU networks.