Transformer Interpretability

Follow this topic Ask about this topic

Momentum

28 papers in the last four weeks, up 180% on the four weeks before. 0.3% of all new papers.

Jul 13Week of Sep 28

Latest papers 224

Sep 7, 2026cs.AI

Unraveling the Real Working Mechanism and Inherent Flaws of GAE: A Method for Interpreting Transformer Processes from an Economic Perspective

We observe a phenomenon that current algorithmic research in the field of explainable artificial intelligence primarily pursues better performance on several proxy metrics. On the one hand, these proxy metrics themselves are more or less flawed and cannot properly measure the quality of methods. On the other hand, metric-oriented research approaches often lead to the neglect of the rationality and interpretability of the methods themselves. Explainable artificial intelligence is abbreviated as XAI. The metric-driven research paradigm has resulted in a lack of interpretability of the relevant XAI methods themselves. Accordingly, there is a need for interpretability research on XAI methods, which can be playfully referred to as XXAI. This paper is one of our works on XXAI. This paper takes Generic Attention-model Explainability (GAE), a widely influential model interpretation method , or rather, XAI method that represents an important technical route, as the research object, and explores the real working mechanism and flaws of this method as well as the technical route it represents. Based on the conclusions of this study, it may be necessary to re-examine or verify GAE-related methods and their domain applications. We argue that GAE is an interpretation method that focuses on the attention process. After pointing out the working mechanism and flaws of GAE, we propose Cumulative Asset Holdings (CAH), a more reasonable Transformer interpretation method integrating both process-based and feature-based ideas from an economic zero-sum games perspective. In addition, it is worth noting that our method is applicable to models with special tokens, where existing methods may suffer from limitations. The model simplification research method and the analysis of additive operations adopted in this study may provide inspiration for other research works in XAI.

Yongjin Cui, Xiaohui Fan

Zhejiang University

Transformer Interpretability Explainable Artificial Intelligence Explainability Evaluation

Sep 7, 2026cs.AI

Encoded Early, Used Late: Where Transformers Begin to Act on an Inferred Partner's Expertise

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

Mika Okamoto, Gabriele Sarti

Georgia Institute of Technology · Northeastern University

Transformer Interpretability Language Model Probing

Sep 1, 2026cs.CL

Sparse Readout Prism: Explaining Logit-Lens Scores in Features Instead of Tokens

A language model's prediction of its next token develops across layers, and lens methods track this process by decoding intermediate hidden states into tokens. But a lens reading reflects both the hidden state and the readout (the unembedding matrix) used to decode it. Many lenses are fit on a corpus, and we show that two lenses differing only in their fitting corpus can report different tokens for the same hidden states. We call this dependence corpus conditionality. To examine readout structure independently of the fitting corpus, we introduce Sparse Readout Prism (SRP), which decomposes the readout using only its weights and expresses any token logit or logit difference as a sum of contributions from sparse readout features. This reveals readout features as a new unit of analysis for lens readings, exposing structure that token identities can obscure and enabling comparisons across tokens, contexts, layers, and lenses. Replacing the original readout with SRP's sparse approximation reconstructs 8.9-17.3 percentage points more of the tested logit differences than the strongest of six baselines built on geometric relations among readout rows. Ablating features shifts logit differences in proportion to their SRP contributions. Although token readings vary with the fitting corpus, the dominant readout feature remains stable. Because SRP uses no corpus in its construction, it provides a control independent of the fitting corpus for lens analyses.

Matteo He, William F. Shen, Xinchi Qiu +1

Department of Computer Science and Technology, University of Cambridge

Transformer Interpretability LLM Interpretability

Sep 1, 2026cs.AI

Looped Transformers under the Jacobian Lens: Does the Global Workspace Survive Recurrence?

Recent work identifies a mid-depth band of verbalisable, causally potent representations in a standard feedforward transformer --- a functional analogue of a global workspace. Whether the same workspace functionality emerges when depth is implemented through recurrence rather than a stack of distinct layers remains unknown. Looped and depth-recurrent transformers provide a direct test of this question because they reuse the same weights across depth. We extend the Jacobian lens to iterated architectures using a virtual-unrolling adapter. We apply the full workspace suite --- lens fitting, readout, and eleven causal experiment families --- to Ouro-2.6B (48 layers looped 4 times, deeply supervised) and Huginn-0125 (a 4-layer core recurred 16 times, trained for latent reasoning), using Qwen3.6-27B (64 untied layers) as the standard baseline. We find that a workspace forms in the iterated part of each architecture, but that recurrence changes how it can be accessed. Ouro reconstructs workspace content in every loop, and linear transport cannot carry that content across loop boundaries; writes and ablations must therefore span every remaining loop. Huginn carries content forward across all sixteen recurrences, while reads, writes, and ablations act only within a sliding window of roughly two recurrences. Whether newly injected content can be verbalised tracks explicit per-iteration supervision; whether existing content can be steered does not.

Wenlong Wang, Fergal Reid

Fin AI Research

Transformer Interpretability Cognitive Modeling Recurrent Transformers

Sep 1, 2026cs.CV

HiLRP: Toward One Trustworthy Explanation for Vision Transformer: Conservation-Valid Attribution via Attention Primitives

Vision Transformer (ViT) design has become increasingly diverse, with backbones combining convolutional stems, windowed, linear, or multi-axis attention, patch merging, and spatial reduction in various configurations. This diversity poses challenges for existing attribution methods, whose assumptions often do not hold across ViT variants: Grad-CAM requires a terminal spatial feature map, attention rollout assumes global softmax attention, and layer-wise relevance propagation (LRP) requires module-specific rules. To the best of our knowledge, no existing method provides a unified attribution framework across this architectural space. We show that this architectural diversity can be captured by a simpler underlying structure. The attention and resolution-reduction operators in current ViTs can be decomposed into four operation types: linear maps, bilinear mixing, normalization or gating, and reindexing. Each operation admits a relevance rule that satisfies conservation. Based on these rules, HiLRP supports new backbones by construction rather than by architecture-specific derivation, and its attribution maps decompose the prediction rather than relying on heuristic assumptions. We prove conservation and conditional equivariance and verify both to machine precision. Across 14 attribution methods and 10 architectures, we find that no prior method remains reliable across ViT families, while Faithfulness Correlation becomes uninformative for backbones robust to spatial masking. HiLRP alone preserves conservation across windowed, spatial-reduction, multi-axis, and linear-attention models, where naive extensions can produce zero or inflated relevance. It also localizes attribution failures in class activation mapping, achieving 0.97 Pointing compared with 0.55 for competing methods on EfficientViT.

Sathiyamohan Nishankar, Pubudu Sanjeewani, Asanka Perera +1

Faculty of Engineering, University of Peradeniya, Sri Lanka · School of Computing Technologies, RMIT University, Melbourne, Australia · School of Engineering & Digital Technologies, University of Southern Queensland, Brisbane, Australia +1

Transformer Interpretability Vision Transformer Layer-Wise Relevance Propagation

Aug 31, 2026cs.CL

Late Transformer Layers Recode Syntax Canonically: Evidence from Greek Scrambling and Cross-Layer Generalisation

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.

Christos Nikolaos Zacharopoulos, Revekka Kyriakoglou, Chara Tsoukala +1

Independent Researcher · Université Paris 8 Vincennes–Saint-Denis Paris, France · Athena Research Center, Athens, Greece +1

Transformer Interpretability Language Modeling Representation Geometry in Language Models

Aug 31, 2026cs.LG

Tracing distinguishability through transformer processing with stochastic LayerNorm

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

Kieran Murphy

New Jersey Institute of Technology

Transformer Interpretability Layer Normalization

Aug 25, 2026cs.AI

RACE: Scalable Statistical Estimation of Functional Consistency in LLM Neurons

Discovering stable neuron behavior across entire domains remains a challenge in mechanistic interpretability. Existing methods often rely on instance-level point estimates or computationally expensive procedures, which either obscure population-level variability or limit scalable domain-wide analysis. We present RACE (Residual Alignment for Consistency Estimation), a forward-pass statistical framework that evaluates the domain-wide functional consistency of Transformer neurons. Compared with gradient-based point estimates, RACE produces neuron rankings that yield more domain-specific effects under perturbation. Token-distribution shifts support the connection between the selected neurons and the target domain, while scoring requires roughly one-hundredth of the computational overhead of the gradient-based methods. Code is available at https://github.com/Nexround/RACE.

Runyu Wang, Bo Liu, Xiaxin Zhang +6

School of Transportation and Civil Engineering, Nantong University · Chongqing University of Post and Telecommunications · China Southern Power Grid Company Limited +1

Transformer Interpretability LLM Interpretability Mechanistic Interpretability

Aug 13, 2026cs.LG

Perturbation-based Regional Interpretability through Subtraction Mapping (PRISM): naming-error dissociations in language models and post-stroke aphasia

Mechanistic interpretability of large language models lacks spatially resolved, falsifiable tools for testing whether internal components are specialized for distinct cognitive operations. We adapt subtraction analysis, the standard framework of human neuroimaging, from biological brains to perturbed transformers, and apply the same logic to both substrates in parallel. Building on the Brain-LLM Unified Model (BLUM), which showed that layer-perturbed LLaVA-1.6-Vicuna-13B error profiles match the lesion patterns of aphasic patients, we develop PRISM (Perturbation-based Regional Interpretability through Subtraction Mapping). PRISM maps the seven clinical Philadelphia Naming Test categories, subtracts error classes pairwise, and treats each perturbation seed as a subject in a group analysis with threshold-free cluster enhancement along the layer axis. We run a structurally matched analysis on 213 chronic post-stroke aphasia patients using correlation-difference lesion-symptom mapping, and replicate both sides on held-out splits. The designs match in subject dimension (seeds, patients), spatial dimension (layers, atlas-parcellated cortex) and thresholding, but the contrast operator differs: a within-subject error-proportion difference for the LLM, a between-subject correlation difference for the cortex. Both substrates recover a robust phonemic-favoring dissociation, a deep layer cluster and a frontal-perisylvian cortical cluster, both replicating; the semantic-favoring direction is a consistently signed but non-significant trend on both. PRISM thus gives a falsifiable, spatially resolved test of functional-specialization claims in transformer language models. A confirmatory ROI-level intervention (PRISM Stage 3) licensing the strongest causal-mechanism claim is left to subsequent work.

Xiang Guan, Roger D. Newman-Norlund, Yong Yang +8

Department of Computer Science and Engineering, University of South Carolina · Department of Communication Sciences and Disorders, University of South Carolina · ALLT.AI, LLC, Columbia, SC +3

Transformer Interpretability LLM Interpretability Neural Network Interpretability

Aug 12, 2026cs.CV

Class Activation Mapping in Explainable Computer Vision: A Method-Centered Review of CNN, Transformer, and Foundation-Model-Era Visual Explanations

Class activation mapping (CAM) is one of the most widely used visual explanation families in explainable artificial intelligence. Its purpose is intuitive: it converts internal model evidence into a heatmap that highlights the image regions, convolutional channels, tokens, or patches that support a target class or concept. Since the first CAM formulation in 2016, the field has moved far beyond global-average-pooled CNN classifiers. CAM-style methods now include gradient-based post-hoc explanations, gradient-free score and ablation methods, high-resolution upscaling, weakly supervised localization and segmentation, transformer token attribution, causal and debiasing methods, and foundation-model-era approaches that use CLIP, DINO, SAM, or feature-distribution comparisons. This review synthesizes a strict corpus of 57 method-centered papers published from 2016 onward. The paper develops a taxonomy that separates methods by attribution mechanism, architectural dependence, and evaluation objective. It then reviews gradient-based CAMs, recent and hybrid CAM-style methods, and model-based or architecture-aware methods. Across the corpus, the main trend is clear: the field is shifting from explaining one class score in one low-resolution CNN layer toward comparative, multi-layer, probabilistic, token-aware, and foundation-model-aware explanations. At the same time, evaluation remains fragmented. Faithfulness, localization, robustness, computational cost, and human trust are often measured with different protocols. The review therefore emphasizes not only what each method contributes, but also which gap it leaves open and which later methods attempt to close that gap.

AmirHossein Eshghi, Hamid Saadatfar, Seyyed Ali Hoseini +2

Department of Computer Engineering, University of Birjand, Iran · Department of Applied Mathematics and Computer Science Visual Computing, Technical University of Denmark, Denmark

Transformer Interpretability Feature Attribution Explainable Image Classification

Aug 12, 2026cs.LG

Geometric and Behavioral Stratification in Transformer Residual Streams

Trained transformer models develop privileged bases: coordinate axes whose statistics differ from the rest of the residual stream. But what kind of direction does such a basis select? We investigate the prediction direction, the unembedding direction of the token a model currently predicts, and find that it functions as a content-defined privileged anchor. Measured with respect to this anchor, residual-stream variation is geometrically and behaviorally stratified by proximity to the prediction. The stratification holds in all eighteen models tested (dense and mixture-of-experts, 7B-120B, base and instruction-tuned). A narrow, scale-invariant prediction interface concentrates readout-relevant structure, while the vast prediction-distal complement expands with model scale. Because the prediction direction sits nearly orthogonal to the principal variance axes, variance-based analyses recover this organization only partly, and the shortfall grows with prompt heterogeneity. Anchoring reveals a steep geometric gradient: prediction-proximal regions are highly structured and cluster related prompts, while the complement is flatter and anti-discriminates among prompt groups. The interface is a narrow slice but functionally decisive. Disrupting the variance directions closest to the prediction causes immediate divergence and frequent task-frame shifts; disrupting the next level down delays divergence and preserves framing. The complement is weakly readout-aligned per direction yet causally and temporally load-bearing, and behavior is driven by direction rather than magnitude. These results establish the prediction direction as a privileged anchor distinct from previously described coordinate axes, and give a geometric account of how high-dimensional computation coexists with linear readout.

Nelson Guda

Transformer Interpretability Transformer Representation Geometry in Language Models

Aug 10, 2026cs.CL

Off-Axis, On Purpose: Where a Transformer Computes Concepts and Why it Does So

A transformer's answer lives on one axis: the direction its unembedding reads. Its intermediate states largely do not, and that off-axis position is usually treated as an obstacle to interpretation. We show it is functional. A 12-layer model computes in two phases. Through the first, every sublayer writes into a subspace held near-orthogonal to the read-out, attention 75 to 96 degrees off it at every depth. Moving attention's values onto the read-out is 64 to 84 times more damaging than a matched random rotation, and the damage is entirely in cross-token mixing: the subspace insulates composition from the vocabulary. Beneath it the frame itself turns rigidly with depth. In the second phase the answer arrives on-axis, late, and by addition rather than by turning accumulated content onto the read-out. Pressing every layer onto the read-out instead, as training for early exit does, matches the baseline on perplexity, LAMBADA and BLiMP while cutting the concept-phase workspace from about twenty-five effective dimensions to fourteen, a change none of those benchmarks register. The geometry can also be imposed, though not by asking for it. Prescribing it through the loss is a lottery: six of eight seeds collapse, because a model told to null its read-out projection obeys most cheaply by discarding dimensions. Inserting one fixed rotation at the phase boundary lands it instead, at baseline quality. A sparse rotation the surrounding weights can absorb converges on all nine seeds, against five of nine for ordinary training. Which rotation is immaterial: twenty-five runs across thirteen distinct ones reach the same quality, and two baselines from different seeds hold their concepts in near-orthogonal frames while agreeing on their read-outs. That freedom is usable: a basis drawn at random and prescribed before training is adopted across the concept phase, with quality unchanged.

Mark Oskin

School of Computer Science and Engineering University of Washington

Transformer Interpretability Transformer Attention Representation Geometry in Language Models

Aug 8, 2026cs.LG

Spectral Outliers Reveal Dominant Learned Structure in Transformer Attention

We apply Marchenko-Pastur (MP) random matrix theory to pre-trained attention weights in order to separate each projection matrix into a random-like bulk and a set of spectral outliers. We validate this decomposition causally: zeroing the MP-identified outliers (signal) in Mistral-7B drives HellaSwag, MMLU, and PIQA close to random-chance performance, whereas zeroing a count-matched subset of bulk singular values causes smaller but non-negligible degradation. Across 11 pre-trained transformers we identify five recurring patterns: spectral outliers encode a dominant component of the learned structure; Q projections carry the most outliers; V projections under grouped-query attention lack a clean signal/noise separation; entry-level outliers form structured row-bands in Q and column-bands in O; and specific residual-stream dimensions persist as band outliers across layers in K and O. We close by outlining how these observations could inform parameter-efficient fine-tuning and structured pruning.

Kasun Dewage, Marianna Pensky, Suranadi De Silva +1

University of Central Florida Orlando, Florida, USA · Virtusa

Transformer Interpretability Transformer Attention

Aug 7, 2026cs.CL

Why Knowing Both Hops Is Not Enough: Understanding Two-Hop Generalization in Language Models

Large language models (LLMs) can solve complex multi-hop problems yet exhibit puzzling failures on simple two-hop queries: although a model may correctly store each individual hop, it often fails to combine them. To understand the internal mechanisms of this phenomenon, we train transformers from scratch in a controlled symbolic environment. Our experiments reveal a pattern in two-hop generalization: models generalize reliably when the second hop follows the training distribution, but always fail when it deviates. Through mechanistic analysis, we provide a complete explanation for these distinct generalization behaviors: in settings where models generalize successfully, performance is driven by the emergence of consistent intermediate representations for the same entities across contexts, whereas failures on settings where the second hop is out-of-distribution arise from a mismatch across layers: lower layers correctly construct these intermediate representations, but upper layers, while trained on corresponding atomic facts, primarily learn to map them to outputs rather than to reason over them. Driven by this insight, we propose a recurrent-style training strategy, which enables transformers to reuse their reasoning circuitry across input forms and substantially improves generalization on out-of-distribution two-hop queries.

Zili Zhang, Yilin Wang, Heng Wang +2

1Xi’an Jiaotong University · University of Illinois Urbana-Champaign

Transformer Interpretability OOD Generalization Multi-Hop Reasoning

Aug 7, 2026cs.LG

Transformer Geometry Observatory TGO-IV: Developmental Topology Observatory

Transformers have had a profound impact on the world of language processing and computer vision. As efforts to answer the million-dollar question of ``How does a Transformer learn?" have been increasing, existing interpretability studies primarily analyze representations at isolated layers or the network as a whole, while the developmental evolution of individual representations and its manifolds across transformer layers remains underexplored. With this work, we aim at providing a comprehensive analysis of the evolution of representations as the representation point cloud transforms across the layers; thereby attempting to isolate layers or establish a trend which comes closer to justifying how and when raw input representations evolve into task-relevant feature representations. Thus, Transformer Geometry Observatory-TGO-IV introduces a topological framework for analysing the evolution of Transformer representations through the lens of Persistent Homology. Rather than studying local geometric properties alone, TGO-IV constructs Vietoris--Rips simplicial complexes from token-level representation point clouds and investigates the evolution of their persistent topological signatures across Transformer layers. The proposed framework comprises complementary topological observatories including Persistence Diagrams, Barcode Diagrams, Betti Curves, Persistence Landscapes, Bottleneck Distance, and Wasserstein Distance, enabling a comprehensive analysis of how the global topology of representation point clouds develops throughout the forward pass.

Kaustubh Kapil, Kishor P. Upla

Neuromorphic Intelligence Research Collective · Department of Electronics Engineering · Sardar Vallabhai National Institute of Technology (SVNIT), Surat, India

Transformer Interpretability Neural Representation Geometry

Aug 6, 2026cs.LG

Beyond Attention: Signed Integrated Gradients Attribution in a BiomeGPT-Style Microbiome Transformer

In a feature-tokenized transformer (arXiv:2106.11959) such as BiomeGPT (doi:10.64898/2026.01.05.697599), each input token is built by fusing a fixed identity with a sample-specific measurement: a fixed species and a variable abundance, T = S + A. To interpret downstream classification in such models, prior work inspects the attention weights of the special [CLS] token (arXiv:2106.11959, arXiv:1810.04805, BiomeGPT) to rank sample tokens by importance. These weights have two critical limitations: they are nonnegative, so they cannot separate disease-supporting from health-supporting evidence (arXiv:2201.12114), and they act after token fusion, obscuring how the input sources S and A each affect the output. To address this we use Integrated Gradients (arXiv:1703.01365), a signed, fusion-aware attribution method, and propose a source-derived baseline T' = S + A_0 for feature-tokenized models such as BiomeGPT, which preserves species identity as a fixed biological coordinate while isolating the effect of abundance variation. Applied to a disease-versus-health decision margin, it yields polarity that explicitly separates pathogenic from protective microbial signals. We show that this gradient-based approach uncovers species-abundance directional relationships and sensitivity diagnostics entirely obscured by unsigned [CLS] attention weights. We further recommend second-order Integrated Hessians (arXiv:2002.04138) to expose microbiome community interaction rules: how a perturbation in one member alters the model's sensitivity to another, and which other species drive ambiguous cases toward disease or health at a given abundance level. This provides a principled approach to explainability in BiomeGPT that generalizes to other smooth and differentiable feature-tokenized transformers. Code is available at https://github.com/nohren/token-source-attribution

Oren Nelson

Department of Computer Science University of California, San Diego San Diego, CA 92122

Gradient-Based Attribution Integrated Gradients Transformer Interpretability

Aug 5, 2026cs.CV

IRIS: A Visual Cortex-Inspired Framework for Analyzing Orientation Selectivity in Vision Transformers

Vision transformers (ViTs) have become the de facto standard for image encoding across many perception tasks. Despite their empirical success, it remains mechanistically unclear how they encode low-level features, given their lack of inductive biases: ViTs process information globally rather than relying on local structure. Biological visual systems, in contrast, build low-level features, such as orientation selectivity in the primary visual cortex, by combining information from small, localized regions of the visual field. These features are general-purpose representations, shared and required across multiple specialized neural pathways, unlike higher-level, task-specific semantic features. This raises the question if such biologically-grounded features arise in ViTs. In this work, we systematically study how orientation selectivity emerges in ViTs by introducing a suite of neuroscience-inspired metrics: representational similarity score (RSS), orientation recruitment score (ORS), and orientation tuning bandwidth to quantify how orientation is encoded in representational geometry and as a function of model depth. Through extensive analysis, we find that: (1) the training paradigm is the strongest determinant of orientation selectivity, with models sharing an objective, peaking at comparable relative depths regardless of scale (2) many units are orientation-selective early in training, with early-to-middle layers recruiting more such units over time, while deeper layers lose selectivity and broaden their tuning toward semantic encoding and (3) our metrics offer a mechanistic heuristic for how many layers to unfreeze for best downstream generalization. Our framework presents a way to track biologically-grounded features during ViT training, probes how desired properties are encoded in transformer representations, and builds a systematic understanding of how ViTs generalize across tasks.

Vaishnavi B Mohan, Vijayakrishna Naganoor, Yashas Annadani +1

University of Washington, USA · Microsoft, USA · TU Munich, Germany +1

Transformer Interpretability Visual Representation Learning Vision Transformer Visual Neuroscience

Aug 5, 2026cs.CV

Grad-CAM for Vision Transformers: A Systematic Taxonomy and Audit of Methodological Ambiguity in Explainable AI

Gradient-weighted Class Activation Mapping (Grad-CAM) is widely used to visualize model decisions, but it was originally formulated for convolutional neural networks, where spatial feature maps and channel dimensions have clear architectural meanings. Vision Transformers (ViTs) do not provide the same structure, instead representing images through tokens, attention, residual streams, and multimodal interactions. This paper presents a systematic taxonomy and literature audit of how Grad-CAM and related methods are adapted, justified, and reported for ViT-based architectures. From an initial search of more than 550 papers, we identify 175 papers that apply Grad-CAM or Grad-CAM-adjacent methods to ViTs. We find that most papers do not provide a full mathematical or implementation-level account of how Grad-CAM is adapted to transformer representations. To characterize this gap, we introduce a descriptive taxonomy of ViT Grad-CAM adaptations that makes explicit the feature locations, gradient targets, spatial reconstruction steps, and aggregation choices that are often left implicit. This taxonomy is not intended to prescribe a single correct adaptation, but to clarify the range of methodological choices being made. The study shows that Grad-CAM on ViTs is often treated as a trivial extension of CNN-based Grad-CAM, despite requiring nontrivial choices that affect rigor, reproducibility, and interpretation.

Casey Wall, Longwei Wang, Rodrigue Rizk +1

USD Artificial Intelligence Research, Department of Computer Science, University of South Dakota

Gradient-Based Attribution Transformer Interpretability Vision Transformer

Aug 4, 2026cs.LG

Sparse Weight Decomposition for Efficient Circuit Extraction

Dense pretrained transformers do not naturally expose interpretable units for circuit extraction. Existing approaches obtain such units by learning auxiliary sparse representations or training sparse models, incurring substantial additional computation while potentially introducing a fidelity gap between the representation being analyzed and the original pretrained model. We propose Sparse Weight Decomposition (SWD), which reparameterizes pretrained linear projections by factorizing each weight matrix into two sparse factors whose shared intermediate coordinates serve as individually addressable circuit units. Without training a separate replacement network, this parametric representation supports the same scoring, selection, and ablation circuit extraction workflow used for methods that learn sparse features. Across single-matrix replacements, SWD matches the held-out fidelity achieved by Transcoder and other strong baselines while using less than 1% of the data that those baselines use to train their replacements. For matched replacement fidelity, SWD reaches the same circuit sufficiency and necessity targets with fewer active read/write edges and selected units across tasks on GPT-2, Qwen2.5, and Qwen3.5-27B. We further show that SWD remains effective for full-model replacement of all attention and MLP weight matrices after fine-tuning the nonzero factor values. Finally, SWD also features a zero-data variant, allowing broader use of mechanistic interpretability analysis (e.g., per-step analysis).

Chuanhao Yan, Xuhan Huang, Yawen Duan +4

IQuest Research · Safe AI Forum · University of Oxford +2

Transformer Interpretability Mechanistic Interpretability Circuits in Language Models Matrix Factorization

Aug 4, 2026cs.AI

Cross-Layer Interaction under Weight-Space Ablation: A Closed-Form Attention Jacobian Bound and a Test on a Real Pretrained Model

A companion paper studies when activation patching and weight-space ablation agree, inside an idealized model where a conditional computation is carried additively through a residual stream. For the one composition in that model where two carriers are architecturally dependent, an attention head and its own layer's normalization-MLP composition, it derives an exact first-order interaction formula, zero when only the MLP is ablated and second-order bounded when the head is also ablated. That result is confined to a single residual block and checked only on small transformers on a synthetic task. This paper extends the result past both limits. First, the interaction from ablating carriers spanning several layers decomposes exactly into same-block terms, one per touched layer, plus a cross-layer remainder on which the decomposition makes no claim of smallness. Second, we isolate that remainder exactly, for two layers, as a double integral of a mixed second derivative, and name the missing ingredient needed to bound it: a Jacobian bound for the attention sub-block. We derive this bound in closed form and verify it, without a single violation, against Qwen2.5-1.5B-Instruct's real weights, though we do not yet chain it across layers. We also give, in closed form, the curvature constant the companion paper's bound leaves unexhibited. Third, on that same model, we search for and find an emergent circuit for indirect object identification, never designed into it, using the original activation-patching method for this task, and test collapse, dissociation, and interaction on it. The result is mixed: a shared carrier emerges across all five tested instances, collapse and dissociation hold on most but not all, and a nonzero interaction is measurable on three of five, at layer pairs outside the same-block case the companion theorem covers.

Abdallah Khemais

ISITCOM, University of Sousse

Transformer Interpretability Weight-Space Analysis Activation Patching

Aug 4, 2026cs.LG

A Theory of Conditional Collapse under Low-Rank Weight-Space Ablations: I. The Single-Block Theory and Synthetic Validation

Activation patching and weight-space ablation both claim a component is causally responsible for a behavior, yet they act on different objects: one forward pass versus the parameters behind every forward pass. We ask when they agree. We study an idealized model where a conditional computation is carried additively through a residual stream,

F(x)=F_0(x)+\sum_iα_i(x)v_i

, read out by a linear functional, and prove three exact results. First, deleting a subset of carriers collapses a matched input pair onto the same unconditional output \emph{if and only if} the removal is symmetric on the pair and leaves no outside contrast; the error is deterministic, and we give its exact form even when the two conditions hold only approximately. Second, patching a carrier moves the readout by its donor-receiver \emph{contrast}, while ablating it moves the readout by its \emph{absolute level}; neither bounds the other, and we construct pairs where every single-carrier patch flips the decision while no single-carrier ablation does. Third, for an attention head composed with its own layer's normalization and MLP, we derive an exact first-order interaction formula with a provably second-order remainder, vanishing identically when only the MLP is ablated but not, in general, when a head is. Small transformers trained on a synthetic conditional task illustrate all three predictions: across thirty-nine ablation configurations the measured interaction is strongly rank-correlated with the idealized model's predictive accuracy (Spearman

-0.83

), and a second task and architecture reproduces the same pattern, including a further polarity reversal. The single-block interaction result extends past one residual block, and the synthetic validation is tested against a real pretrained model, in a companion paper that takes this theory further along both axes.

Abdallah Khemais

ISITCOM, University of Sousse

Transformer Interpretability Weight-Space Analysis Causal Interventions in Language Models Activation Patching

Aug 4, 2026cs.CL

Probing Character-level Transformers for the Spanish L-shaped Morphome

When a transformer learns an irregular morphological pattern, what has it learned? Our test case is the Spanish \emph{L-shaped morphome}, a complex irregular pattern in which the verb's stem alternates in exactly the first-person singular indicative and all subjunctive forms, and whose membership no phonological, semantic, or syntactic feature predicts. Prior studies have shown that character-level transformers can reproduce this pattern, but that evidence describes what models produce, not what they represent. Probing five architectures, twelve trained models each, under lemma-disjoint cross-validation with controls and surface baselines, we show that the models encode the L-shaped class itself, not just its visible alternations. It is decodable above every surface baseline, survives instances in which every form shows the same stem, and probes trained on alternating instances still classify non-alternating ones. The encoding is localized where the stem choice is made, at the stem-final consonant position of the middle decoder, before the alternant is read. And it is item-specific: which verbs a model learned matters far more than which architecture it is. The models store the morphome as an item-specific lexical abstraction, sufficient to reproduce the pattern but not to generalize it as humans do.

Akhilesh Kakolu Ramarao, Kevin Tang, Wiebke Petersen +1

Department of English Language and Linguistics, Heinrich Heine University Düsseldorf · Department of Linguistics, College of Liberal Arts and Sciences, University of Florida · Institute of Linguistics and Information Science, Heinrich Heine University Düsseldorf +1

Transformer Interpretability Transformer Representation Probing

Aug 3, 2026cs.CV

Does Explainability Transfer? A Controlled Benchmark of Attribution Methods on Vision Transformers and CNNs

Most evidence on the effectiveness of explainable artificial intelligence (XAI) attribution methods has been established on convolutional neural networks (CNNs), with limited investigation into whether these conclusions generalize to the diverse Vision Transformer (ViT) architectures that now dominate computer vision. This paper presents a controlled benchmark that evaluates attribution quality across five dimensions: faithfulness, localization, robustness, complexity, and computational cost. A standardized framework assesses 13 attribution methods from four algorithmic families on eight representative backbones spanning CNNs, isotropic ViTs, hierarchical transformers, hybrid architectures, and linear-attention transformers. The results show that attribution performance is strongly architecture-dependent and that rankings established on CNNs do not reliably transfer to transformer-based models. CAM-based methods achieve the highest scores under the conventional bounding-box localization metric on CNNs and most ViTs but perform poorly on linear-attention architectures. Pixel-level dense-mask evaluation further reveals that these gains largely reflect metric saturation rather than accurate localization. CAM-based methods also exhibit limited robustness on global-attention transformers, whereas attention rollout provides consistently stable explanations with poor localization. Furthermore, faithfulness correlation offers limited discrimination between attribution methods, highlighting the limitations of single-metric evaluation. These findings challenge prevailing conclusions on attribution performance and demonstrate the need for architecture-aware, multi-dimensional evaluation. The open-source code for the evaluation framework and benchmark results is available at https://github.com/Nishan-Charlie/VIT_XAI_Bench.

Sathiyamohan Nishankar, Nethmi Pathirana, Pubudu Sanjeewani +2

aFaculty of Computing, Sabaragamuwa University of Sri Lanka, Sri Lanka · bFaculty of Engineering, University of Moratuwa, Katubedda, Sri Lanka · cSchool of Computing Technologies, RMIT University, Melbourne, Australia +3

Transformer Interpretability Feature Attribution Vision Transformer Explainability Evaluation

Aug 3, 2026cs.LG

Feed-Forward Steering in Transformer Residual Dynamics

Attention-only dynamical theories model Transformer residual directions as particles aggregating on a sphere. We extend this framework by incorporating the feed-forward network (FFN) term as a local steering field acting on each token state. The resulting theory predicts that the tangential component of the FFN field is necessary for motion in residual-direction space, that critical residual directions correspond to nonlinear projective equilibria, and that a commutator defect determines when a finite attention--FFN block can be accurately approximated by a parallel, additive flow. Across GPT-2, Pythia, Mistral, and Llama models, the extended theory improves one-step angular prediction relative to an attention-only baseline, with the contribution of the FFN increasing from GPT-2 to Llama-3-8B. Intervention experiments show that retaining only the tangential FFN component preserves most model quality, whereas retaining only the radial component causes performance to collapse. The tangential component also preserves output diversity under aggregation pressure. As a practical application, layers with small commutator defects can be approximately parallelized with only a modest increase in loss, whereas layers with large defects degrade rapidly. These findings support the interpretation of FFN layers as directional steering fields that shape Transformer residual geometry and govern the feasibility of block-level interventions.

Timur Mudarisov, Mikhail Burtsev, Radu State

University of Luxembourg · 2London Institute of Mathematical Sciences

Transformer Interpretability Transformer FFNs Transformer Transformer Attention

Aug 3, 2026cs.CV

Transformer Geometry Observatory TGO-III: Semantic Geometry Observatory

With the widespread adoption of Vision Transformers in modern AI, the need to analyze their inherent representational behavior has become increasingly important. While most existing studies emphasize token geometries and training dynamics, the evolution of representational covariance structures and class-level geometric organization remains comparatively underexplored. In this work, we investigate semantic geometry and class separability as representations evolve across the layers of ViT-Small/16 through TGO-III: Semantic Geometry Observatory. It is a framework designed to analyze the emergence of semantic organization, feature evolution, and class-wise representation geometry throughout training. The framework employs multiple complementary observatories, including Linear Probe Accuracy, Fisher Ratio, Class Centroid Distances, Local Intrinsic Dimension, and Local PCA Rank, to quantify the progressive evolution of discriminative representations. Our analysis reveals that class representations become progressively more linearly separable, Fisher discriminability increases, class centroids move farther apart, and local representation manifolds exhibit structured class-dependent geometric complexity. These observations provide empirical evidence supporting the Semantic Expansion Hypothesis, suggesting that the manifold expansion observed in previous observatories is accompanied by the progressive organization of representations into increasingly discriminative semantic structures. Collectively, TGO-III extends the Transformer Geometry Observatory framework by establishing a direct connection between manifold geometry, covariance evolution, and semantic organization during Transformer training.

Kaustubh Kapil, Kishor P. Upla

Department of Electronics Engineering Sardar Vallabhai National Institute of Technology (SVNIT), Surat, India

Transformer Interpretability Neural Representation Geometry Vision Transformer

Aug 3, 2026cs.CL

Divergent large language model predictions from convergent representations in ambiguous word pairs

In this work we investigate how decoder-only transformers resolve lexical ambiguity through layer-by-layer analysis of three models spanning three parameter sizes (GPT-2-Small-117M, Llama-3.2-3B, Qwen2.5-32B). For both homonyms and polysemes, we find that representations become maximally distinct in middle layers, then partially reconverge in late layers, while the KL divergence between their next-token predictions reaches its maximum in the final layers. The activation patching experiment provides causal evidence that late-layer representational differences directly determine outputs despite apparent increased similarity in embedding space. Our single-layer ablation experiment indicates that models achieve equivalent disambiguation despite qualitatively different layer-wise vulnerabilities. These findings offer a mechanism for recent observations where models' internal embedding similarities show low correlation with their behavioural outputs despite strong performance. The semantic distinctions therefore remain present but become increasingly invisible to similarity measures over the embeddings, with implications for embedding-based methods such as semantic search, retrieval, and clustering that rely on late-layer cosine similarity.

K. Jack Scott, Narun Pat, Veronica Liesaputra

Department of Psychology, University of Otago, Dunedin, New Zealand · School of Computing, University of Otago, Dunedin, New Zealand

Transformer Interpretability Decoder-Only Language Models Representation Geometry in Language Models

Jul 31, 2026cs.CV

Interpretability-Guided Soft Pruning of Attention Heads in Vision Transformers

Vision foundation models, such as DINOv2, learn highly expressive representations but rely on massive, opaque architectures that demand substantial computational power and memory. To provide an interpretable-guided and efficient solution to this issue, we first propose a spectral analysis and new visualization technique for individual attention heads based on the Laplacian eigenvectors of their attention maps. Building upon recent observations regarding the block structure of Vision Transformers, we perform semantic clustering of attention heads and identify functional redundancies. Leveraging these insights, we introduce SAPER (Soft Attention PrunER), an end-to-end differentiable pruning framework based on the LapSum Soft Top-K approach. Extensive experiments on ImageNet-1K demonstrate that SAPER achieves a highly favorable accuracy-efficiency trade-off, outperforming the competitive RAPTOR baseline in FLOPs reduction while preserving strong classification performance.

Kamil Książek, Piotr Suszyński, Michał Jan Włodarczyk +2

Centre for Credible Artificial Intelligence Warsaw University of Technology Warsaw, Poland · Faculty of Mathematics and Computer Science Jagiellonian University Krakow, Poland · University of Warsaw Warsaw, Poland

Transformer Interpretability Attention Head Analysis Vision Transformer Efficient ViTs Attention Head Pruning Attention Head Specialization

Jul 30, 2026cs.AI

HealthCAT: An Interpretable Encoder-only Transformer Framework for Health Indicator Prediction and Temporal Interpretation of Wearable Sensor Data

Wearable sensors continuously capture fine-grained multivariate time-series data, providing opportunities to model behavioural patterns associated with health outcomes. However, existing deep learning methods prioritise predictive accuracy over interpretability, limiting their application in health research. In this study, we present HealthCAT, a flexible framework that integrates an Encoder-only Transformer with an Attentive Class Activation Token (AttentiveCAT) to generate class-specific, time-step-level interpretations. These interpretations can be mapped back onto behavioural cycles that are relevant to the domain (e.g., time-of-day), supporting individual-level analysis of wearable sensor data. We evaluated HealthCAT using two real-world wearable sensor datasets (306 participants in total). HealthCAT outperformed deep learning baselines by up to 17% in F1-score and 12% in accuracy on both datasets (

p<0.05

). In masking experiments, the time steps identified by HealthCAT carried significantly more predictive value than random selection across all masking conditions (

p<0.05

), indicating that the identified time steps are predictively informative. By coupling predictive performance with validated time-step-level interpretability, HealthCAT moves wearable sensor analysis beyond aggregated metrics towards temporal patterns that support health monitoring, behavioural pattern analysis, and intervention design in health research. The significance of this work is that it enables accurate prediction of health indicators from wearable sensor data while providing insights into when and how physical activity patterns occur, rather than relying solely on aggregated summary measures.

Xiaotong Yu, Joshua Y. Kim, HaeJin Lee +1

Transformer Interpretability Explainable Artificial Intelligence Time Series Classification Wearable Health Monitoring Time Series Transformers

Jul 28, 2026cs.LG

Emergent Latent-State Computation under Stochastic Volatility

Mechanistic interpretability has largely focused on language models and deterministic toy tasks. Much less is known about how sequence models internally represent latent stochastic dynamics under noisy, partially observed observations. We study this question in a controlled multivariate stochastic volatility setting, where models observe only returns while the ground-truth latent volatility state is known to the researcher. This setting provides a useful benchmark for mechanistic interpretability under partial observability: the latent state is hidden from the model but directly available for evaluation. Across architectures, losses, and output heads, we find evidence for a two-stage computation. Hidden representations encode substantial information about the next latent volatility state, and the output head maps this representation to squared return forecasts. Furthermore, in Transformers, latent-state decodability emerges at identifiable architectural stages whose location depends on the volatility period. In long-cycle regimes, this computation simplifies into an explicit latent-state filter consisting of a learned linear projection followed by

\ell^2

normalization. Output-head replacement further shows that part of the degradation under noisy MSE training arises from readout misalignment rather than representation failure. These results suggest that stochastic volatility models provide a useful benchmark for mechanistic interpretability under noisy latent dynamics and partial observability.

Xiaoyu Huang, Lulu Wang

Department of Mathematics Temple University Philadelphia, PA, USA · Department of Data Analytics Dickinson College Carlisle, PA, USA

Transformer Interpretability Mechanistic Interpretability State Representation Learning

Jul 27, 2026cs.LG

Interpretable GOHR Agents via Sparse Autoencoders

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

Shiwei Tan, Yusong Zhao, Weiyi Qin +6

1Rutgers University · 3Paul B Kantor Consultant · University of Illinois Urbana-Champaign

Transformer Interpretability Sparse Autoencoders Interpretability in RL