Transformer Interpretability

Follow this topic Ask about this topic

Transformer interpretability asks how models process information internally, from language and vision to speech, biology, and time series. Researchers use probes and representation geometry to find encoded features, sparse autoencoders to decompose activations, and circuit tracing, ablation, and activation patching to test causal roles; attribution methods explain inputs, while steering tests how internal signals can control outputs. Work also explores interpretable architectures that expose computation by design and controlled algorithmic tasks where the expected solution is known. A central challenge is that information being readable in a representation does not prove the model uses it, and findings can vary with architecture, task, and intervention, making causal validation and reliable evaluation essential.

Latest papers

340 papers34 papers in the last four weeks, down 8% on the four weeks before. 0.2% of all new papers.

Apr 25, 2026

From Edges to Depth: Probing the Spatial Hierarchy in Vision Transformers

Vision Transformers trained only on image classification routinely transfer to tasks that demand spatial understanding, yet they receive no spatial supervision during pretraining. We ask where and how robustly such structure is encoded. Probing a frozen ViT-B/16 layerwise for two complementary properties, local patch boundaries (BSDS500) and per-patch depth (NYU Depth V2), reveals a clear hierarchy: boundary structure becomes linearly decodable at layers 5-6 (AP = 0.833), while depth, which requires integrating global cues, peaks two to three layers later at layer 8 (MAE = 0.0875). Both signals collapse at the final classification layer, and random-weight controls confirm the encodings are learned rather than architectural. Causal interventions add specificity: ablating the single direction a linear depth probe reads degrades depth decoding by up to 165%, while ablating any other direction changes it by less than 1%. Targeted activation patching along that direction shows the depth signal is partially re-derived at each layer rather than passively carried in the residual stream, with mid-layer interventions persisting most strongly downstream. The result is that a classification-trained ViT develops an actively maintained spatial hierarchy that mirrors the early-to-late progression observed in the primate visual cortex.

Jainum Sanghavi

Northeastern University

Transformer Interpretability Visual Spatial Reasoning Visual Representation Learning Vision Transformer

Apr 25, 2026

When Chain-of-Thought Fails, the Solution Hides in the Hidden States

Whether intermediate reasoning is computationally useful or merely explanatory depends on whether chain-of-thought (CoT) tokens contain task-relevant information. We present a mechanistic causal analysis of CoT on GSM8K using activation patching: transferring token-level hidden states from a CoT generation to a direct-answer run for the same question, then measuring the effect on final-answer accuracy. Across models, generating after patching yields substantially higher accuracy than both direct-answer prompting and the original CoT trace, revealing that individual CoT tokens can encode sufficient information to recover the correct answer, even when the original trace is incorrect. This task-relevant information is more prevalent in correct than incorrect CoT runs and is unevenly distributed across tokens, concentrating in mid-to-late layers and appearing earlier in the reasoning trace. Moreover, patching language tokens such as verbs and entities carry task-solving information that steers generation toward correct reasoning, whereas mathematical tokens encode answer-proximal content that rarely succeeds. Patched outputs are often shorter and yet exceed the accuracy of a full CoT trace, suggesting complete reasoning chains are not always necessary. Together, these findings demonstrate that CoT encodes recoverable, token-level problem-solving information, offering new insight into how reasoning is represented and where it breaks down.

Houman Mehrafarin, Amit Parekh, Ioannis Konstas

Heriot-Watt University, Edinburgh, United Kingdom · University of Edinburgh, Edinburgh, United Kingdom

Transformer Interpretability CoT Reasoning Causal Interventions in Language Models Reasoning Trace Analysis

Apr 24, 2026

Real-Time AttentionBender: Granular Interactive Network Bending of Video Diffusion Transformers

Generative video models have achieved remarkable visual fidelity, yet their prompt-only interface offers thin creative agency and obscures the model's material process from the artists working with it. We present Real-Time AttentionBender, a tool that extends the practice of network bending across the full depth of the video diffusion transformer (DiT) and brings it into live, interactive generation. Built as a plugin within the DayDream Scope ecosystem and wrapping open-source real-time Wan pipelines, the tool exposes self-attention, cross-attention, and the feed-forward network as independently manipulable surfaces, with targeting down to individual diffusion steps, DiT layers, prompt tokens, and hidden neurons. The immediacy of live manipulation affords what we call "material intimacy" with the model: a responsive, near-mechanistic feel for how specific layers and neurons shape generated video. We position the tool as simultaneously an XAIxArts probe into transformer internals and an expressive instrument for discovering aesthetics outside the model's default representational space.

Adam Cole, Rebecca Fiebrink, Mick Grierson

Creative Computing Institute University of the Arts London London, UK

Transformer Interpretability Video Diffusion Models Real-Time Interactive Video Generation Computational Creativity Controllable Video Generation

Apr 24, 2026

Dissociating Decodability and Causal Use in Bracket-Sequence Transformers

When trained on tasks requiring an understanding of hierarchical structure, transformers have been found to represent this hierarchy in distinct ways: in the geometry of the residual stream, and in stack-like attention patterns maintaining a last-in, first-out ordering. However, it remains unclear whether these representations are causally used or merely decodable. We examine this gap in transformers trained on the Dyck language (a formal language of balanced bracket sequences), where the hierarchical ground truth is explicit. By probing and intervening on the residual stream and attention patterns, we find that depth, distance, and top-of-stack signals are all decodable, yet their causal roles diverge. Specifically, masking attention to the true top-of-stack position causes a sharp drop in long-distance accuracy, while ablating low-dimensional residual stream subspaces has comparatively little effect. These results, which extend to a templated natural language setting, suggest that even in a controlled setting where the relevant hierarchical variables are known, decodability alone does not imply causal use.

Aryan Sharma, Cutter Dawes, Shivam Raval

1Yale University · 2Independent · 3Harvard University

Transformer Interpretability Self-Attention Transformer Attention Attention Mechanisms Causal Interpretability Causal Interventions in Language Models

Apr 23, 2026

Evaluating Post-hoc Explanations of the Transformer-based Genome Language Model DNABERT-2

Explaining deep neural network predictions on genome sequences enables biological insight and hypothesis generation-often of greater interest than predictive performance alone. While explanations of convolutional neural networks (CNNs) have been shown to capture relevant patterns in genome sequences, it is unclear whether this transfers to more expressive Transformer-based genome language models (gLMs). To answer this question, we adapt AttnLRP, an extension of layer-wise relevance propagation to the attention mechanism, and apply it to the state-of-the-art gLM DNABERT-2. Thereby, we propose strategies to transfer explanations from token and nucleotide level. We evaluate the adaption of AttnLRP on genomic datasets using multiple metrics. Further, we provide an extensive comparison between the explanations of DNABERT-2 and a baseline CNN. Our results demonstrate that AttnLRP yields reliable explanations corresponding to known biological patterns. Hence, like CNNs, gLMs can also help derive biological insights. This work contributes to the explainability of gLMs and addresses the comparability of relevance attributions across different architectures.

Isabel Kurth, Paulo Yanez Sarmiento, Bernhard Y. Renard

Hasso Plattner Institute University of Potsdam

Transformer Interpretability Attention Mechanisms Genomics Layer-Wise Relevance Propagation

Apr 22, 2026

AttentionBender: Manipulating Cross-Attention in Video Diffusion Transformers as a Creative Probe

We present AttentionBender, a tool that manipulates cross-attention in Video Diffusion Transformers to help artists probe the internal mechanics of black-box video generation. While generative outputs are increasingly realistic, prompt-only control limits artists' ability to build intuition for the model's material process or to work beyond its default tendencies. Using an autobiographical research-through-design approach, we built on Network Bending to design AttentionBender, which applies 2D transforms (rotation, scaling, translation, etc.) to cross-attention maps to modulate generation. We assess AttentionBender by visualizing 4,500+ video generations across prompts, operations, and layer targets. Our results suggest that cross-attention is highly entangled: targeted manipulations often resist clean, localized control, producing distributed distortions and glitch aesthetics over linear edits. AttentionBender contributes a tool that functions both as an Explainable AI style probe of transformer attention mechanisms, and as a creative technique for producing novel aesthetics beyond the model's learned representational space.

Adam Cole, Mick Grierson

University of the Arts London London, UK

Transformer Interpretability Video Diffusion Models Cross-Attention Attention Control in Diffusion Models Controllable Video Generation

Apr 21, 2026

Cognitive Alignment At No Cost: Inducing Human Attention Biases For Interpretable Vision Transformers

For state-of-the-art image understanding, Vision Transformers (ViTs) have become the standard architecture but their processing diverges substantially from human attentional characteristics. We investigate whether this cognitive gap can be shrunk by fine-tuning the self-attention weights of Google's ViT-B/16 on human saliency fixation maps. To isolate the effects of semantically relevant signals from generic human supervision, the tuned model is compared against a shuffled control. Fine-tuning significantly improved alignment across five saliency metrics and induced three hallmark human-like biases: tuning reversed the baseline's anti-human large-object bias toward small-objects, amplified the animacy preference and diminished extreme attention entropy. Bayesian parity analysis provides decisive to very-strong evidence that this cognitive alignment comes at no cost to the model's original classification performance on in- (ImageNet), corrupted (ImageNet-C) and out-of-distribution (ObjectNet) benchmarks. An equivalent procedure applied to a ResNet-50 Convolutional Neural Network (CNN) instead degraded both alignment and accuracy, suggesting that the ViT's modular self-attention mechanism is uniquely suited for dissociating spatial priority from representational logic. These findings demonstrate that biologically grounded priors can be instilled as a free emergent property of human-aligned attention, to improve transformer interpretability.

Ethan Knights

Cambridge, UK

Transformer Interpretability Visual Attention AI Alignment Vision Transformer Attention Mechanisms

Apr 20, 2026

Decision-Aware Attention Propagation for Vision Transformer Explainability

Vision Transformers (ViTs) have become a dominant architecture in computer vision, yet their prediction process remains difficult to interpret because information is propagated through complex interactions across layers and attention heads. Existing attention based explanation methods provide an intuitive way to trace information flow. However, they rely mainly on raw attention weights, which do not explicitly reflect the final decision and often lead to explanations with limited class discriminability. In contrast, gradient based localization methods are more effective at highlighting class specific evidence, but they do not fully exploit the hierarchical attention propagation mechanism of transformers. To address this limitation, we propose Decision-Aware Attention Propagation (DAP), an attribution method that injects decision-relevant priors into transformer attention propagation. By estimating token importance through gradient based localization and integrating it into layer wise attention rollout, the method captures both the structural flow of attention and the evidence most relevant to the final prediction. Consequently, DAP produces attribution maps that are more class sensitive, compact, and faithful than those generated by conventional attention based methods. Extensive experiments across Vision Transformer variants of different model scales show that DAP consistently outperforms existing baselines in both quantitative metrics and qualitative visualizations, indicating that decision aware propagation is an effective direction for improving ViT interpretability.

Sehyeong Jo, Gangjae Jang, Haesol Park

Aarhus University · University of Colorado Boulder · KIST

Transformer Interpretability Vision Transformer Neural Network Interpretability

Apr 18, 2026

One Mask to Rule Them All: On Hidden Facts after Editing and How to Find Them

Knowledge editing methods such as ROME and MEMIT update factual associations in transformer models by modifying MLP weights. While evaluated mainly by output behavior, their internal mechanism remains underexplored. We investigate whether edits rely on a common mechanism, regardless of which fact is modified. Despite fact-specific weight changes, we argue that ROME and MEMIT target the same subset of weights critical for maintaining edits. To isolate this subset, we train a compact binary mask over the edited weights. The mask reverses 80% of edits on the training set and over 70% on the test set, confirming that diverse edits share a common functional structure. Our analysis reveals that the mask reverses edits by eliminating overattention in later layers. Additionally, we show that injecting the mask during editing drops editing success from 98% to 38%, demonstrating that this mechanism is necessary for edits to succeed. Our finding that edits suppress rather than overwrite knowledge explains why ROME and MEMIT fail to propagate changes to related facts. The identified common functional subspace informs detection and defense against unwanted edits.

Ali Holmov, Paul Youssef, Nandi Schoots +1

Technical University of Munich · Marburg University · University of Oxford

Knowledge Editing Transformer Interpretability LLM Auditing Factual Knowledge in Language Models Model Editing Language Model Editing+1

Apr 18, 2026

Structural Instability of Feature Composition

Sparse Autoencoders (SAEs) have emerged as a powerful paradigm for disentangling feature superposition in transformer-based architectures, enabling precise control via activation steering. However, the theoretical foundations of compositional steering -- the simultaneous activation of distinct semantic latents -- remain under-explored. The prevailing Linear Representation Hypothesis often abstracts away non-linear interference effects that arise in overcomplete dictionaries. We present a geometric framework for analyzing the instability of feature unions. Modeling the activation space as a high-dimensional sparse cone manifold, we derive an asymptotic compositional-collapse threshold under a spherical dictionary model, characterized by the Gaussian mean width (statistical dimension) of the signal cone. We further show that, in the high-bias regime, ReLU rectification converts microscopic correlation-induced variance fluctuations into a systematic drift that accumulates under composition, yielding interference growth consistent with a ratchet effect. We validate the predicted scaling trends on structured semantic features extracted from CLEVR, where hierarchical correlations accelerate the transition relative to random baselines. Together, our results highlight geometric constraints on the scalability of union-based steering and motivate composition mechanisms that explicitly manage interference beyond naive linear superposition.

Yunpeng Zhou

Whiteknights House, Whiteknights, Reading, RG6 6UR

Feature Superposition Transformer Interpretability Representation Disentanglement Sparse Autoencoders

Apr 17, 2026

Towards Intrinsic Interpretability of Large Language Models:A Survey of Design Principles and Architectures

While Large Language Models (LLMs) have achieved strong performance across many NLP tasks, their opaque internal mechanisms hinder trustworthiness and safe deployment. Existing surveys in explainable AI largely focus on post-hoc explanation methods that interpret trained models through external approximations. In contrast, intrinsic interpretability, which builds transparency directly into model architectures and computations, has recently emerged as a promising alternative. This paper presents a systematic review of the recent advances in intrinsic interpretability for LLMs, categorizing existing approaches into five design paradigms: functional transparency, concept alignment, representational decomposability, explicit modularization, and latent sparsity induction. We further discuss open challenges and outline future research directions in this emerging field. The paper list is available at: https://github.com/PKU-PILLAR-Group/Survey-Intrinsic-Interpretability-of-LLMs.

Yutong Gao, Qinglin Meng, Yuan Zhou +1

MOE Key Lab of Computational Linguistics, Peking University · Nanjing University of Science and Technology · Purdue University +1

Transformer Interpretability LLM Interpretability

Apr 16, 2026

What Is the Minimum Architecture for Prolepsis? Early Irrevocable Commitment Across Tasks in Small Transformers

When do transformers commit to a decision, and what prevents them from correcting it? We introduce prolepsis: a transformer commits early, task-specific attention heads sustain the commitment, and no layer corrects it. Replicating \citeauthor{lindsey2025biology}'s (\citeyear{lindsey2025biology}) planning-site finding on open models (Gemma~~2 2B, Llama~~3.2 1B), we ask five questions. (Q1)~Planning is invisible to six residual-stream methods; CLTs are necessary. (Q2)~The planning-site spike replicates with identical geometry. (Q3)~Specific attention heads route the decision to the output, filling a gap flagged as invisible to attribution graphs. (Q4)~Search requires

{\leq}16

layers; commitment requires more. (Q5)~Factual recall shows the same motif at a different network depth, with zero overlap between recurring planning heads and the factual top-10. Prolepsis is architectural: the template is shared, the routing substrates differ. All experiments run on a single consumer GPU (16,GB VRAM).

Éric Jacopin

Cosmic AI

Transformer Interpretability Attention Head Analysis Transformer LLM Planning Mechanistic Interpretability Attention Mechanisms+1

Apr 16, 2026

Hallucination as Trajectory Commitment: Causal Evidence for Asymmetric Attractor Dynamics in Transformer Generation

We present causal evidence that hallucination in autoregressive language models is an early trajectory commitment governed by asymmetric attractor dynamics. Using same-prompt bifurcation, in which we repeatedly sample identical inputs to observe spontaneous divergence, we isolate trajectory dynamics from prompt-level confounds. On Qwen2.5-1.5B across 61 prompts spanning six categories, 27 prompts (44.3%) bifurcate with factual and hallucinated trajectories diverging at the first generated token (KL = 0 at step 0, KL > 1.0 at step 1). Activation patching across 28 layers reveals a pronounced causal asymmetry: injecting a hallucinated activation into a correct trajectory corrupts output in 87.5% of trials (layer 20), while the reverse recovers only 33.3% (layer 24); both exceed the 10.4% baseline (p = 0.025) and 12.5% random-patch control. Window patching shows correction requires sustained multi-step intervention, whereas corruption needs only a single perturbation. Probing the prompt encoding itself, step-0 residual states predict per-prompt hallucination rate at Pearson r = 0.776 at layer 15 (p < 0.001 against a 1000-permutation null); unsupervised clustering identifies five regime-like groups (eta^2 = 0.55) whose saddle-adjacent cluster concentrates 12 of the 13 bifurcating false-premise prompts, indicating that the basin structure is organized around regime commitments fixed at prompt encoding. These findings characterize hallucination as a locally stable attractor basin: entry is probabilistic and rapid, exit demands coordinated intervention across layers and steps, and the relevant basins are selected by clusterable regimes already discernible at step 0.

G. Aytug Akarlar

Chimera Research Initiative Istanbul, Türkiye · Chimera Research Initiative is an independent research effort exploring causal and neuro-symbolic approaches to AI systems.

Transformer Interpretability Hallucination in Language Models Autoregressive Language Modeling

Apr 15, 2026

Seeing Through Circuits: Faithful Mechanistic Interpretability for Vision Transformers

Transparency of neural networks' internal reasoning is at the heart of interpretability research, adding to trust, safety, and understanding of these models. The field of mechanistic interpretability has recently focused on studying task-specific computational graphs, defined by connections (edges) between model components. Such edge-based circuits have been defined in the context of large language models, yet vision-based approaches so far only consider neuron-based circuits. These tell which information is encoded, but not how it is routed through the complex wiring of a neural network. In this work, we investigate whether useful mechanistic circuits can be identified through computational graphs in vision transformers. We propose an effective method for Automatic Visual Circuit Discovery (Vi-CD) that recovers class-specific circuits for classification, identifies circuits underlying typographic attacks in CLIP, and discovers circuits that lend themselves for steering to correct harmful model behavior. Overall, we find that insightful and actionable edge-based circuits can be recovered from vision transformers, adding transparency to the internal computations of these models.

Nina Żukowska, Wolfgang Stammer, Bernt Schiele +1

Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbrücken, Germany

Transformer Interpretability Mechanistic Interpretability Circuit Discovery Vision Transformer VLM Interpretability

Apr 15, 2026

Applied Explainability for Large Language Models: A Comparative Study

Large language models (LLMs) achieve strong performance across many natural language processing tasks, yet their decision processes remain difficult to interpret. This lack of transparency creates challenges for trust, debugging, and deployment in real-world systems. This paper presents an applied comparative study of three explainability techniques: Integrated Gradients, Attention Rollout, and SHAP, on a fine-tuned DistilBERT model for SST-2 sentiment classification. Rather than proposing new methods, the focus is on evaluating the practical behavior of existing approaches under a consistent and reproducible setup. The results show that gradient-based attribution provides more stable and intuitive explanations, while attention-based methods are computationally efficient but less aligned with prediction-relevant features. Model-agnostic approaches offer flexibility but introduce higher computational cost and variability. This work highlights key trade-offs between explainability methods and emphasizes their role as diagnostic tools rather than definitive explanations. The findings provide practical insights for researchers and engineers working with transformer-based NLP systems. This is a preprint and has not undergone peer review.

Venkata Abhinandan Kancharla

NRI Institute of Technology

LLM Interpretability Explainability Evaluation Transformer Interpretability Feature Attribution Faithfulness of Language Model Explanations

Apr 14, 2026

Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision

Understanding the internal activations of Vision Transformers (ViTs) is critical for building interpretable and trustworthy models. While Sparse Autoencoders (SAEs) have been used to extract human-interpretable features, they operate on individual layers and fail to capture the cross-layer computational structure of Transformers, as well as the relative significance of each layer in forming the last-layer representation. Alternatively, we introduce the adoption of Cross-Layer Transcoders (CLTs) as reliable, sparse, and depth-aware proxy models for MLP blocks in ViTs. CLTs use an encoder-decoder scheme to reconstruct each post-MLP activation from learned sparse embeddings of preceding layers, yielding a linear decomposition that transforms the final representation of ViTs from an opaque embedding into an additive, layer-resolved construction that enables faithful attribution and process-level interpretability. We train CLTs on CLIP ViT-B/32 and ViT-B/16 across CIFAR-100, COCO, and ImageNet-100. We show that CLTs achieve high reconstruction fidelity with post-MLP activations while preserving and even improving, in some cases, CLIP zero-shot classification accuracy. In terms of interpretability, we show that the cross-layer contribution scores provide faithful attribution, revealing that the final representation is concentrated in a smaller set of dominant layer-wise terms whose removal degrades performance and whose retention largely preserves it. These results showcase the significance of adopting CLTs as an alternative interpretable proxy of ViTs in the vision domain.

Gerasimos Chatzoudis, Konstantinos D. Polyzos, Zhuowei Li +4

Rutgers University · University of California San Diego

Transformer Interpretability Vision Transformer VLM Interpretability

Apr 14, 2026

Do Transformers Use their Depth Adaptively? Evidence from a Relational Reasoning Task

We investigate whether transformers use their depth adaptively across tasks of increasing difficulty. Using a controlled multi-hop relational reasoning task based on family stories, where difficulty is determined by the number of relationship hops that must be composed, we monitor (i) how predictions evolve across layers via early readouts (the logit lens) and (ii) how task-relevant information is integrated across tokens via causal patching. For pretrained models, we find some limited evidence for adaptive depth use: some larger models need fewer layers to arrive at plausible answers for easier tasks, and models generally use more layers to integrate information across tokens as chain length increases. For models finetuned on the task, we find clearer and more consistent evidence of adaptive depth use, with the effect being stronger for less constrained finetuning regimes that do not preserve general language modeling abilities.

Alicia Curth, Rachel Lawrence, Sushrut Karmalkar +1

Microsoft Research Cambridge

Adaptive Computation Transformer Interpretability

Apr 14, 2026

Distinct mechanisms underlying in-context learning in transformers

Modern distributed networks, notably transformers, acquire a remarkable ability (termed `in-context learning') to adapt their computation to input statistics, such that a fixed network can be applied to data from a broad range of systems. Here, we provide a complete mechanistic characterization of this behavior in transformers trained on a finite set

S

of discrete Markov chains. The transformer displays four algorithmic phases, characterized by whether the network memorizes and generalizes, and whether it uses 1-point or 2-point statistics. We show that the four phases are implemented by multi-layer subcircuits that exemplify two qualitatively distinct mechanisms for implementing context-adaptive computations. Minimal models isolate the key features of both motifs. Memorization and generalization phases are delineated by two boundaries that depend on data diversity,

K = |S|

. The first (

K_1^\ast

) is set by a kinetic competition between subcircuits and the second (

K_2^\ast

) is set by a representational bottleneck. A symmetry-constrained theory of a transformer's training dynamics explains the sharp transition from 1-point to 2-point generalization and identifies key features of the loss landscape that allow the network to generalize. Put together, we show that transformers develop distinct subcircuits to implement in-context learning and identify conditions that favor certain mechanisms over others.

Cole Gibson, Wenping Cui, Gautam Reddy

Lewis-Sigler Institute for Integrative Genomics, Princeton University, Princeton, NJ 08544, USA · Department of Physics, Princeton University, Princeton, NJ 08544, USA

In-Context Learning Transformer Training Dynamics Transformer Interpretability Neural Network Training Dynamics Mechanistic Interpretability Statistical Physics of Learning

Apr 13, 2026

The Linear Centroids Hypothesis: Features as Directions Learned by Local Experts

The Linear Representation Hypothesis (LRH) identifies features of a trained deep network (DN) as linear directions in the activation spaces, i.e., output spaces of intermediate layers. This characterization decouples the input-output maps learned by a DN from the organization of feature directions in its activation spaces. We introduce the Linear Centroids Hypothesis (LCH), which instead identifies features with linear directions among a DN's centroid spaces -- where any vector denotes a centroid or summary of a local affine expert characterizing the learned input-output maps of the DN exactly (e.g., for piecewise-affine DNs) or approximately (e.g., for smooth DNs like transformers). We show that replacing intermediate activations with centroids yields a functional drop-in alternative for standard interpretability tools. Empirically, this change yields sparser, more downstream-useful feature dictionaries on DINO ViTs, suppresses spurious directions on a controlled task, recovers interpretable circuits in GPT2-Large, and produces faithful gradient-based saliency maps. LCH unifies dictionaries, probing, circuits, and saliency maps into a single geometric object grounded in the network's input-output map -- making interpretability mechanistic by construction rather than post hoc. Code to study the LCH https://github.com/ThomasWalker1/LinearCentroidsHypothesis .

Thomas Walker, Ahmed Imtiaz Humayun, Randall Balestriero +1

Rice University · Google Research · Brown University

Neural Network Interpretability Mechanistic Interpretability Transformer Interpretability Representation Geometry in Language Models

Apr 13, 2026

Symmetry Reveals Layerwise Dynamics: How Transformers Perform In-Context Classification

Transformers can perform in-context classification from a few labeled examples, yet the inference-time algorithm remains opaque. We study multi-class linear classification in the hard no-margin regime and make the computation identifiable by enforcing feature- and label-permutation equivariance at every layer. This enables interpretability while maintaining functional equivalence and yields highly structured weights. From these models we extract an explicit depth-indexed recursion: an end-to-end identified, emergent update rule inside a softmax transformer, to our knowledge the first of its kind. Attention matrices formed from mixed feature-label Gram structure drive coupled updates of training points, labels, and the test probe. The resulting dynamics implement a geometry-driven algorithmic motif, which can provably amplify class separation and yields robust expected class alignment.

Patrick Lutz, Themistoklis Haris, Arjun Chandra +2

Boston University, Departments of Computer Science and Electrical & Computer Engineering.

In-Context Learning Transformer Interpretability Transformer Attention Linear Classification

Apr 11, 2026

Tracing the Thought of a Grandmaster-level Chess-Playing Transformer

While modern transformer neural networks achieve grandmaster-level performance in chess and other reasoning tasks, their internal computation process remains largely opaque. Focusing on Leela Chess Zero (LC0), we introduce a sparse decomposition framework to interpret its internal computation by decomposing its MLP and attention modules with sparse replacement layers, which capture the primary computation process of LC0. We conduct a detailed case study showing that these pathways expose rich, interpretable tactical considerations that are empirically verifiable. We further introduce three quantitative metrics and show that LC0 exhibits parallel reasoning behavior consistent with the inductive bias of its policy head architecture. To the best of our knowledge, this is the first work to decompose the internal computation of a transformer on both MLP and attention modules for interpretability. Combining sparse replacement layers and causal interventions in LC0 provides a comprehensive understanding of advanced tactical reasoning, offering critical insights into the underlying mechanisms of superhuman systems. Our code is available at https://github.com/JacklE0niden/Leela-SAEs.

Rui Lin, Zhenyu Jin, Guancheng Zhou +7

Shanghai Innovation Institute, Shanghai, China · School of Computer Science, Fudan University, Shanghai, China · School of Mathematics and Statistics, Xi’an Jiaotong University.

Transformer Interpretability Mechanistic Interpretability Transformer Attention Sparse Coding

Apr 11, 2026

Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation

As the foundational architecture of modern machine learning, Transformers have driven remarkable progress across diverse AI domains. Despite their transformative impact, a persistent challenge across various Transformers is Attention Sink (AS), in which a disproportionate amount of attention is focused on a small subset of specific yet uninformative tokens. AS complicates interpretability, significantly affecting the training and inference dynamics, and exacerbates issues such as hallucinations. In recent years, substantial research has been dedicated to understanding and harnessing AS. However, a comprehensive survey that systematically consolidates AS-related research and offers guidance for future advancements remains lacking. To address this gap, we present the first survey on AS, structured around three key dimensions that define the current research landscape: Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation. Our work makes a pivotal contribution by highlighting the key concepts and main trends in the field, guiding researchers through the evolution of AS-related studies. We envision this survey as a valuable resource, empowering researchers to effectively manage AS within the current Transformer paradigm, while simultaneously inspiring innovative advancements for the next generation of Transformers. The paper list of this work is available at https://github.com/ZunhaiSu/Awesome-Attention-Sink.

Zunhai Su, Hengyuan Zhang, Wei Wu +24

Tsinghua University · Meituan LongCat Team · The University of Hong Kong +6

Attention Sinks Transformer Attention Transformer Interpretability

Apr 10, 2026

Relational Preference Encoding in Looped Transformer Internal States

We investigate how looped transformers encode human preference, training lightweight evaluator heads on frozen Ouro-2.6B loop-iteration states on Anthropic HH-RLHF. v2: an erratum is prepended; the original manuscript is unchanged. A post-publication audit found the three headline results inflated by two independent evaluation errors. The 95.2% pairwise evaluator accuracy is a canonical-ordering artifact: the data were correctly split, but the evaluator learned to prefer the first-presented argument; its strict antisymmetrized accuracy on the full 8,552-pair test set is 63.9%. The 84.5% pairwise probe and the below-chance 21.75% pointwise probe were source-item leaks (orientation rows and pair partners crossing the train/test split); corrected pair-disjoint values are 56.5% and 54.2% -- above chance, so the "inverted polarity" finding is withdrawn. The central finding survives at much smaller magnitude: preference is decoded more accurately relationally than pointwise (paired +2.3 points, 95% CI [+1.3, +3.3]), and the antisymmetrized evaluator still beats the linear probe, but no corrected readout rivals end-to-end reward models. The methodological findings stand (constant-output degeneracy, flip test, swap-protocol metric deflation), with one correction: antisymmetrized accuracy, not antisymmetry correlation, certifies relational discrimination. The two errors are mutually invisible -- a split audit cannot see an ordering prior, antisymmetrization cannot see a leak -- so both checks are required. Full audit in the follow-up work (Kirin, 2026, in preparation).

Jan Kirin

Pairwise Preference Learning Reward Modeling Transformer Interpretability Pairwise Comparison Pairwise Preference Evaluation Pairwise Ranking+2

Apr 8, 2026

Selective Neuron Amplification in Transformer Language Models

Large language models often fail on tasks they seem to already understand. In our experiments, this appears to be less about missing knowledge and more about certain internal circuits not being strongly activated during inference. We explore Selective Neuron Amplification, which increases the influence of task relevant neurons without changing the model's parameters. The method works at inference time and does not permanently alter the model. SNA helps mainly when the model is uncertain, while having low effect when the model is already confident. This suggests that some model failures are due to weak activation rather than lack of capability.

Ryyan Akhtar, Payal Pahwa, Monika Arora

Department of Artificial Intelligence and Data Science, Guru Gobind Singh Indraprastha University, BPIT, New Delhi, India

Causal Interventions in Language Models Activation Steering Transformer Interpretability

Apr 7, 2026

LAG-XAI: A Lie-Inspired Affine Geometric Framework for Interpretable Paraphrasing in Transformer Latent Spaces

Modern Transformer-based language models achieve strong performance in natural language processing tasks, yet their latent semantic spaces remain largely uninterpretable black boxes. This paper introduces LAG-XAI (Lie Affine Geometry for Explainable AI), a novel geometric framework that models paraphrasing not as discrete word substitutions, but as a structured affine transformation within the embedding space. By conceptualizing paraphrasing as a continuous geometric flow on a semantic manifold, we propose a computationally efficient mean-field approximation, inspired by local Lie group actions. This allows us to decompose paraphrase transitions into geometrically interpretable components: rotation, deformation, and translation. Experiments on the noisy PIT-2015 Twitter corpus, encoded with Sentence-BERT, reveal a "linear transparency" phenomenon. The proposed affine operator achieves an AUC of 0.7713. By normalizing against random chance (AUC 0.5), the model captures approximately 80% of the non-linear baseline's effective classification capacity (AUC 0.8405), offering explicit parametric interpretability in exchange for a marginal drop in absolute accuracy. The model identifies fundamental geometric invariants, including a stable matrix reconfiguration angle (~27.84°) and near-zero deformation, indicating local isometry. Cross-domain generalization is confirmed via direct cross-corpus validation on an independent TURL dataset. Furthermore, the practical utility of LAG-XAI is demonstrated in LLM hallucination detection: using a "cheap geometric check," the model automatically detected 95.3% of factual distortions on the HaluEval dataset by registering deviations beyond the permissible semantic corridor. This approach provides a mathematically grounded, resource-efficient path toward the mechanistic interpretability of Transformers.

Olexander Mazurets, Olexander Barmak, Leonid Bedratyuk +1

Department of Computer Science, Khmelnytskyi National University, Khmelnytskyi, 29016, Ukraine · Department of Theoretical Cybernetics, Taras Shevchenko National University of Kyiv, Kyiv, 03187, Ukraine · Laboratory of Communicative Information Technologies, V.M. Glushkov Institute of Cybernetics of NAS of Ukraine, Kyiv, 03187, Ukraine

Transformer Interpretability Representation Geometry in Language Models

Apr 7, 2026

Disentangling MLP Neuron Weights in Vocabulary Space

Interpreting the information encoded in language model weights remains a fundamental challenge in mechanistic interpretability. In this work, we introduce ROTATE (Rotation-Optimized Token Alignment in weighT spacE), a data-free method requiring no forward passes that disentangles MLP neurons directly in weight space. Our approach relies on a key statistical observation: neurons that encode coherent, monosemantic concepts exhibit high kurtosis when projected onto the model's vocabulary. By optimizing rotations of neuron weights to maximize their vocabulary-space kurtosis, our method recovers sparse, interpretable directions which we name vocabulary channels. Experiments on Llama-3.1-8B-Instruct and Gemma-2-2B-it demonstrate that ROTATE consistently recovers vocabulary channels that are faithful to the neuron's behavior; ablating individual channels selectively disables corresponding input activations or the promotion of specific concepts. Moreover, aggregating channel-level descriptions yields comprehensive neuron descriptions that outperform optimized activation-based baselines by 2-3x in head-to-head comparisons. By providing a data-free decomposition of neuron weights, ROTATE offers a scalable, fine-grained building block for interpreting language models.

Asaf Avrahamy, Yoav Gur-Arieh, Mor Geva

Blavatnik School of Computer Science and AI, Tel Aviv University

LLM Interpretability Transformer Interpretability Mechanistic Interpretability

Apr 7, 2026

Selective Aggregation of Attention Maps Improves Diffusion-Based Visual Interpretation

Numerous studies on text-to-image (T2I) generative models have utilized cross-attention maps to boost application performance and interpret model behavior. However, the distinct characteristics of attention maps from different attention heads remain relatively underexplored. In this study, we show that selectively aggregating cross-attention maps from heads most relevant to a target concept can improve visual interpretability. Compared to the diffusion-based segmentation method DAAM, our approach achieves higher mean IoU scores. We also find that the most relevant heads capture concept-specific features more accurately than the least relevant ones, and that selective aggregation helps diagnose prompt misinterpretations. These findings suggest that attention head selection offers a promising direction for improving the interpretability and controllability of T2I generation.

Jungwon Park, Jungmin Ko, Dongnam Byun +1

Department of Intelligence and Information, Seoul National University, Seoul, Republic of Korea · Interdisciplinary Program in Artificial Intelligence, Seoul National University, Seoul, Republic of Korea

T2I Diffusion Models Cross-Attention Transformer Interpretability Text-Guided Image Segmentation

Apr 7, 2026

Learn to Rank: Visual Attribution by Learning Importance Ranking

Interpreting the decisions of complex computer vision models is crucial to establish trust and accountability, especially in safety-critical domains. An established approach to interpretability is generating visual attribution maps that highlight regions of the input most relevant to the model's prediction. However, existing methods face a three-way trade-off. Propagation-based approaches are efficient, but they can be biased and architecture-specific. Meanwhile, perturbation-based methods are causally grounded, yet they are expensive and for vision transformers often yield coarse, patch-level explanations. Learning-based explainers are fast but usually optimize surrogate objectives or distill from heuristic teachers. We propose a learning scheme that instead optimizes deletion and insertion metrics directly. Since these metrics depend on non-differentiable sorting and ranking, we frame them as permutation learning and replace the hard sorting with a differentiable relaxation using Gumbel-Sinkhorn. This enables end-to-end training through attribution-guided perturbations of the target model. During inference, our method produces dense, pixel-level attributions in a single forward pass with optional, few-step gradient refinement. Our experiments demonstrate consistent quantitative improvements and sharper, boundary-aligned explanations, particularly for transformer-based vision models. Code and pretrained models are available at https://github.com/dschinagl/AHA.

David Schinagl, Christian Fruhwirth-Reisinger, Alexander Prutsch +2

Institute of Visual Computing (IVC), Graz University of Technology, Graz, Austria · Nomadic AI, USA · Amazon, USA

Feature Attribution Explainable Image Classification Transformer Interpretability

Apr 6, 2026

Noise Immunity in In-Context Tabular Learning: An Empirical Robustness Analysis of TabPFN's Attention Mechanisms

Tabular foundation models (TFMs) such as TabPFN (Tabular Prior-Data Fitted Network) are designed to generalize across heterogeneous tabular datasets through in-context learning (ICL). They perform prediction in a single forward pass conditioned on labeled examples without dataset-specific parameter updates. This paradigm is particularly attractive in industrial domains (e.g., finance and healthcare) where tabular prediction is pervasive. Retraining a bespoke model for each new table can be costly or infeasible in these settings, while data quality issues such as irrelevant predictors, correlated feature groups, and label noise are common. In this paper, we provide strong empirical evidence that TabPFN is highly robust under these sub-optimal conditions. We study TabPFN and its attention mechanisms for binary classification problems with controlled synthetic perturbations that vary: (i) dataset width by injecting random uncorrelated features and by introducing nonlinearly correlated features, (ii) dataset size by increasing the number of training rows, and (iii) label quality by increasing the fraction of mislabeled targets. Beyond predictive performance, we analyze internal signals including attention concentration and attention-based feature ranking metrics. Across these parametric tests, TabPFN is remarkably resilient: ROC-AUC remains high, attention stays structured and sharp, and informative features are highly ranked by attention-based metrics. Qualitative visualizations with attention heatmaps, feature-token embeddings, and SHAP plots further support a consistent pattern across layers in which TabPFN increasingly concentrates on useful features while separating their signals from noise. Together, these findings suggest that TabPFN is a robust TFM capable of maintaining both predictive performance and coherent internal behavior under various scenarios of data imperfections.

James Hu, Mahdi Ghelichi

Model Development Innovation, Risk Management TD Bank, Toronto, Canada

Tabular ML Tabular ICL Tabular Foundation Models Transformer Interpretability

Apr 4, 2026

Automated Attention Pattern Discovery at Scale in Large Language Models

Large language models have found success by scaling up capabilities to work in general settings. The same can unfortunately not be said for interpretability methods. The current trend in mechanistic interpretability is to provide precise explanations of specific behaviors in controlled settings. These often do not generalize, or are too resource intensive for larger studies. In this work we propose to study repeated behaviors in large language models by mining completion scenarios in Java code datasets, through exploiting the structured nature of code. We collect the attention patterns generated in the attention heads to demonstrate that they are scalable signals for global interpretability of model components. We show that vision models offer a promising direction for analyzing attention patterns at scale. To demonstrate this, we introduce the Attention Pattern - Masked Autoencoder(AP-MAE), a vision transformer-based model that efficiently reconstructs masked attention patterns. Experiments on StarCoder2 show that AP-MAE (i) reconstructs masked attention patterns with high accuracy, (ii) generalizes across unseen models with minimal degradation, (iii) reveals recurring patterns across inferences, (iv) predicts whether a generation will be correct without access to ground truth, with accuracies ranging from 55% to 70% depending on the task, and (v) enables targeted interventions that increase accuracy by 13.6% when applied selectively, but cause collapse when applied excessively. These results establish attention patterns as a scalable signal for interpretability and demonstrate that AP-MAE provides a transferable foundation for both analysis and intervention in large language models. Beyond its standalone value, AP-MAE also serves as a selection procedure to guide fine-grained mechanistic approaches. We release code and models to support future work in large-scale interpretability.

Jonathan Katzy, Razvan-Mihai Popescu, Erik Mekkes +2

Delft University of Technology

Transformer Interpretability LLM Interpretability Attention Mechanisms Masked Autoencoders