Transformer Interpretability

Ask about this topic

Overview

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

CardsList

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

Mar 9, 2026

SYNAPSE: Framework for Neuron Analysis and Perturbation in Sequence Encoding

In recent years, Artificial Intelligence has become a powerful partner for complex tasks such as data analysis, prediction, and problem-solving, yet its lack of transparency raises concerns about its reliability. In sensitive domains such as healthcare or cybersecurity, ensuring transparency, trustworthiness, and robustness is essential, since the consequences of wrong decisions or successful attacks can be severe. To overcome these limitations, this work proposes SYNAPSE, a systematic, training-free framework (requiring no retraining of the base model) for understanding and stress-testing the internal behavior of Transformer models across domains. SYNAPSE extracts per-layer [CLS] representations, trains a lightweight linear probe to obtain global and per-class neuron rankings, and applies forward-hook interventions during inference. This design enables controlled, repeatable experiments on internal representations without altering the original model, thereby allowing weaknesses, stability patterns, and label-specific sensitivities to be measured and compared directly across tasks and architectures. SYNAPSE reveals a consistent organization of internal representations across the two evaluated domains, in which task-relevant information is encoded in broad, overlapping neuron subsets. This redundancy provides a strong degree of functional stability, while class-wise asymmetries expose heterogeneous specialization patterns and enable fine-grained, label-aware analysis. In contrast, small structured manipulations in weight or logit space are sufficient to redirect predictions, highlighting complementary vulnerability profiles and illustrating how SYNAPSE can guide the development of more robust and interpretable Transformer models.

Jesús Sánchez Ochoa, Enrique Tomás Martínez Beltrán, Alberto Huertas Celdrán

Department of Information and Communications Engineering, University of Murcia, Murcia, 30100, Spain

Neuron-Level InterpretabilityTransformer InterpretabilityNeural Network RobustnessMechanistic Interpretability
AskPDF
Mar 9, 2026

Sensivity of LLMs' Explanations to the Training Randomness:Context, Class & Task Dependencies

Transformer models are now a cornerstone in natural language processing. Yet, explaining their decisions remains a challenge. It was shown recently that the same model trained on the same data with a different randomness can lead to very different explanations. In this paper, we investigate how the (syntactic) context, the classes to be learned and the tasks influence this explanations' sensitivity to randomness. We show that they all have statistically significant impact: smallest for the (syntactic) context, medium for the classes and largest for the tasks.

Romain Loncour, Jérémie Bogaert, François-Xavier Standaert

UCLouvain, Louvain-la-Neuve, Belgium.

LLM InterpretabilityTransformer Interpretability
AskPDF
Mar 8, 2026

Interpretable-by-Design Transformers via Architectural Stream Independence

While transformers achieve strong performance, their internal decision-making processes remain opaque. We investigate whether architectural constraints can enforce interpretability by design through architectural stream independence: maintaining a token stream (carrying symbolic structure) and contextual semantics in separated streams that remain independently observable throughout processing, with integration delayed until output. We validate this principle through the Late Fusion Architecture (LFA), which demonstrates interpretable symbolic heads through all the final layers, while standard transformers show dissolution by the third of six layers; we quantify this effect by introducing the Token-Position Dependence Score (PDS), with PDSmaxPDS_{max} = 0.276 and 0.058, respectively. Crucially, intervention experiments demonstrate functional modularity: suppressing LFA's recency heads causes minimal semantic damage (Cohen's d = -0.158) versus catastrophic entanglement in baselines (d = -0.672). LFA demonstrates that architectural constraints improve underlying learning mechanisms, averaging 42% stability versus 19% and 11% for baseline comparisons, with extremes from 50% on LFA's best pairs (6 of 12 heads position-invariant) down to 0% complete collapse in over-constrained cases. By preventing premature entanglement, architectural independence steers models toward semantic understanding over positional heuristics, establishing interpretability as an architectural design criterion enforceable through structural constraints rather than post-hoc analysis.

Clayton Kerce, Alexis Fox

Georgia Tech Research Institute

Transformer InterpretabilityNeural Network InterpretabilityMechanistic InterpretabilityCausal InterpretabilityAttention Head Analysis
AskPDF
Mar 8, 2026

Engineering Verifiable Modularity in Transformers via Per-Layer Supervision

Transformers resist surgical control. Ablating an attention head identified as critical for capitalization produces minimal behavioral change because distributed redundancy compensates for damage. This Hydra effect renders interpretability illusory: we may identify components through correlation, but cannot predict or control their causal role. We demonstrate that architectural interventions can expose hidden modularity. Our approach combines dual-stream processing separating token and contextual representations, per-layer supervision providing independent gradient signal at each depth, and gated attention regularizing toward discrete activation patterns. When trained with per-layer supervision, models produce ablation effects 5 to 23 times larger than architecturally identical controls trained with standard objectives. This enables 4 times greater control leverage on targeted behaviors: scaling identified attention heads produces smooth, predictable changes in model output. The key finding is architectural. Without per-layer supervision, ablation damage concentrates near zero with low variance (Winograd standard deviation 0.63%). With per-layer supervision, effects spread widely (standard deviation 6.32%), revealing which predictions depend on which circuits. The larger variance is not measurement noise but the signature of unmasked modularity. We validate our approach through three components: engineered features that capture computational dynamics rather than vocabulary structure (validated by near-zero correlation with raw activation clustering), an architecture providing positive control for modularity, and causal experiments demonstrating functional reorganization where different tasks route through different attention heads. This es tablishes a methodology for transforming interpretability from passive observation to active control.

J. Clayton Kerce

Georgia Tech Research Institute

Transformer InterpretabilityMechanistic InterpretabilityCausal Interventions in Language ModelsAttention Head Analysis
AskPDF
Mar 8, 2026

The Dual-Stream Transformer: Channelized Architecture for Interpretable Language Modeling

Standard transformers entangle all computation in a single residual stream, obscuring which components perform which functions. We introduce the Dual-Stream Transformer, which decomposes the residual stream into two functionally distinct components: a token stream updated by attention and a context stream updated by feed-forward networks. Information flow between attention heads is controlled through a hierarchy of mixing strategies, from fully independent (maximum interpretability) to dense (standard transformer behavior). This design exposes a tunable tradeoff between interpretability and performance. We measure this tradeoff on language modeling tasks at 29M parameters. Fully independent head mixing increases validation loss by 8% relative to dense baselines. The recommended Kronecker mixing strategy, which permits scalar communication between heads while preserving within-head structure, costs only 2.5%. All configurations maintain functional generation under attention amplification (scaling logits by factors up to 16 at inference time), with degradation ranging from 16% to 27%. This robustness suggests the architectures learn discrete algorithms that operate independently of soft probabilistic mixing. The architecture provides a foundation for interpretable language models where internal structure is exposed by design. \footnote{This work was partially supported by DARPA Contract HR001125C0302.}

J. Clayton Kerce, Alexis Fox

Georgia Tech Research Institute

TransformerTransformer InterpretabilityTransformer AttentionLanguage Modeling
AskPDF
Mar 7, 2026

Class Visualizations and Activation Atlases for Enhancing Interpretability in Deep Learning-Based Computational Pathology

The rapid adoption of transformer-based models in computational pathology has enabled prediction of molecular and clinical biomarkers from H&E whole-slide images, yet interpretability has not kept pace with model complexity. While attribution- and generative-based methods are common, feature visualization approaches such as class visualizations (CVs) and activation atlases (AAs) have not been systematically evaluated for these models. We developed a visualization framework and assessed CVs and AAs for a transformer-based foundation model across tissue and multi-organ cancer classification tasks with increasing label granularity. Four pathologists annotated real and generated images to quantify inter-observer agreement, complemented by attribution and similarity metrics. CVs preserved recognizability for morphologically distinct tissues but showed reduced separability for overlapping cancer subclasses. In tissue classification, agreement decreased from Fleiss k = 0.75 (scans) to k = 0.31 (CVs), with similar trends in cancer subclass tasks. AAs revealed layer-dependent organization: coarse tissue-level concepts formed coherent regions, whereas finer subclasses exhibited dispersion and overlap. Agreement was moderate for tissue classification (k = 0.58), high for coarse cancer groupings (k = 0.82), and low at subclass level (k = 0.11). Atlas separability closely tracked expert agreement on real images, indicating that representational ambiguity reflects intrinsic pathological complexity. Attribution-based metrics approximated expert variability in low-complexity settings, whereas perceptual and distributional metrics showed limited alignment. Overall, concept-level feature visualization reveals structured morphological manifolds in transformer-based pathology models and provides a framework for expert-centered interrogation of learned representations across label granularities.

Marco Gustav, Fabian Wolf, Christina Glasner +6

Else Kroener Fresenius Center for Digital Health, Faculty of Medicine, TUD Dresden University of Technology, Dresden, Germany · Institute of Pathology, University Medical Center Mainz, Mainz, Germany · Pathology, Faculty of Medicine, University of Augsburg, Augsburg, Germany +4

Feature VisualizationExplainable Medical Image AnalysisComputational PathologyTransformer InterpretabilityHistopathology Image Classification
AskPDF
Mar 6, 2026

Safe Transformer: An Explicit Safety Bit For Interpretable And Controllable Alignment

Current safety alignment methods encode safe behavior implicitly within model parameters, creating a fundamental opacity: we cannot easily inspect why a model refuses a request, nor intervene when its safety judgments fail. We propose Safe Transformer, a modular approach that augments pre-trained language models by inserting a discrete information bottleneck containing an explicit safety bit between transformer layers. The safety bit serves as both an interpretable signal of the model's safety classification and a controllable switch: through contrastive training, the model learns disentangled representations where the safety bit governs the behavioral mode - producing helpful responses when s=1s=1 and refusals when s=0s=0 - while additional unsupervised bits uu encode semantic content for generation. Additional unsupervised bits in the information bottleneck allow semantic information to flow through, preserving the model's generation capabilities. This design achieves both interpretability (the safety decision is directly readable) and controllability (the safety bit can be manually overridden), requiring only lightweight fine-tuning without pre-training from scratch. In red-team benchmarks, Safe Transformer achieves near-zero Attack Success Rate, substantially outperforming base models and safety fine-tuning baselines.

Jingyuan Feng, Andrew Gambardella, Gouki Minegishi +3

The University of Tokyo.

LLM Safety AlignmentLLM AlignmentTransformer Interpretability
AskPDF
Mar 5, 2026

The Spike, the Sparse and the Sink: Anatomy of Massive Activations and Attention Sinks

We study two recurring phenomena in Transformer language models: massive activations, in which a small number of tokens exhibit extreme outliers in a few channels, and attention sinks, in which certain tokens attract disproportionate attention mass regardless of semantic relevance. Prior work observes that these phenomena frequently co-occur and often involve the same tokens, but their functional roles and causal relationship remain unclear. Through systematic experiments, we show that the co-occurrence is largely an architectural artifact of modern Transformer design, and that the two phenomena serve related but distinct functions. Massive activations operate globally: they induce near-constant hidden representations that persist across layers, effectively functioning as implicit parameters of the model. Attention sinks operate locally: they modulate attention outputs across heads and bias individual heads toward short-range dependencies. We identify the pre-norm configuration as the key choice that enables the co-occurrence, and show that ablating it causes the two phenomena to decouple.

Shangwen Sun, Alfredo Canziani, Yann LeCun +1

New York University.

Massive ActivationsAttention SinksTransformer Interpretability
AskPDF
Mar 4, 2026

Feature-level Interaction Explanations in Multimodal Transformers

Multimodal Transformers often produce predictions without clarifying how different modalities jointly support a decision. Most existing multimodal explainable AI (MXAI) methods extend unimodal saliency to multimodal backbones, highlighting important tokens or patches within each modality, but they rarely pinpoint which cross-modal feature pairs provide complementary evidence (synergy) or serve as reliable backups (redundancy). We present Feature-level I2MoE (FL-I2MoE), a structured Mixture-of-Experts layer that operates directly on token/patch sequences from frozen pretrained encoders and explicitly separates unique, synergistic, and redundant evidence at the feature level. We further develop an expert-wise explanation pipeline that combines attribution with top-K% masking to assess faithfulness, and we introduce Monte Carlo interaction probes to quantify pairwise behavior: the Shapley Interaction Index (SII) to score synergistic pairs and a redundancy-gap score to capture substitutable (redundant) pairs. Across three benchmarks (MMIMDb, ENRICO, and MMHS150K), FL-I2MoE yields more interactionspecific and concentrated importance patterns than a dense Transformer with the same encoders. Finally, pair-level masking shows that removing pairs ranked by SII or redundancy-gap degrades performance more than masking randomly chosen pairs under the same budget, supporting that the identified interactions are causally relevant. Code is available at https://github.com/dut0817/FL-I2MoE.

Yeji Kim, Housam Khalifa Bashier Babiker, Mi-Young Kim +1

University of Alberta, Edmonton, AB, Canada · University of Waikato, Hamilton, New Zealand · University of Alberta, Camrose, AB, Canada

Multimodal UnderstandingFeature AttributionExplainable Artificial IntelligenceTransformer InterpretabilityMultimodal Model InterpretabilityMixture-of-Experts Models+1
AskPDF
Mar 3, 2026

Sparse autoencoders reveal organized biological knowledge but minimal regulatory logic in single-cell foundation models: a comparative atlas of Geneformer and scGPT

Background: Single-cell foundation models such as Geneformer and scGPT encode rich biological information, but whether this includes causal regulatory logic rather than statistical co-expression remains unclear. Sparse autoencoders (SAEs) can resolve superposition in neural networks by decomposing dense activations into interpretable features, yet they have not been systematically applied to biological foundation models. Results: We trained TopK SAEs on residual stream activations from all layers of Geneformer V2-316M (18 layers, d=1152) and scGPT whole-human (12 layers, d=512), producing atlases of 82525 and 24527 features, respectively. Both atlases confirm massive superposition, with 99.8 percent of features invisible to SVD. Systematic characterization reveals rich biological organization: 29 to 59 percent of features annotate to Gene Ontology, KEGG, Reactome, STRING, or TRRUST, with U-shaped layer profiles reflecting hierarchical abstraction. Features organize into co-activation modules (141 in Geneformer, 76 in scGPT), exhibit causal specificity (median 2.36x), and form cross-layer information highways (63 to 99.8 percent). When tested against genome-scale CRISPRi perturbation data, only 3 of 48 transcription factors (6.2 percent) show regulatory-target-specific feature responses. A multi-tissue control yields marginal improvement (10.4 percent, 5 of 48 TFs), establishing model representations as the bottleneck. Conclusions: These models have internalized organized biological knowledge, including pathway membership, protein interactions, functional modules, and hierarchical abstraction, yet they encode minimal causal regulatory logic. We release both feature atlases as interactive web platforms enabling exploration of more than 107000 features across 30 layers of two leading single-cell foundation models.

Ihor Kendiukhov

Department of Computer Science, University of Tübingen, Tübingen, Germany

Genomic Foundation ModelsSingle-Cell TranscriptomicsNeural Network InterpretabilityTransformer Interpretability
AskPDF
Mar 2, 2026

What Helps---and What Hurts: Bidirectional Explanations for Vision Transformers

Vision Transformers (ViTs) achieve strong performance in visual recognition, yet their decision-making remains difficult to interpret. We propose BiCAM, a bidirectional class activation mapping method that captures both supportive (positive) and suppressive (negative) contributions to model predictions. Unlike prior CAM-based approaches that discard negative signals, BiCAM preserves signed attributions to produce more complete and contrastive explanations. BiCAM further introduces a Positive-to-Negative Ratio (PNR) that summarizes attribution balance and enables lightweight detection of adversarial examples without retraining. Across ImageNet, VOC, and COCO, BiCAM improves localization and faithfulness while remaining computationally efficient. It generalizes to multiple ViT variants, including DeiT and Swin. These results suggest the importance of modeling both supportive and suppressive evidence for interpreting transformer-based vision models.

Qin Su, Tie Luo

Department of Electrical and Computer Engineering University of Kentucky, Lexington, KY, USA

Transformer InterpretabilityExplainable Image ClassificationFeature AttributionVision Transformer
AskPDF
Feb 27, 2026

GLUScope: A Tool for Analyzing GLU Neurons in Transformer Language Models

We present GLUScope, an open-source resource for analyzing neurons in Transformer-based language models (LMs), intended for interpretability researchers. We focus on the now-standard GLU-based activation functions such as SwiGLU, used in virtually all recent open-weights LMs. This introduces a new challenge: understanding positive activations is not enough. Instead, both the gate and the in activation of a neuron can be positive or negative, leading to four different possible sign combinations (activation quadrants) that in some cases have quite different functionalities (e.g. promoting vs. suppressing a specific set of tokens). Accordingly, for any neuron, our tool shows text examples for each of the four sign combinations, and indicates how often each combination occurs. Our resource can lead to novel insights: in particular, thanks to analyzing all four activation quadrants, it can surface interpretable behavior that other tools would miss. We also release the code used to create the resource, thus enabling researchers to build on it by adding more models, text corpora, or neuron pages.

Sebastian Gerstner, Leonor Veloso, Hinrich Schütze

LMU Munich and Munich Center for Machine Learning (MCML), Munich, Germany

Neuron-Level InterpretabilityTransformer InterpretabilityLLM InterpretabilityMechanistic InterpretabilityTransformer FFNs
AskPDF
Feb 26, 2026

MetaOthello: A Controlled Study of Multiple World Models in Transformers

Foundation models must handle multiple generative processes, yet mechanistic interpretability largely studies capabilities in isolation; it remains unclear how a single transformer organizes multiple, potentially conflicting "world models". Previous experiments on Othello playing neural-networks test world-model learning but focus on a single game with a single set of rules. We introduce MetaOthello, a controlled suite of Othello variants with shared syntax but different rules or tokenizations, and train small GPTs on mixed-variant data to study how multiple world models are organized in a shared representation space. We find that transformers trained on mixed-game data do not partition their capacity into isolated sub-models; instead, they converge on a mostly shared board-state representation that transfers causally across variants. Linear probes trained on one variant can intervene on another's internal state with effectiveness approaching that of matched probes. For isomorphic games with token remapping, representations are equivalent up to a single orthogonal rotation that generalizes across layers. When rules partially overlap, early layers maintain game-agnostic representations while a middle layer identifies game identity, and later layers specialize. MetaOthello offers a path toward understanding not just whether transformers learn world models, but how they organize many at once.

Aviral Chawla, Galen Hall, Juniper Lovato

Vermont Complex Systems Institute, University of Vermont, Burlington, VT · University of Michigan, Ann Arbor, MI

Transformer InterpretabilityWorld Model LearningTransformer
AskPDF
Feb 26, 2026

Transformers converge to invariant algorithmic cores

Training selects for behavior, not circuitry: many weight configurations can implement the same function. Studying any single trained neural network thus risks describing accidents of one training run rather than the computation itself. This work shifts focus from what transformers happen to do to what they must do by extracting algorithmic cores, compact subspaces that are necessary and sufficient for a task and that recur across independently trained models. Here, Algorithmic Core Extraction (ACE) is introduced to isolate these subspaces, causally validate them, and recover the algorithms they implement across settings ranging from synthetic tasks to large-scale pretrained models. Markov-chain transformers embed three-dimensional cores in nearly orthogonal subspaces yet recover identical transition spectra. Modular-addition transformers form compact cyclic cores at grokking that later inflate under continued regularization, redundantly distributing the same computation across many functionally equivalent modes. This functional redundancy is found to accelerate the transition from memorization to generalization, yielding an inverse scaling law for grokking time. In six language models spanning more than two orders of magnitude in scale (GPT-2 Small/Medium/Large, LLaMA-3.1, Gemma-2, and Qwen2.5), subject-verb agreement is governed by a single, steerable axis that aligns across architectures. Flipping this axis inverts grammatical number throughout open-ended generation. Together these results suggest that beneath the apparent complexity of trained transformers lies a simpler, shared computational structure, and that targeting invariants rather than parameterizations may offer a more tractable path to mechanistic understanding and control. Code: https://github.com/joshseth/cores

Joshua S. Schiffman

New York Genome Center, New York, NY, USA.

Transformer InterpretabilityInvariant Representation LearningGrokking
AskPDF
Feb 25, 2026

How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Processing

When a decoder-only transformer is forced to process matched correct and incorrect single-token continuations of a factual query, the two pathways through hidden-state space diverge: displacement vectors from the query-only representation keep near-equal magnitude but rotate apart, with angular separation growing through mid-depth before late layers resolve an asymmetric outcome. A logit-lens preference in the incorrect run falls far below the equal-probability prior (roughly 11.5x more mass on the incorrect token than the correct one). We read this pattern, rotational divergence then late-layer asymmetric commitment, as the geometric signature of the model externally appearing to reject a wrong continuation, while staying explicit that it is observational, not causal: the incorrect run could equally reflect the model conforming to the token it is forced to carry, which only a random-token control can settle. It holds across six decoder-only transformers spanning four architecture families (Llama, Mistral, Gemma, StableLM) from 1B to 13B parameters; a seventh (Qwen2 1.5B) is flat under our protocol, plausibly a tokenizer artefact, leaving an emergence threshold open. Linear probes recover the distinction at intermediate depth, and cross-domain transfer is structurally asymmetric, a financial-medical corridor transferring far better than transport pairs. Where single-layer activation patching is cleanly interpretable (LLaMA-2 13B, Mistral 7B) it yields no layer band of consistent recovery; a third model (StableLM-2 1.6B) recovers uniformly above the ceiling, which we diagnose as a code-path artefact and exclude. Under this scoped null the late-layer asymmetry is not localized to a single component, fitting a distributed-by-trajectory account rather than single-layer localized recall. We document this with forced-completion probing across seven models, three domains, and 300 queries.

Javier Marín

Groundlens-dev

Transformer InterpretabilityLLM InterpretabilityFactual Knowledge in Language ModelsRepresentation Geometry in Language ModelsLanguage Model Probing
AskPDF
Feb 24, 2026

Multi-Dimensional Spectral Geometry of Biological Knowledge in Single-Cell Transformer Representations

Single-cell foundation models such as scGPT learn high-dimensional gene representations, but what biological knowledge these representations encode remains unclear. We systematically decode the geometric structure of scGPT internal representations through 63 iterations of automated hypothesis screening (183 hypotheses tested), revealing that the model organizes genes into a structured biological coordinate system rather than an opaque feature space. The dominant spectral axis separates genes by subcellular localization, with secreted proteins at one pole and cytosolic proteins at the other. Intermediate transformer layers transiently encode mitochondrial and ER compartments in a sequence that mirrors the cellular secretory pathway. Orthogonal axes encode protein-protein interaction networks with graded fidelity to experimentally measured interaction strength (Spearman rho = 1.000 across n = 5 STRING confidence quintiles, p = 0.017). In a compact six-dimensional spectral subspace, the model distinguishes transcription factors from their target genes (AUROC = 0.744, all 12 layers significant). Early layers preserve which specific genes regulate which targets, while deeper layers compress this into a coarser regulator versus regulated distinction. Repression edges are geometrically more prominent than activation edges, and B-cell master regulators BATF and BACH2 show convergence toward the B-cell identity anchor PAX5 across transformer depth. Cell-type marker genes cluster with high fidelity (AUROC = 0.851). Residual-stream geometry encodes biological structure complementary to attention patterns. These results indicate that biological transformers learn an interpretable internal model of cellular organization, with implications for regulatory network inference, drug target prioritization, and model auditing.

Ihor Kendiukhov

Department of Computer Science University of Tübingen Tübingen, Germany

Transformer InterpretabilitySingle-Cell Transcriptomics
AskPDF
Feb 24, 2026

Beyond Accuracy: An Explainability-Driven Analysis of Harmful Content Detection

Although automated harmful content detection systems are frequently used to monitor online platforms, moderators and end users frequently cannot understand the logic underlying their predictions. While recent studies have focused on increasing classification accuracy, little focus has been placed on comprehending why neural models identify content as harmful, especially when it comes to borderline, contextual, and politically sensitive situations. In this work, a neural harmful content detection model trained on the Civil Comments dataset is analyzed explainability-drivenly. Two popular post-hoc explanation methods, Shapley Additive Explanations and Integrated Gradients, are used to analyze the behavior of a RoBERTa-based classifier in both correct predictions and systematic failure cases. Despite strong overall performance, with an area under the curve of 0.93 and an accuracy of 0.94, the analysis reveals limitations that are not observable from aggregate evaluation metrics alone. Integrated Gradients appear to extract more diffuse contextual attributions while Shapley Additive Explanations extract more focused attributions on explicit lexical cues. The consequent divergence in their outputs manifests in both false negatives and false positives. Qualitative case studies reveal recurring failure modes such as indirect toxicity, lexical over-attribution, or political discourse. The results suggest that explainable AI can foster human-in-the-loop moderation by exposing model uncertainty and increasing the interpretable rationale behind automated decisions. Most importantly, this work highlights the role of explainability as a transparency and diagnostic resource for online harmful content detection systems rather than as a performance-enhancing lever.

Trishita Dhara, Siddhesh Sheth

Upper Hand, Indianapolis, IN 46204, USA · Ace Rent a Car, Indianapolis, IN 46204, USA

Content ModerationNeural Network InterpretabilityTransformer Interpretability
AskPDF
Feb 22, 2026

Beyond Behavioural Trade-Offs: Mechanistic Tracing of Pain-Pleasure Decisions in an LLM

Prior behavioural work suggests that some LLMs alter choices when options are framed as causing pain or pleasure, and that such deviations can scale with stated intensity. To bridge behavioural evidence (what the model does) with mechanistic interpretability (what computations support it), we investigate how valence-related information is represented and where it is causally used inside a transformer. Using Gemma-2-9B-it and a minimalist decision task modelled on prior work, we (i) map representational availability with layer-wise linear probing across streams, (ii) test causal contribution with activation interventions (steering; patching/ablation), and (iii) quantify dose-response effects over an epsilon grid, reading out both the 2-3 logit margin and digit-pair-normalised choice probabilities. We find that (a) valence sign (pain vs. pleasure) is perfectly linearly separable across stream families from very early layers (L0-L1), while a lexical baseline retains substantial signal; (b) graded intensity is strongly decodable, with peaks in mid-to-late layers and especially in attention/MLP outputs, and decision alignment is highest slightly before the final token; (c) additive steering along a data-derived valence direction causally modulates the 2-3 margin at late sites, with the largest effects observed in late-layer attention outputs (attn_out L14); and (d) head-level patching/ablation suggests that these effects are distributed across multiple heads rather than concentrated in a single unit. Together, these results link behavioural sensitivity to identifiable internal representations and intervention-sensitive sites, providing concrete mechanistic targets for more stringent counterfactual tests and broader replication. This work supports a more evidence-driven (a) debate on AI sentience and welfare, and (b) governance when setting policy, auditing standards, and safety safeguards.

Francesca Bianco, Derek Shiller

Independent Researcher Future Impact Group Fellowship · Rethink Priorities Future Impact Group Mentor

LLM Decision-MakingEmotion Representation in Language ModelsTransformer Interpretability
AskPDF
Feb 19, 2026

The Anxiety of Influence: Bloom Filters in Transformer Attention Heads

Some transformer attention heads appear to function as membership testers, dedicating themselves to answering the question "has this token appeared before in the context?" We identify these heads across four language models (GPT-2 small, medium, and large; Pythia-160M) and show that they form a spectrum of membership-testing strategies. Two heads (L0H1 and L0H5 in GPT-2 small) function as high-precision membership filters with false positive rates of 0-4% even at 180 unique context tokens -- well above the dhead=64d_\text{head} = 64 bit capacity of a classical Bloom filter. A third head (L1H11) shows the classic Bloom filter capacity curve: its false positive rate follows the theoretical formula p≈(1−e−kn/m)kp \approx (1 - e^{-kn/m})^k with R2=1.0R^2 = 1.0 and fitted capacity m≈5m \approx 5 bits, saturating by n≈20n \approx 20 unique tokens. A fourth head initially identified as a Bloom filter (L3H0) was reclassified as a general prefix-attention head after confound controls revealed its apparent capacity curve was a sequence-length artifact. Together, the three genuine membership-testing heads form a multi-resolution system concentrated in early layers (0-1), taxonomically distinct from induction and previous-token heads, with false positive rates that decay monotonically with embedding distance -- consistent with distance-sensitive Bloom filters. These heads generalize broadly: they respond to any repeated token type, not just repeated names, with 43% higher generalization than duplicate-token-only heads. Ablation reveals these heads contribute to both repeated and novel token processing, indicating that membership testing coexists with broader computational roles. The reclassification of L3H0 through confound controls strengthens rather than weakens the case: the surviving heads withstand the scrutiny that eliminated a false positive in our own analysis.

Peter Balogh

Attention Head AnalysisAttention Head SpecializationTransformer Interpretability
AskPDF
Feb 18, 2026

A Structural Theory of Position Bias in Transformers

Transformer models systematically favor certain token positions, yet the architectural origins of this position bias remain poorly understood. This bias is closely connected to the Lost-in-the-Middle phenomenon, where models underutilize information placed in the middle of the context. We show that causal Transformer architectures induce a systematic position-bias prior that can help explain the influence profiles associated with Lost-in-the-Middle. To characterize this prior, we develop a structural theory based on cumulative attention rollout with layerwise residual mixing. At finite depth, we isolate the competing positional drifts induced by causal masking, residual mixing, and positional encoding. Their interaction produces broad, often U-shaped profiles. At infinite depth, prior work predicts inevitable collapse under attention-only theory and identifies the discrepancy with observed Transformer behavior as an open problem. Under the stated boundedness assumptions, our residual-aware analysis replaces inevitable collapse with a dichotomy governed by cumulative attention mixing. Empirically, across 30 model-dataset settings, our adaptive residual-aware rollout closely matches measured input-token influence and reduces the average profile discrepancy by approximately 85% relative to both the attention-only and fixed-residual rollout baselines.

Hanna Herasimchyk, Robin Labryga, Tomislav Prusina +1

University of Hamburg Hamburg, Germany

Lost-in-the-Middle EffectPosition BiasTransformer InterpretabilityAttention Residual Connections
AskPDF
Feb 18, 2026

Explainable AI: Context-Aware Layer-Wise Integrated Gradients for Explaining Transformer Models

Transformer models achieve state-of-the-art performance across domains and tasks, yet their deeply layered representations make their predictions difficult to interpret. Existing explainability methods rely on final-layer attributions, capture either local token-level attributions or global attention patterns without unification, and lack context-awareness of inter-token dependencies and structural components. They also fail to capture how relevance evolves across layers and how structural components shape decision-making. To address these limitations, we proposed the Context-Aware Layer-wise Integrated Gradients (CA-LIG) Framework, a unified hierarchical attribution framework that computes layer-wise Integrated Gradients within each Transformer block and fuses these token-level attributions with class-specific attention gradients. This integration yields signed, context-sensitive attribution maps that capture supportive and opposing evidence while tracing the hierarchical flow of relevance through the Transformer layers. We evaluate the CA-LIG Framework across diverse tasks, domains, and transformer model families, including sentiment analysis and long and multi-class document classification with BERT, hate speech detection in a low-resource language setting with XLM-R and AfroLM, and image classification with Masked Autoencoder vision Transformer model. Across all tasks and architectures, CA-LIG provides more faithful attributions, shows stronger sensitivity to contextual dependencies, and produces clearer, more semantically coherent visualizations than established explainability methods. These results indicate that CA-LIG provides a more comprehensive, context-aware, and reliable explanation of Transformer decision-making, advancing both the practical interpretability and conceptual understanding of deep neural models.

Melkamu Abay Mersha, Jugal Kalita

College of Engineering and Applied Science, University of Colorado Colorado Springs, Colorado Springs, 80918, CO, USA

Transformer InterpretabilityGradient-Based AttributionIntegrated GradientsExplainability Evaluation
AskPDF
Feb 17, 2026

Quantifying LLM Attention-Head Stability: Implications for Circuit Universality

In mechanistic interpretability, recent work scrutinizes transformer "circuits" - sparse, mono or multi layer sub computations, that may reflect human understandable functions. Yet, these network circuits are rarely acid-tested for their stability across different instances of the same deep learning architecture. Without this, it remains unclear whether reported circuits emerge universally across labs or turn out to be idiosyncratic to a particular estimation instance, potentially limiting confidence in safety-critical settings. Here, we systematically study stability across-refits in increasingly complex transformer language models of various sizes. We quantify, layer by layer, how similarly attention heads learn representations across independently initialized training runs. Our rigorous experiments show that (1) middle-layer heads are the least stable yet the most representationally distinct; (2) deeper models exhibit stronger mid-depth divergence; (3) unstable heads in deeper layers become more functionally important than their peers from the same layer; (4) applying weight decay optimization substantially improves attention-head stability across random model initializations; and (5) the residual stream is comparatively stable. Our findings establish the cross-instance robustness of circuits as an essential yet underappreciated prerequisite for scalable oversight, drawing contours around possible white-box monitorability of AI systems.

Karan Bali, Jack Stanley, Praneet Suresh +1

Mila - Quebec Artificial Intelligence Institute, Montreal, Canada · McGill University, Montreal, Canada

Transformer InterpretabilityAttention Head AnalysisCircuits in Language Models
AskPDF
Feb 16, 2026

Inner Loop Inference for Pretrained Transformers: Unlocking Latent Capabilities Without Training

Deep Learning architectures, and in particular Transformers, are conventionally viewed as a composition of layers. These layers are actually often obtained as the sum of two contributions: a residual path that copies the input and the output of a Transformer block. As a consequence, the inner representations (i.e. the input of these blocks) can be interpreted as iterative refinement of a propagated latent representation. Under this lens, many works suggest that the inner space is shared across layers, meaning that tokens can be decoded at early stages. Mechanistic interpretability even goes further by conjecturing that some layers act as refinement layers. Following this path, we propose inference-time inner looping, which prolongs refinement in pretrained off-the-shelf language models by repeatedly re-applying a selected block range. Across multiple benchmarks, inner looping yields modest but consistent accuracy improvements. Analyses of the resulting latent trajectories suggest more stable state evolution and continued semantic refinement. Overall, our results suggest that additional refinement can be obtained through simple test-time looping, extending computation in frozen pretrained models.

Jonathan Lys, Vincent Gripon, Bastien Pasdeloup +4

IMT Atlantique, Lab-STICC, UMR CNRS 6285, Brest, France · Sony Europe Ltd. Stuttgart Technology Center, EUREC, Germany

Transformer InferenceTest-Time ScalingTransformer Interpretability
AskPDF
Feb 9, 2026

Central Dogma Transformer II: An AI Microscope for Understanding Cellular Regulatory Mechanisms

Motivation: Interpretability is not optional in biology: understanding gene regulation requires models whose learned structure can be directly interrogated, not merely accurate predictors whose internals resist mapping onto regulatory relationships. We ask whether an architecture mirroring the central dogma yields attention and gradient maps that recover known regulatory elements and networks in inspectable form. Results: Central Dogma Transformer II (CDT-II) mirrors the central dogma in its architecture -- DNA self-attention, RNA self-attention, and DNA-to-RNA cross-attention -- requiring only genomic embeddings and raw per-cell expression. On K562 CRISPR interference (CRISPRi) data with five genes held out entirely, CDT-II predicts perturbation effects (per-gene mean r = 0.84), recovers the GFI1B regulatory network (6.6-fold enrichment, P = 3.5 x 10^-17), and concentrates cross-attention on ENCODE regulatory elements including CTCF sites (mean 7.67x across 28 target genes, P < 0.001). Gradient attribution predicts consequences of perturbing therapeutic targets (mean r = 0.82). For TFRC, target of the anti-TfR1 antibody PPMX-T003, it identifies erythrocyte-structure, iron-dependent DNA-synthesis and oxidative-stress genes, matching anemia and ferroptosis reported clinically and preclinically -- without clinical data as input. CDT-II acts as an AI microscope, surfacing clinically relevant regulatory structure from perturbation experiments alone. Availability: Source code is available at https://github.com/nobusama/CDT2. Pre-computed embeddings, training data, and model weights are available at https://huggingface.co/datasets/nobusama17/CDT2-data.

Nobuyuki Ota

Independent Researcher, Burlingame, CA, USA

Transformer InterpretabilitySingle-Cell Perturbation ModelingTransformer AttentionGene Regulatory Network InferenceGenomics
AskPDF
Feb 8, 2026

Patches of Nonlinearity: Instruction Vectors in Large Language Models

Despite the recent success of instruction-tuned language models and their ubiquitous usage, very little is known of how models process instructions internally. In this work, we address this gap from a mechanistic point of view by investigating how instruction-specific representations are constructed and utilized in different stages of post-training: Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). Via causal mediation, we identify that instruction representation is fairly localized in models. These representations, which we call Instruction Vectors (IVs), demonstrate a curious juxtaposition of linear separability along with non-linear causal interaction, broadly questioning the scope of the linear representation hypothesis commonplace in mechanistic interpretability. To disentangle the non-linear causal interaction, we propose a novel method to localize information processing in language models that is free from the implicit linear assumptions of patching-based techniques. We find that, conditioned on the task representations formed in the early layers, different information pathways are selected in the later layers to solve that task, i.e., IVs act as circuit selectors.

Irina Bigoulaeva, Jonas Rohweder, Subhabrata Dutta +1

Ubiquitous Knowledge Processing Lab (UKP Lab) · Department of Computer Science, Technical University of Darmstadt · National Research Center for Applied Cybersecurity ATHENE, Germany

Transformer InterpretabilityLLM InterpretabilityInstruction TuningCausal Interventions in Language Models
AskPDF
Feb 4, 2026

Hierarchical Latent Structures in Data Generation Process Unify Mechanistic Phenomena across Scale

Contemporary studies in mechanistic interpretability have uncovered many puzzling phenomena in the neural information processing of Transformer-based language models, such as induction heads, function vectors, and the Hydra effect. Some of these individual phenomena have been independently tied to different data distributional properties, while some have been loosely associated with model architecture and how Transformers process information. However, a unified understanding of the relationship between data, model architecture, and optimization remains lacking, failing to answer the fundamental question: why do these three phenomena appear universally across different model families and scales, despite their seeming disconnect? In this work, we answer this question by unifying these three phenomena as consequences of hierarchical latent structures in the data generation process, coupled with decorrelated gradients across additive model components and directional concavity in the representation geometry. We validate our theoretical results in a toy model regime and in a large-scale synthetic data regime, comparing them with language models trained on natural language data.

Jonas Rohweder, Subhabrata Dutta, Iryna Gurevych

UKP Lab, TU Darmstadt

Transformer InterpretabilityLatent Variable ModelsInduction HeadsFunction Vectors
AskPDF
Feb 2, 2026

Universal Redundancies in Time Series Foundation Models

Time Series Foundation Models (TSFMs) leverage extensive pretraining to accurately predict unseen time series during inference, without the need for task-specific fine-tuning. Through large-scale evaluations on standard benchmarks, we find that leading transformer-based TSFMs exhibit redundant components in their intermediate layers. We introduce a set of tools for mechanistic interpretability of TSFMs, including ablations of specific components and direct logit attribution on the residual stream. Our findings are consistent across several leading TSFMs with diverse architectures, and across a diverse set of real-world and synthetic time-series datasets. We discover that all models in our study are robust to ablations of entire layers. Furthermore, we develop a theoretical framework framing transformers as kernel regressors, motivating a purely intrinsic strategy for ablating heads based on the stable rank of the per-head projection matrices. Using this approach, we uncover the specific heads responsible for degenerate phenomena widely observed in TSFMs, such as parroting of motifs from the context and seasonality bias. Our study sheds light on the universal properties of this emerging class of architectures for continuous-time sequence modeling.

Anthony Bao, Venkata Hasith Vattikuti, Jeffrey Lai +1

ECE Department, UT Austin · Department of Physics, UT Austin · Oden Institute, UT Austin

Transformer InterpretabilityKernel RegressionMechanistic InterpretabilityTransformer PruningTime-Series Foundation ModelsNeural Network Pruning
AskPDF
Jan 9, 2026

The Grammar of Transformers: A Systematic Review of Interpretability Research on Syntactic Knowledge in Language Models

We present a systematic review of 337 articles evaluating the syntactic abilities of Transformer-based language models (TLMs), reporting on over 3,000 datapoints spanning a wide range of syntactic phenomena, languages, models, and methods. We take the data to collectively show that TLMs encode a non-trivial amount of syntactic knowledge. Behavioral evidence shows strong performance on formal syntactic phenomena, but weaker and more variable performance on phenomena at the syntax-semantics interface. Performance is also consistently lower for languages with less digital support. Probing and mechanistic studies further support the presence of syntactic knowledge in TLMs. Yet, because most work remains observational and methodologically heterogeneous, insight into the detailed computational mechanisms underlying syntactic processing remains limited. At the same time, the literature remains heavily concentrated on English and BERT-like models. We discuss the implications of our results and provide recommendations for future research.

Nora Graichen, Iria de-Dios-Flores, Gemma Boleda

Universitat Pompeu Fabra1 · ICREA2

Transformer InterpretabilityLLM Evaluation
AskPDF
Dec 23, 2025

Probing Spectrum-Like Organization of States of Mind in Transformer Representation Spaces

We investigate whether graded states of mind form spectrum-like structure in transformer representation spaces. To do so, we construct a dataset of 636 short natural-language sentences annotated with both a continuous score from −5-5 to 55 and one of seven ordered tiers, ranging from collapsed or scarcity-driven expressions to more coherent, reflective, and integrative ones. We evaluate five frozen transformer representations: four sentence-embedding models and one decoder-only residual-stream representation. Across all representations, simple probes reliably recover both the continuous score and the discrete tier labels, and permutation tests show that performance significantly exceeds shuffled-label baselines. Additional analyses reveal a consistent geometric pattern: UMAP projections show low-to-high organization, confusion matrices concentrate errors between neighboring tiers, and directional ablation identifies a prominent score-aligned component. These results suggest that transformer representations contain statistically significant, spectrum-like organization aligned with the annotated state-of-mind structure. The annotations are used only as an operational framework for representation analysis, not as a clinical or diagnostic measure.

Sophie Zhao

School of Computer Science Georgia Institute of Technology

Transformer InterpretabilityNeural Representation Geometry
AskPDF
Nov 22, 2025

Equivalence of Context and Parameter Updates in Modern Transformer Blocks

Recent research has established that the impact of context in a vanilla transformer can be represented implicitly by forming a token-dependent, rank-1 patch to its MLP weights. This work extends that foundational theory to the diverse architectures of modern Large Language Models. We first demonstrate a precise, analytical solution for a Gemma-style transformer block, proving that the entire effect of a context can be perfectly mapped to rank-1 patches on its MLP weight matrices and a patch to the RMSNorm scale. We then generalize this result, providing a constructive proof and algorithm for multi-layer models. To unify these findings, we introduce a general framework centered on two core properties: input controllability and output controllability. We prove that a perfect implicit weight patch is possible for any MLP block where the inner function is input-controllable and the outer function is output-controllable. This provides a simpler and more powerful lens for understanding how transformer models transmute prompts into effective weights. This setup generalizes to a wide range of modern LLM architectures including gating, pre-/post-norm, mixture of experts and sequential/parallel transformer blocks.

Adrian Goldwaser, Michael Munn, Javier Gonzalvo +1

University of Cambridge, UK · Google Research

Transformer InterpretabilityTransformer FFNsTransformerIn-Context Learning
AskPDF