Mechanistic Interpretability

Model organisms (MOs) - language models trained to exhibit undesired or unnatural behaviours - are frequently used as testbeds for evaluating white-box interpretability techniques. Current MOs are typically constructed via post-hoc supervised fine-tuning (SFT) on behavioural transcripts or synthetic documents. Prior research has shown that interpretability methods can easily identify hidden behaviours in these MOs. However, recent work suggests that such post-hoc training methods may make interpretability unrealistically easy. We investigate this claim by constructing a suite of 54 $\verb|OLMo2-1B|$- and $\verb|gemma-3-1b-it|$-based MOs trained with seven different techniques, including standard post-hoc SFT, post-hoc DPO, and more realistic integration of MO data into the OLMo post-training DPO phase. We use these MO variants to benchmark activation oracles, activation steering, logit lens, and sparse autoencoders. Our findings show that (i) MO interpretability depends strongly on training objective, target behaviour, model architecture, and training data generation pipeline; (ii) substantial variance remains even after controlling for differences in the strength of target behaviour expression; and (iii) our more realistic $\textit{integrated training}$ often yields less interpretable MOs than standard post-hoc methods. Our results cast substantial doubt on the validity of current MOs as interpretability proxies.

In light of strides in Arti cial Intelligence (AI) and its wide spread application, challenges persist in the interpretability of AI models, particularly within specialized domains like healthcare, such as electro cardiograph (ECG) recognition. Rather than relying solely on end-to-end convolutional neural networks, this paper introduces a novel approach using a domain knowledge-based graph convolution network for ECG recognition. Key landmarks points of PRQST, vital to ECG interpreta tion, are incorporated as domain knowledge. The double-stream directed graph is employed to model both intra and inter ECG cycles. Speci cally, spatial directed graphs capture the positional relationships among key points, while temporal directed graphs delineate temporal dependencies between adjacent cycles in extended ECG sequences. Experimental re sults on the First Chinese ECG Intelligent Competition dataset, which speci cally classify ECG into nine categories, prove the e cacy of the proposed model. The overall average F1 score is 88.1%, the average F1 score of rare categories is 76.3%, both outperform the state-of-the-art models. The introduction of domain knowledge did enhance the detec tion performance, especially for rare categories.

Concept Bottleneck Models (CBMs) provide interpretability by routing predictions through a layer of human-understandable concepts. However, defining an optimal concept set for a specific dataset remains an open challenge. Existing approaches rely on expensive expert annotations or LLM-generated lists based solely on class names. Even "open-vocabulary" variants typically depend on static concept sets, which restrict discovery and introduce label bias. Furthermore, traditional CBMs often suffer from information leakage, where unmodeled visual features bypass the bottleneck and compromise the integrity of the explanations. To overcome these limitations, we propose Caption Bottleneck Models (CaBM), a framework that circumvents the need for predefined concept sets by replacing rigid concept layers with free-form natural language. By representing images via LMM-generated captions and training a classifier strictly on this text, CaBM ensures a leakage-free architecture by construction. Additionally, by analyzing the text classifier post-training, CaBM autonomously discovers high-quality, dataset-specific concepts. Our results across fine- and coarse-grained benchmarks demonstrate that CaBM achieves competitive accuracy while preserving interpretability without the constraints of external dictionaries or manual labeling.

The research proposes a multilayer Q-matrix-embedded neural network for cognitive diagnosis (M-QCDNet), which integrates the structural interpretability of cognitive diagnostic models (CDMs) with the deep learning neural network (NN). M-QCDNet structures the item-skill relationship using the Q-matrix as a structural prior, ensuring latent mastery profiles remain interpretable and consistent with cognitive theory, followed by the proposed loss function with an L2 penalty to penalize skills not aligned with the Q-matrix and to balance predictive performance and structural alignment. Corresponding evaluation matrices, the interpretable alignment-based metrics that quantify the degree to which predicted skill activations correspond to item-level skills, were further developed. M-QCDNet offers practical benefits for classroom practice, enabling early detection of learning difficulties and supporting mastery-based interventions. By embedding diagnostic validity into model design, M-QCDNet bridges psychometric transparency and neural flexibility, advancing interpretable, fair, and actionable AI for cognitive diagnostics.

Mechanistic interpretability has produced a rich inventory of component-level analyses that characterise what neural-network components encode and how they interact. Their outputs, however, are not easily reusable: selectivity tables, circuit diagrams, and feature lists remain locked in per-study notebooks - non-composable, not queryable in natural language, and not directly actionable for downstream audit or intervention. We study the representation layer that sits between these analyses and downstream use as a bottleneck that can be evaluated independently, and introduce Manifestation Units, a typed tuple protocol (E, S, R, D, G) extended with attention-head primitives (T) for transformer architectures, organising per-component statistics into structured fields populated automatically and queried through hybrid retrieval. Instantiated across generative vision (beta-VAE), discriminative vision (CNN), and language (GPT-2), the protocol supports two findings: typed structure substantially outperforms unstructured baselines on retrieval, and CNN filters retrieved by the schema satisfy causal sufficiency and necessity criteria under matched-budget controls. The schema absorbs attention-head primitives without modification, set-recovers known IOI circuit members under retrieval-budget-matched controls, and reveals an irreducible two-field core (S+R) with remaining fields either redundant or actively interfering. We present this as schema infrastructure for mechanistic interpretability rather than frontier-scale validation.

Mechanistic interpretability (MI) requires full access to model internals, yet the APIs for most widely deployed language models at best expose log-probabilities over output tokens. This creates a surrogate problem: when do measurements made on open models allow us to make claims about a closed model? We evaluate surrogate fidelity at the prediction, attribution, and representation levels. For binary classification tasks, log-odds provide an API-compatible scalar readout of the model's representation space, and leave-one-out attributions provide insight into model behavior. Across eleven models spanning four families (Llama, Qwen, GPT, and Gemini), we find that prediction fidelity substantially overstates attribution fidelity: models that agree on what the answer is often disagree on why. We document an access-validity inversion: white-box signals like attention patterns and perturbation magnitudes are highly stable across models but only weakly predictive of causal attributions, which black-box input ablations capture by design. Mechanistic insight does not automatically transfer to closed targets, and prediction-level agreement is insufficient to warrant such transfer. Code and results are available at https://github.com/facebookresearch/surrogate.

Artificial intelligence is transforming our capability to solve biological challenges. In dimensionality bottleneck regimes exacerbated by high-dimensional biological data, neural networks force distinct concepts into the lower dimensions known as superposition. Although this superposition is widely known to hinder interpretability, its impact on corrupting the geometry of latent spaces remains critically overlooked. Here, we utilized sparse autoencoders (SAEs) trained on over 100,000 multiplexed images of patient-derived Parkinson's disease and healthy neurons to resolve superposition. This approach bypasses the mathematical non-uniqueness of feature attribution by shifting to interpretable latent representation analysis. We theoretically and empirically demonstrate that superposition contaminates representational metric spaces, and thereby SAEs successfully recover geometric fidelity. By treating these geometrically purified representations as single-cell state vectors, we adapted single-cell RNA sequencing (scRNA-seq) data analysis methodologies directly to the image domain. Finally, we introduce GW-map, utilizing Gromov-Wasserstein optimal transport to align these image representations with authentic scRNA-seq data de novo. This coupling reconstructs hierarchical neuronal pathology pathways such as Calcium-AIS scaffold, without reference spatial transcriptomics, establishing a scalable foundation for spatial biology. Code is available at https://github.com/jijihihi/Bio\_superposition

Recent work has shown that well-optimized individual decision trees can match complex black box models in some settings, primarily in noisy domains. For the remaining settings, however, complex ensembled compositions of trees often achieve higher accuracy at the cost of interpretability, leaving practitioners with difficult modeling decisions along an accuracy-interpretability tradeoff. Ideally, we would like to classify as much of the data as possible with one or a small number of trees, achieving interpretability for most samples while maintaining state-of-the-art accuracy. We introduce Multistage Defer Trees: a sequence of sparse decision trees that each make predictions for most samples, while deferring a small proportion to the next tree in the sequence or, ultimately, to a black box. We demonstrate that we can train this model class to match the performance of complex tree-based ensembles while routing most samples through only one or a small number of sparse decision trees. We discuss a range of techniques for training these models while maintaining simplicity. Our method expands the accuracy--interpretability frontier in settings where single-tree methods remain insufficient, demonstrating that even when complex models are necessary, they need not be fully opaque.

Explainability in Many-Objective Optimization (MaO) is currently hindered by the escalating complexity of the Pareto front, which renders the relationship between high-dimensional decision variables and objective outcomes increasingly opaque. As the number of objectives exceeds the limits of traditional visualization, decision-makers encounter a cognitive drought'' in identifying relevant trade-offs or specifying target regions without a priori knowledge. To bridge this interpretability gap, we introduce the {Partition-Guided Distance Saliency (PGDS)} framework, a novel XAI approach designed for continuous optimization landscapes. Our framework automates the explanation process through a three-stage pipeline that prioritizes geometric intuition over abstract rules. First, we employ a surrogate model that learns how geometric distances in the decision space map to proximity in the objective space. Second, to address the difficulty of manual target selection in high dimensions, the framework automatically partitions the objective landscape into distinct regions and identifies local Dominating Points'' to serve as automated targets for improvement. Third, we quantify how sensitive a solution's position is to each decision variable by measuring the distance shifts induced by perturbations to each variable. This allows PGDS to categorize features as either Drivers'' which facilitate convergence toward preferred regions, or Blockers'' which represent geometric constraints hindering further progress. Validation on 10-objective benchmarks and a physics-informed engineering problem (Welded Beam) demonstrates that PGDS provides differentiated, actionable insights that traditional visualization and rule-based XAI methods fail to provide.

Interpretable Mesomorphic Neural Networks (IMNs) offer a promising framework that combines the predictive power of deep neural networks with the interpretability of linear models. However, the original formulation lacks safeguards to ensure that the learned interpretations are in fact reliable. In particular, the network is free to concentrate all explanatory variance into a single weight of the linear output layer, achieving strong predictive performance while producing interpretations that are largely meaningless. Paradoxically, the L1 penalty proposed to encourage sparse solutions exacerbates this problem by further incentivizing such degenerate configurations. To address this vulnerability, we introduce Local Fidelity Regularization (LFR), a novel penalty term that prevents degenerate weight collapse by aligning the linear output weights with local data variations. This structural constraint guarantees faithful explanations and substantially improves the reliability of model interpretations. Furthermore, empirical evaluations across the OpenML benchmark suite demonstrate that LFR does not compromise accuracy for explainability; rather, it achieved improved AUROC over the unregularized IMN. By yielding results highly competitive with state-of-the-art black-box models, LFR provides the dual benefit of reliable interpretability and superior predictive performance. Source code and usage instructions are available at https://github.com/hugohammer/LFR-IMN.git.

Earth system infrastructures relying on satellite-based technologies, such as Global Positioning System (GPS) communications, are affected by ionospheric Total Electron Content (TEC) gradients. Modeling these gradients under physical constraints remains challenging due to their dynamic and transient nature. While existing machine learning (ML) models can predict hourly TEC variations, it remains unclear whether their temporal resolution is sufficient to preserve small-scale TEC irregularities within predicted signals. To address this gap, we introduce an interpretable ML-based model, t-STEP, designed to predict TEC at a 30-second resolution and estimate irregularity signatures from the modeled signals. This high cadence enables the derivation of Rate of TEC changes (ROT) and the ROT Index (ROTI) as diagnostic indicators of ionospheric variability. The model is developed using GPS observations from solar cycle 24 at a station located at 5.49°S, 47.49°W. A multi-metric evaluation framework, including dynamic time warping, is used for robustness assessment, while SHAP (SHapley Additive exPlanations) provides insight into feature contributions. The 30-second TEC predictions achieve 91% accuracy with a mean absolute error (MAE) of 4.38 TECU during high solar activity (2015). Compared with the International Reference Ionosphere (IRI-2020), the hourly model improves accuracy by 35%, reduces absolute errors by 57%, and increases prediction skill by 54%. More importantly, the 30-second model captures TEC irregularity dynamics and morphologies during geomagnetic storms of different intensities, outperforming an attention-based Long Short-Term Memory model under the same experimental conditions. This study demonstrates the potential of a single TEC prediction framework for scalable irregularity monitoring without requiring separate models for individual transient events.

While existing data attribution methods can identify which training examples build specific mechanistic circuits, they cannot explain how training data shapes the high-level behavioral decisions a model learns to make. To bridge this gap, we introduce Symbolic Mechanistic Data Attribution (SMDA), a framework that attributes training pairs to the interpretable symbolic policies governing model behavior. SMDA fits a closed-form Ridge regression over sparse autoencoder (SAE) features to model a target behavior, then analytically decomposes how each supervised fine-tuning example shifts that policy through feature-activation Delta_X and output-probability Delta_Y pathways. We distill a symbolic policy for refusal behavior in Llama-3.2-3B-Instruct and analyze 200 SFT training pairs. Our analysis reveals that (1) the symbolic policy's coefficients expose systematic gaps in the base model's safety behavior for categories like religious stereotyping; (2) per-feature Delta_X/Delta_Y decomposition can mechanistically explain why harmful and harmless pairs exert qualitatively different influences on certain features; and (3) individual training pairs routinely exhibit cross-feature interference, allowing SMDA to identify training pairs whose dominant effect falls on unintended features. These results demonstrate that combining mechanistic interpretability with data attribution yields a diagnostic tool that is both more fine-grained than black-box influence functions and more scalable than manual circuit analysis.

Large Language Models (LLMs) have demonstrated the ability to simulate human-like OCEAN personality traits in generated text. Previous efforts have focused on prompt engineering or fine-tuning to shape LLM personality. In this work, we propose a mechanistic interpretability approach that directly intervenes on the model's latent features. Our method identifies latent directions in the residual stream corresponding to a target OCEAN trait using sparse autoencoders (SAEs) and contrastive activation analysis. We formalize an additive steering vector in activation space and demonstrate how applying a small additive shift to the hidden states enhances the target trait while preserving overall language modeling performance. To determine the optimal combination of feature shifts, we explore a linear weighting heuristic with grid search optimization that balances personality expression with task performance. Our approach shows promise in controllably steering personality traits at the mechanistic level while maintaining high performance on standard benchmarks.

While interpretable models such as concept bottleneck models (CBMs) and program synthesis methods enable verification of model decisions, their evaluation is typically limited to simple tasks, leaving complex reasoning on real-world images largely unexplored. We introduce COCOLogic-V2, an object-centric dataset for visual inductive reasoning on real-world images covering a broad subset of first-order logic. By categorizing samples into positive variants, near-boundary (NB), and far-from-boundary (FB) negatives, COCOLogic-V2 enables fine-grained diagnosis of model accountability. Our evaluations show that models tend to separate positive and FB samples well but fail on NB samples, while perceptual noise and large rule-induced search spaces pose additional challenges in few-shot settings. Together, these results highlight that visual inductive reasoning remains an open challenge and COCOLogic-V2 provides a concrete foundation for advancing methods in this direction.

Wearable devices produce large, high dimensional training logs for everyday runners, and interpretation rather than data collection is now the limiting step. This paper evaluates five dimensionality reduction models, three autoencoder variants, PCA, and a Variational Autoencoder, on their ability to compress nine sensor runner profiles into a single scalar performance indicator, the latent score. Because the setting is fully unsupervised, model quality is assessed along two complementary axes: reconstruction error (Mean Squared Error) and latent score interpretability, measured via Spearman and Kendall rank correlations, Mutual Information, and Permutation Importance. These are combined into a composite selection criterion that prevents selecting models on reconstruction accuracy alone. Feature rankings from the four metrics are aggregated via a modified Borda count, and their stability is confirmed by bootstrap validation. A two feature linear baseline is included to anchor the comparison. Deep autoencoder achieved the lowest reconstruction error and the highest composite score. Once the PCA hidden layers were widened, the deeper variants became closely competitive with Deep AE on the composite criterion, indicating that the limiting factor was hidden layer capacity rather than the one dimensional bottleneck. Running pace, aerobic decoupling, and average heart rate emerged as the dominant latent score drivers across all models and resampling runs, consistent with established physiology.

LLM-as-a-Judge and self-evaluation pipelines implicitly assume that evaluation is easier than generation. We test this in a controlled in-context QA setting where a context passage is the sole information source and each model judges the answer it generated, removing the parametric-knowledge confound of open-domain comparisons. Across four benchmarks (SQuAD 2.0, DROP, HotpotQA, MuSiQue) and two models, evaluation is not uniformly easier: generation accuracy exceeds self-evaluation on three of four, with multi-hop MuSiQue the exception. Attention analysis reveals why: evaluation attends to context 3--5x less than generation does and barely reads the candidate answer. LoRA fine-tuning confirms the asymmetry is not a training artifact: generation fine-tuning induces over-acceptance and evaluation fine-tuning degrades generation. These findings challenge core assumptions in self-evaluation pipelines.

Activation patching is the primary tool in mechanistic interpretability. It attributes causal responsibility for a model behavior to each of its individual components by estimating its natural indirect effect (NIE). Re-deriving the activation patching estimand from causal mediation analysis, we find that the NIE does not solely capture the causal effect through the specific component. It also contains interaction effects (INT) that measure how much the component's causal effect itself depends on the state of other components in the model. A natural response may be to try to eliminate INT by adjusting the estimator or unit of analysis, but each of these potential remedies has predictable failure modes. We demonstrate these failure modes in the GPT-2 IOI circuit; components whose causal importance is conditional on the state of other components are either invisible or artificially inflated, and INT variance explains the previously documented instability of faithfulness scores. We prove that INT scales with the distance between clean and patched component activations, is negligible when the model is locally affine, and decomposes combinatorially into pairwise and higher-order group interactions. Despite its inevitability, INT is not a nuisance to be eliminated, but rather a diagnostic for interpretability studies. Its individual and group-level magnitude and sign signal when causal conclusions are prompt-dependent, and when greedy NIE-based component ranking will miss mechanisms only discoverable through combinatorial search.

Foundation models for structural biology have achieved remarkable performance in predicting biomolecular structure and show promise for the design of proteins and small molecules. Yet understanding which internal features drive their outputs remains challenging. Standard sparse autoencoders (SAEs), effective on transformer-style sequence embeddings, do not transfer cleanly to pairformer-like architectures: naively operating on pairwise representations yields a quadratic blow-up of features and obscures concepts distributed jointly across sequence and pair representations. We introduce PairSAE, which summarizes pairwise tensors via an N-mode SVD into token-wise interaction roles, then uses a sparse autoencoder to learn a shared set of token-level features that decode into both sequence and pair representations. Evaluated on Boltz-2 activations for PLINDER protein-ligand complexes, PairSAE yields interpretable features that align with UniProt annotations and predict Boltz-2 affinity values. These results indicate that PairSAE links the latent space of foundation models for structural biology to interpretable structural concepts, clarifying what the model "knows" while avoiding pairformer-induced pitfalls that limit conventional SAEs.

Language models (LMs) capture large amounts of factual knowledge applicable to a wide range of tasks, motivating the view of their parameters as a knowledge base. An important property of knowledge bases is that different queries for the same fact return consistent results, drawing on a single source of truth. We investigate whether LMs satisfy this property through behavioral and mechanistic analyses. Our results suggest that they encode knowledge in a task-specific manner. Behaviorally, facts acquired on one task frequently fail to co-emerge on others during training. Parameter localization experiments suggest a mechanistic explanation, revealing distinct parameter subsets underlying different tasks for the same fact. Finally, we show that chain-of-thought reasoning draws part of its effectiveness from engaging task-specific parameters beyond those tied to the evaluation task. Our findings suggest that what the model knows and how it is asked are intertwined in parameter space, undermining the "knowledge base" analogy and carrying implications for the reliability and controllability of factual knowledge in LMs.

Edible insects offer an efficient source of alternative protein, requiring less land, water and emitting less greenhouse gas than conventional livestock. However, their successful integration into the food supply chain demands reliable species authentication to control allergen exposure, prevent adulteration, and meet regulatory standards. Near-infrared spectroscopy provides a rapid analytical tool, but its performance drops when applied to production batches unseen during training due to batch-to-batch variation in spectral measurements. We introduce the Batch-Invariant Spectral Network (BISN), an end-to-end framework that combines a learnable preprocessing module, initialised with Savitzky-Golay filtering, with an entropy-regularised adversarial objective to suppress batch-specific spectral variation. In contrast to Domain-Adversarial Neural Networks, which enforce domain adaptation only after feature extraction, BISN suppress batch-effects before species-specific features are learned. Using 2,700 spectra from three species (Acheta domesticus, Hermetia illucens, and Tenebrio molitor) collected across three independent production batches, BISN achieves a mean leave-one-batch-out accuracy of 0.93 (standard deviation 0.04), outperforming the strongest baseline by four percent. Further insights gained by using explainable AI confirm that model decisions consistently rely on the lipid and protein absorption regions across all folds, connecting predictive performance to known insect biochemistry. BISN addresses both cross-batch robustness and biochemical interpretability for automated insect species authentication under realistic industrial conditions. The source code and dataset are publicly available at https://github.com/majharB/bisn.

We develop a framework for interpreting AI systems as agents, drawing on the philosophical tradition of radical interpretation and the tools of mechanistic interpretability. The core question is: given the computational facts about a system, how do we solve for its beliefs, desires, and meanings? This matters increasingly for safety. We want to be able to trust the systems we deploy, whether by understanding their goals or, more modestly, by reliably detecting deception. Interpretability researchers are building tools to read beliefs and desires off a model's internals, but there is no settled account of when such a tool has succeeded. This book supplies one. We propose criteria on both representationalist and interpretationist approaches, and tie each to tests current interpretability methods can carry out. A central lesson is that these attributions cannot be made piecemeal. Beliefs, desires, and the propositional structure they presuppose are jointly constrained, and a method that fixes one while measuring the others inherits whatever distortions that introduces. This holism becomes pressing for AI systems, which may not share the interpreter's concepts. However, it also provides leverage: a system's attitudes constrain its propositional structure, that structure constrains which attitudes can be attributed, and mechanistic interpretability can help us measure both.

We review the concepts of interpretability and explainability as they apply to machine learning in physics. We define interpretability as concerning the structural transparency of a model (the ability to understand or approximate its inner workings) and explainability as concerning the scientific content of a model (the ability to map it onto domain knowledge). We discuss the trade-offs each entails (interpretability vs. expressivity; explainability vs. adaptability), the contexts in which each is needed, and the intrinsic and post-hoc tools available for achieving them. Throughout, we emphasize that machine-learned models are subject to the same scientific questions as classical models, differing only in scale, and that interpretability and explainability are best understood as deliberate modeling choices rather than inherent properties. We also emphasize the importance of task specification and intervention plans as a core aspect of model design.

Increasing demand for precise and reliable control in complex scenarios has led to the development of increasingly sophisticated controllers, including data-driven approaches employing closed box models and mathematically rigorous yet complex designs. This complexity highlights the needs for explainable control that can provide human-understandable insights into controller behavior. In this paper, an explainable control framework (XCF) along with supporting algorithms and user interface are proposed to explain how controllers determine their control actions and their underlying working mechanism. The novel contributions of this work are threefold: First, the XCF is designed to provide model-agnostic explanations for controllers in closed-loop systems and can optionally refine local explanations by system response dynamics. Second, a novel explanation method, hierarchical fuzzy model-agnostic explanation for control systems (HFMAE-C), is proposed based on the designed framework. The HFMAE-C employs a fuzzy logic system to approximate the controller's behavior and system dynamics, providing sample, local, domain and universe level explanations via IF-THEN rules revealing the controller's decision logic and salience values quantifying the contribution of system states to control actions. Third, a large language model agent-supported user interface is developed to automatically analyze user requirements, select appropriate algorithms, interpret the generated explanations to a natural language report, and provide interactive consultation. Case studies on inverted pendulum system and Turtlebot obstacle avoidance demonstrate the effectiveness of the proposed method through simulated user experiments and quantitative comparisons with mainstream explainable control approaches.

Given the widespread prevalence of depression and its consequential impact on individuals and society, it is crucial to obtain objective measures for early diagnosis and intervention. As a multidisciplinary topic, these objective measures should be interpretable and accessible to health care professionals, ensuring effective collaboration and treatment planning in the realm of mental health care. Even though current automated depression diagnosis approaches improved over the last decade, a critical gap exists as they often lack affect-specificity and interpretability, limiting their practical application and potential impact on mental health care. In particular, interpretability from temporal activities from videos when deep models are used is not fully explored. In this study, we present a novel framework for analyzing Deep Neural Networks' decisions when trained on facial videos, specifically focusing on automatic depression severity diagnosis. By fine-tuning Deep Convolutional Neural Networks (DCNN) pre-trained on Action Recognition datasets on depression severity facial videos from AVEC depression dataset, our framework is able to interpret the model's saliency maps by examining face regions and temporal expression semantics. Our approach generates both visual and quantitative explanations for the model's decisions, providing greater insight into its reasoning. In addition to this interpretability, our video-based modeling has improved upon previous single-face benchmarks for visual depression diagnosis, resulting in enhanced predictive performance. Overall, our work demonstrates the successful development of a framework capable of generating hypotheses from a facial model's decisions while simultaneously improving depression's predictive capabilities.

We address the problem of explainability in machine learning models through feature attribution methods. In particular, we consider a variant of Shapley values known as Asymmetric Shapley Values (ASV), which enables the incorporation of causal knowledge into model-agnostic explanations through the use of a causal graph. We show that in certain contexts in which the computation of SHAP is $\#P$-hard, the exact computation of ASV can be done in polynomial time. To extend this algorithmic result, we introduce a notion of equivalence classes over the topological orderings of the underlying causal graph, which is useful to reduce the time to compute ASV. In particular, we present a polynomial-time algorithm (in the number of equivalence classes) to compute it whenever the causal graph is a rooted directed tree. Finally, we develop an algorithm for approximating ASV in arbitrary causal DAGs which relies on a procedure to sample topological orderings uniformly at random. To implement this sampling mechanism we leverage known algorithms as well as simpler alternatives. Our experimental results demonstrate the practical viability of the proposed approach in realistic causal structures.

Sparse autoencoders (SAEs) are increasingly used to extract interpretable concepts from vision and vision language models, yet existing evaluation methods largely rely on proxy metrics or qualitative inspection rather than measuring semantic correspondence. We present a human-grounded evaluation framework that quantifies alignment between SAE latents and human-annotated concepts, without requiring user studies, and validate this matching through targeted attribute perturbations. To enable this intervention-style evaluation in vision, we construct synCUB and synCOCO, synthetic benchmarks of paired images that differ in exactly one attribute. We introduce Fully-Binary Matching Pursuit (FBMP), a coalition-based matching procedure that supports many-to-one mappings between SAE latents and annotated concepts, and consistently outperforms one-to-one baselines. For functional validation, we propose a Targeted Attribute Perturbation Alignment Score (TAPAScore), which tests whether matched concepts respond selectively and in the expected direction under targeted image-level attribute perturbations. Under sanity checks, our matching and TAPAScore are the only evaluated metrics that reliably distinguish trained SAEs from untrained ones. Across SAEs trained on CLIP and DINOv2 embeddings, we find that increased overcompleteness can reduce perturbation alignment, indicating a reduction in interpretability. Our evaluation framework suggests that moderate dictionary sizes provide the best trade-off, yielding the most interpretable SAEs. Code and datasets are available at https://github.com/JonasKlotz/sae-concept-eval.

Pruning Large Language Models (LLMs) reduces memory and inference costs by removing parts of the network, producing smaller models that retain most of their accuracy. As attention layers are the most resource-intensive parts of LLMs, pruning them is a promising compression strategy. Prior work shows that up to 33% of attention layers can be pruned with minimal accuracy loss. Nevertheless, the impact of attention pruning on model interpretability, specifically faithfulness and confidence calibration, remains unstudied. To address this gap, we study how pruning attention layers affects explanation faithfulness and confidence calibration across five LLMs and eight datasets. While the pruned models often maintain high accuracy, we find that their faithfulness and calibration often degrade. Notably, faithfulness and calibration can fluctuate significantly, even when accuracy remains stable, highlighting a misalignment between model confidence, interpretability, and accuracy. Our findings suggest that layer pruning can affect LLMs' interpretability and reliability in ways not captured by accuracy and efficiency measures alone. We recommend including explainability and calibration metrics when evaluating pruned models.

Understanding the features of large language models (LLMs) is a central goal of interpretability. LLMs are commonly assumed to use superposition to represent more features than they have dimensions. They may not only represent features in superposition but also perform computation in superposition. Theory predicts that computing in superposition requires error correction that privileges feature directions over generic ones, but this prediction has not been tested empirically. We propose an empirical test of error correction in LLMs based on activation perturbations. Perturbing residual-stream activations, we find that they are robust to small perturbations--forming activation plateaus consistent with error correction--but less robust along candidate feature directions ("pure" directions, constructed from contrastive prompt pairs) than along mixtures of two such directions, indicating that the pure directions are privileged. We quantify this privilegedness by modeling the perturbation effect as a function of the $L^p$-norm of its decomposition into feature components. For $p=2$ the response is a quadratic form with at most as many nonzero eigenvalues as the residual-stream dimension, which cannot privilege the many feature directions superposition requires. $p>2$ lifts this constraint and is consistent with feature-specific error correction. We find $p>2$ for contrastive, MELBO, and SAE-decoder directions, and $p\approx2$ for random and PCA directions (controls). These results replicate across Gemma-2-9B, Qwen3-1.7B, Llama-3.1-8B, Mistral-7B-v0.3, Aya-Expanse-8B, and Yi-1.5-9B. We further validate our method on a toy model of error correction with known ground-truth features, recovering $p>2$ for true feature directions, degrading toward $2$ as we rotate away from them.

A central aspiration of mechanistic interpretability is controllability: if we know where a behavior is represented in a model's activations, we should be able to modify it. This rests on a hidden premise -- that the direction which detects a behavior and the direction which controls it are the same, or close. We test this geometrically: what is the angle between the direction that best detects a behavior and the one that best causes it? If detection implies control the cosine is near 1; otherwise it quantifies a detection-intervention gap. On Gemma 2-2B-it, output format (clean JSON vs markdown fencing) collapses both roles onto one axis. Hallucination does not: the model detects fake entities with perfect linear separability (AUC = 1.000 from layer 5), yet that direction sits at cos = 0.12 (about 83 degrees) from the direction producing a refusal -- a small, reproducible alignment, far from the cos = 1 that "detection is control" would require. A detector built from activations, with no chosen tokens, likewise fails to align (cos = -0.06). The gap generalizes: across four models from three families and two scales (1B-9B), cos stays in [0.12, 0.20], identical before and after instruction tuning (0.1197 vs 0.1200), placing its origin in pretraining. A 15-degree rotation toward the refusal direction partially bridges it -- 73% and 60% refusal on two held-out fake-entity categories at 1.8% false positives. We then ask whether this cosine predicts steerability, and it does not: detection is a high-dimensional class, not a single direction, and what separates the steerable case is functional, not readable from a static angle. The cosine is a weight-computable signature of the dissociation between knowing and steering, not a predictor of it.

This work investigates the interpretability of a Wav2Vec 2.0based speech intelligibility assessment model for oral and oropharyngeal cancer patients through canonical correlation analysis. By measuring the correlation between the model embeddings and eGeMAPS low-level descriptors (LLDs) as an interpretable reference, we analyze how acoustic information is encoded across the model layers. The analysis is conducted at two levels: individual LLDs layer-wise, and group-level: prosodic, spectral, and voice quality. Results show that the learned representations are most strongly correlated with spectral and prosodic features, with the first MFCC coefficient yielding the highest correlations across all layers. At the group level, spectral and prosodic groups achieve correlations of 0.77 and 0.71 respectively, while voice quality reaches 0.65. Beyond model interpretability, this work also offers practical guidance on acoustic feature selection for pathological speech assessment.

More than a billion people communicate in Indic languages, yet the natural language processing infrastructure serving them remains fragmented and underdeveloped. The cause is structural: the field organizes its tools and benchmarks around individual languages or small subsets of genealogical language families, building separate analyzers, parsers, and datasets for each language and starting over for the next. This overlooks a deep regularity. Through more than two millennia of convergence around Sanskrit, Indic languages came to share a morphosyntactic architecture formalized in Pānini's grammar, the Astādhyāyī. This cuts across genealogical lines, uniting languages through a common framework. We argue that this Pāninian framework supplies a unifying computational architecture the field has lacked, and that benchmarks grounded explicitly in it would make Indic language systems more accurate, more data-efficient, and more transferable, effectively merging many apparently disparate and sparse Indic language resources into a single high-resource metalanguage bedrock. We propose a four-part benchmark suite to render this shared architecture explicit, measurable, and ready to be leveraged for practical applications. Moreover, we underscore the question it raises for interpretability research: whether neural models trained on these languages come to represent Pānini's categories on their own.

Mechanistic interpretability has made substantial progress in automatically localizing circuits, but explaining what localized components do remains labor-intensive and difficult to standardize. In this work, we study whether language model (LM) agents can assist with this explanation problem once a circuit has already been identified. We introduce AgenticInterpBench, a benchmark for circuit explanation built from 84 semi-synthetic transformer circuits with 163 component-level annotations. We propose HyVE (Hypothesize, Validate, Explain), an agentic explainer that analyzes each component through an iterative loop of observation, hypothesis generation, and causal validation, eventually producing a component-level explanation and a circuit-level task description. Across four LM backbones, HyVE recovers useful component- and task-level explanations, but no backbone is uniformly best. Our analysis shows that strong backbones usually form observation-grounded hypotheses, while failures more often arise later in the validation loop, through incomplete validation plans, code execution errors, or unresolved hypotheses. A case study on an arithmetic circuit in Llama-3-8B shows that the same formulation can extend beyond semi-synthetic benchmarks to naturally trained models. Overall, LM agents are promising circuit explainers, but reliable validation remains the key obstacle.

Coding agents based on large language models (LLMs) demonstrate remarkable autonomous capabilities, but they also introduce significant safety and misuse risks during multi-turn interactions with external environments. Existing safety mechanisms mainly rely on external guardrails, which have a limited ability to perform fine-grained behavioral control during execution. Meanwhile, recent mechanistic interpretability methods for LLM safety are mostly confined to single-turn or jailbreak-style QA settings, limiting their ability to capture the evolving risk dynamics of multi-turn agent execution. In this paper, we investigate the safety of multi-turn coding agents from an internal perspective. We propose AgentLens (Mechanistic Subspace Intervention and Steering), a white-box defense framework that performs runtime safety detection and representation-level mitigation for coding agents. Unlike conventional agent guardrails, AgentLens detect harmful execution states from step-level hidden representations and mitigate unsafe behavior by intervening in a 10-dimensional subspace within a single layer. To support this research, we introduce the Mechanistic Agent Safety (MAS) benchmark, comprising comprehensively annotated multi-turn execution trajectories across 194 tasks using LLaMA-3.1-8B, Qwen-2.5-7B, and Gemma-2-9B. Extensive experiments show that AgentLens achieves strong safety detection performance, provides preliminary evidence for lookahead risk anticipation, and substantially reduces harmful actions of the coding agent, establishing a foundation for applying mechanistic interpretability to dynamic LLM agent safety. The code is available at: https://github.com/EddyLuo1232/AgentLens

Deep learning has become an important tool in computational pathology, enabling automated analysis of histopathological images. While convolutional neural networks (CNNs) have traditionally dominated this field, transformer-based and hybrid architectures have recently demonstrated promising performance. However, comprehensive comparisons of these approaches for colorectal histopathology remain limited. This study evaluated twelve ImageNet-pretrained CNN, transformer, and hybrid architectures using the Kather colorectal histopathology dataset containing 5,000 image tiles from eight tissue classes. All models were trained using a standardized transfer-learning and fine-tuning protocol and assessed using multiple performance metrics, including accuracy, precision, sensitivity, specificity, F1-score, ROC-AUC, Cohen's kappa, and Matthews correlation coefficient. All evaluated models achieved high classification performance, with accuracies ranging from 93.2% to 97.1%. EVA-02 achieved the highest overall performance (97.1% accuracy, 97.0% F1-score), closely followed by ViT-B/16. Among CNNs, ResNet34 and ConvNeXt-Tiny demonstrated highly competitive performance, achieving accuracies of 96.4% and 96.3%, respectively. Transformer architectures generally produced the strongest results across evaluation metrics, although the performance gap between the best transformer and CNN models was relatively small. Per-class analysis showed consistently strong classification performance across all tissue categories, with Complex Stroma representing the most challenging class. Overall, transformer-based architectures achieved the highest predictive performance, whereas modern CNNs provided a favorable balance between accuracy and model complexity. These findings provide a comprehensive benchmark of major deep learning paradigms for colorectal histopathology classification.

Electroencephalogram (EEG) analysis remains the clinical gold standard for epilepsy diagnosis and seizure detection. While Deep Learning (DL) has significantly advanced automated EEG interpretation, its transition from controlled experimental settings to routine clinical deployment is severely bottlenecked by fundamental architectural flaws. Standard DL models operate as opaque black-boxes lacking clinical interpretability, demand massive amounts of balanced annotated data, and incur steep computational costs incompatible with resource-constrained wearable or implantable neuromodulation devices. This paper presents a comprehensive review of these prevailing limitations and explores Kolmogorov-Arnold Networks (KANs) as a emerging paradigm for EEG-based seizure detection. By replacing the fixed activation functions of traditional neurons with flexible, learnable functions along the network's connections, KANs bridge the critical gap between predictive accuracy and mathematical transparency. We systematically analyze how KAN architectures resolve the shortcomings of traditional DL-based models by offering exceptional parameter efficiency, inherent interpretability for physician trust, and robust performance under data scarcity. Ultimately, this review establishes KANs not merely as an incremental algorithmic update, but as a fundamental paradigm shift necessary to actualize next-generation, patient-specific, and thoroughly transparent clinical EEG monitoring systems.

Recent latent reasoning methods, such as CODI and COCONUT, face a fundamental interpretability problem: they maintain multiple superimposed candidate traces in the hidden space at each step, unlike explicit- CoT, which follows a single transparent reasoning trace. Existing mechanistic methods show compression, shortcuts, and superposition without explaining how reasoning evolves across latent steps. To address this gap, we model latent token sequences as trajectories in representation space and apply dynamical systems analysis to characterize the evolution of reasoning. Using quantitative measures, such as step-to-step change, direction consistency, and Lyapunov sensitivity, alongside qualitative projections, such as UMAP and DMD/PHATE, we show that latent CoT exhibits structured, non-random dynamics with two distinct stability classes. CODI behaves as a stable attractor, while COCONUT behaves as an unstable expanding system, and SIM-CoT supervision tightens both behaviors without changing the underlying dynamics. This framework advances the interpretability of latent CoT reasoning dynamics and provides actionable insights for improving latent reasoning performance. Code1 and Project page2 available online.

High-stakes reasoning tasks necessitate transparent and verifiable workflows, yet conventional single-model large language models (LLMs) often struggle with hallucination and low interpretability under zero-shot paradigms. To address this general AI challenge, we propose a Multi-Agent Audit Framework that simulates a collaborative, multi-step verification process. We empirically validate this architecture in the sensitive domain of clinical mental health screening using a modular LangChain workflow. Our framework decomposes the reasoning process into a Perception Agent, Knowledge Retrieval-Augmented Generation (RAG), Chain-of-Thought (CoT) clinical inference, and a critical Audit verification stage. We evaluated this framework on the DAIC-WOZ dataset using locally deployed open-source models. Experimental results demonstrate that our multi-agent pipeline significantly outperforms single-agent baselines, reducing the Mean Absolute Error (MAE) for PHQ-8 depression severity prediction from 5.35 to 5.02. By exposing cross-agent validation traces, the framework mitigates reasoning drift and provides highly interpretable diagnostic rationales, offering a generalizable paradigm for reliable AI-assisted decision support beyond isolated model scaling. We make data and code open access on GitHub for replicability.

LLM reasoning transparency is a critical affordance for understanding model decisions, mitigating misuse and misalignment, and debugging surprising model behaviors. However, DiffusionGemma performs a larger fraction of its computation in a continuous latent space; does this make its reasoning less transparent? We study this question by decomposing transparency into two components: variable transparency, whether we understand intermediate snapshots of a model's computational state; and algorithmic transparency, whether we can use these snapshots to reconstruct the process by which the model arrived at its outputs. Naively, DiffusionGemma has poor variable transparency: its opaque serial depth, the amount of serial computation that occurs in between interpretable model states, seems at first 28.6X higher than the corresponding autoregressive Gemma 4 model. However, we show that we can map the information flowing between denoising steps through an interpretable token bottleneck with no decrease in downstream performance. Treating these intermediate states as interpretable reduces the opaque serial depth to just 1.1X that of Gemma 4. Algorithmic transparency is harder for diffusion models than for autoregressive models because all token predictions in the canvas can change at every denoising step, giving the model the power to implement complicated distributed algorithms during the denoising process. To begin bridging this gap, we conduct a suite of interpretability case studies, uncovering initial evidence of novel diffusion-specific phenomena such as non-chronological reasoning, token and sequence smearing, and intermediate-context reasoning. Finally, we test monitorability, a key application of transparency that measures whether model outputs are useful for downstream tasks. We find that DiffusionGemma is similarly monitorable to Gemma 4.

Neural networks learn features that reflect the hierarchical, multi-scale structure of natural data. Synthetic datasets used to evaluate interpretability methods typically lack this structure, limiting their value as realistic toy models. To close this gap, we introduce a family of synthetic datasets consisting of hierarchical functions defined on critical mean-field percolation clusters embedded in a high-dimensional data space. The percolation data consists of sparse, low-dimensional fractal clusters with a power-law size distribution. Latent variables modeling a taxonomic hierarchy generate each data point's target value. The data model is analytically tractable with known critical exponents that fix its properties without requiring hyperparameter tuning. We leverage a mapping between percolation clusters, random trees, and additive coalescence to propose an almost linear-time algorithm to jointly sample a random tree and its hierarchical latent decomposition, enabling data generation at arbitrary scale. Using probing experiments, we find that the model's ground-truth latent variables can be linearly decoded from neural network activations. Together, sparsity, self-similarity, power-law statistics, and analytical tractability make critical percolation a principled testbed for interpretability research.

Concept Bottleneck Models (CBMs) enhance the interpretability of deep learning networks by aligning the features extracted from images with natural concepts. However, existing CBMs are constrained in their ability to generalize beyond a fixed set of predefined classes and the risk of non-concept information leakage, where predictive signals outside the intended concepts are inadvertently exploited. In this paper, we propose Multimodal Concept Bottleneck Model (MM-CBM) to address these issues and extend CBMs into CLIP. MM-CBM utilizes dual Concept Bottleneck Layers (CBLs) to align both the image and text embeddings into interpretable features. This allows us to perform new vision tasks like zero-shot classification or image retrieval in an interpretable way. Compared to existing methods, MM-CBM achieves up to 51.26% accuracy improvement on average across four standard benchmarks. Our method maintains high accuracy, staying within ~5% of black-box performance while offering greater interpretability.

Pre-trained language models such as BERT achieve strong text classification performance but lack transparency, limiting their use in high-stakes settings. The Tsetlin Machine (TM) offers fully interpretable, clause-based reasoning but captures little semantic information, and prior attempts to bridge the two rely on static word embeddings that miss contextual meaning. We propose a semantic pre-training framework that transfers knowledge from a pre-trained language model into a TM without using embeddings. Text samples are grouped into semantically coherent clusters with K-means or Top2Vec, and the resulting cluster-sample pairs pre-train a non-negated TM with enhanced Type I feedback. The TM thereby learns interpretable semantic keywords that are fine-tuned on downstream tasks. Across five datasets, our method substantially outperforms vanilla and embedding-based TMs and reaches performance competitive with BERT while remaining interpretable.

Linear probes are widely used in interpretability research and often compared by cosine similarity. The Mahalanobis cosine similarity (MCS) between two directions, which reweights the inner product by test data covariance, is a natural task-aware refinement. Ying et al. (2026) report that a probe's MCS to a reference probe trained on the out-of-distribution (OOD) data near-perfectly linearly predicts the probe's OOD AUROC (R^2 = 0.98). Here, we extend this empirical finding across models, layers, and concept domains, and prove this general phenomenon in closed form: For balanced classes whose projections are Gaussian, OOD AUROC and MCS to the reference probe are linear because both are sigmoid-shaped functions of the probe's signal-to-noise ratio (SNR) on the test data. The theory also predicts when this linearity fails, which we verify empirically. MCS offers a theoretically grounded and empirically effective alternative to Euclidean cosine similarity for comparing linear probes.

Concept Bottleneck Models (CBMs) enhance interpretability by projecting learned features into a human-understandable concept space. Recent approaches leverage vision-language models to generate concept embeddings, reducing the need for manual concept annotations. However, these models suffer from a critical limitation: as the number of concepts approaches the embedding dimension, information leakage increases, enabling the model to exploit spurious or semantically irrelevant correlations and undermining interpretability. In this work, we propose Concept Flow Models (CFMs), which replace the flat bottleneck with a hierarchical, concept-driven decision tree. Each internal node in the hierarchy focuses on a localized subset of discriminative concepts, progressively narrowing the prediction scope. Our framework constructs decision hierarchies from visual embeddings, distributes semantic concepts at each hierarchy level, and trains differentiable concept weights through probabilistic tree traversal. Extensive experiments on diverse benchmarks demonstrate that CFMs match the predictive performance of flat CBMs, while substantially mitigating information leakage by reducing effective concept usage. Furthermore, CFMs yield stepwise decision flows that enable transparent and auditable model reasoning with hierarchical class structures.

While Audio Large Language Models (Audio LLMs) excel at multimodal understanding, they suffer from text dominance, a bias where models blindly favor text over acoustic evidence, causing hallucinations. However, the internal mechanisms underlying how these models behave when audio and textual inputs contradict each other remain unexplored. In this work, we present the first mechanistic analysis of this phenomenon by tracing the propagation of internal representations across layers. Our investigation reveals three key findings: (i) text dominance is systematically and empirically across models; (ii) while text and audio rely on functionally distinct pathways, they ultimately converge into a shared semantic space in late layers; and (iii) the text pathway does not erase audio information, but rather actively suppresses intact audio representations. Building on these insights, we leverage back-patching, a training-free intervention that routes late-layer audio activations back into earlier layers. This amplifies the audio representations, enabling them to overcome textual suppression. Our evaluation shows that back-patching consistently reduces text dominance, paving the way for mechanistic multimodal alignment under conflict.

Transformer-based architectures have significantly advanced the generation of complex symbolic sequences, yet a significant gap remains in achieving fine-grained, interpretable control over discrete signal attributes. This paper investigates the mechanistic interpretability of the Multitrack Music Transformer (MMT) and proposes a framework for deterministic attribute modulation without retraining to bridge this gap via inference-time activation steering. Utilizing the Difference-in-Means (DiffMean) methodology, we isolate latent directions for signal attributes, specifically Pitch and Duration, within the residual stream. We validate the Linear Representation Hypothesis in this domain, achieving high correlation between steering magnitude and attribute shift. To address the inherent feature entanglement in multi-attribute steering, we introduce a Dual Steering framework utilizing Gram-Schmidt Orthogonalization. Experimental results demonstrate that this geometric decoupling reduces conceptual interference and signal degradation compared to naive vector addition, enabling independent deterministic control even against strong autoregressive conditioning.

This paper presents a methodology-centered transfer learning framework for fragility adaptation under domain shift, class imbalance, and scarce target labels while preserving engineering interpretability and supporting decision-making under uncertainty. Four transfer learning strategies (instance-based, parameter-based, hierarchical Bayesian, and multi-source) are demonstrated through three complementary case studies: (i) instance-based transfer learning via importance weighting, demonstrated on coastal bridge fragility using Hurricane Katrina observations; (ii) parameter-based transfer learning together with hierarchical Bayesian transfer learning, enabling partial pooling across strata and posterior uncertainty quantification, demonstrated on residential building fragility using Hurricane Ian observations; and (iii) multi-source transfer learning that fuses multiple analytical fragility models with learned source weights and regularized target-domain adaptation, demonstrated on seismic bridge fragility using observations from the 2001 Nisqually earthquake. Across these case studies, direct transfer of source models (i.e. using existing state-of-the-art models) fails under domain shift and severe class imbalance, while targeted adaptation substantially improves failure detection and predictive stability in low-data regimes. These findings highlight the need for systematic guidance on diagnostics, strategy selection, and uncertainty reporting when developing and adapting fragility models.

One of the major difficulties in the mechanistic interpretability of neural networks is the occurrence of polysemanticity, which suggests that each neuron is typically responsible for multiple different tasks, impeding a clean interpretation of their function. The seminal paper of Elhage et al. (2022) argues that this occurs due to superposition, a phenomenon where the neural network represents distinct features as non-orthogonal directions in a lower-dimensional space, a strategy that allows much greater compression of the data without sacrificing fidelity due to the feature sparsity of input vectors. Elhage et al. (2022) empirically validates these hypotheses in a rather natural and simple autoencoder with sparse inputs. The contribution of the present work is to analyze the mathematical basis for the occurrence and optimality of superposition, while rigorously corroborating some of their findings. In particular, we provide upper and lower bounds for the L2 reconstruction loss, tight in the very sparse regime, for power activation functions. A short list of interesting open problems are also included at the end.

Sparse autoencoders (SAEs) are increasingly used to extract interpretable features from language models (LMs), yet a central question remains: when can an SAE-based explanation be treated as a faithful view of an underlying frozen LM We study this through a post-hoc generalization framework that certifies the LM via a sparse proxy, obtained by replacing a native hidden activation with its pretrained SAE reconstruction. Our framework derives an upper bound on the base model's expected risk using four measurable quantities: proxy risk, SAE reconstruction gap, concept-pool mismatch, and sparse complexity. We interpret this certificate as an operational criterion for explanatory faithfulness. In particular, a non-vacuous bound indicates that the extracted sparse features retain meaningful predictive information, while small reconstruction and mismatch errors indicate that the proxy remains behaviorally close to the original model. Empirically, we show that the bound becomes non-vacuous on GPT-2 Small, Gemma-2B, and Llama-3-8B at practical sample sizes. A detailed layerwise analysis of Llama-3-8B reveals a strong depth dependence, with later layers becoming much easier to certify, associated with both stronger local fidelity and weaker downstream error amplification. Finally, through feature-shuffling ablations, we show that the decomposition distinguishes genuine semantic alignment from mere statistical sparsity, providing a useful diagnostic for when SAE-based explanations become less reliable.

Delirium is common in hospitalized patients and is often missed in routine care. We present a user-centered interactive machine learning (UC-iML) framework for delirium detection support that combines physician-guided feature refinement with interpretable modeling. Using 3,862 labeled admissions from six Toronto hospitals in the General Medicine Inpatient Initiative (GEMINI), we integrate administrative variables, laboratory results, medications, and a radiology-derived text indicator. Physicians guide feature refinement and model evaluation, and Shapley Additive exPlanations (SHAP) are used to summarize feature attribution. We evaluate standard supervised classifiers with temporally separated holdout testing and a later-phase validation cohort. Compared with automated and baseline variants, the proposed framework shows better overall discrimination and stronger temporal robustness, while the explanations highlight clinically meaningful signals. These results support UC-iML as a practical human-in-the-loop framework for clinically relevant delirium modeling.

A prominent research direction in mechanistic interpretability is learning sparse circuits over LLM components to reveal how they jointly produce model behavior. However, raw neurons are polysemantic, making learned circuits hard to interpret. Sparse autoencoder (SAE) features alleviate this, but their high dimensionality makes existing intervention-based circuit learning methods computationally prohibitive. We propose CircuitLasso, a scalable circuit-learning approach based on sparse linear regression. CircuitLasso recovers circuits whose structural accuracy matches that of state-of-the-art intervention-based methods on the benchmark data, at a fraction of the computational cost. For interpretability, CircuitLasso efficiently uncovers relationships among SAE features, showing how human-interpretable semantic features propagate through the model and influence its predictions. Finally, we validate the utility of our learned circuits by leveraging their insights to achieve comparable performance at substantially lower cost on a domain-generalization task.

Circuit discovery is a key technique in mechanistic interpretability to pinpoint the model components that are crucial for performing a given task. Although the current state-of-the-art method (EAP-IG) performs well on the metric of (un)faithfulness, it suffers from substantial variability. This includes resampling variance, where the circuit changes when we probe with a new batch of data from the same distribution; rephrasing variance, where the discovered circuit shifts when the prompts are rephrased; and sample-wise variance, where a circuit with low population unfaithfulness exhibits large fluctuations in unfaithfulness across individual samples. This paper studies the roots of these variances. We demonstrate that CEAP, our new circuit discovery method that improves upon EAP-IG with a theoretical guarantee, can substantially lessen resampling variance. We further show that rephrasing variance arises because prompts with different templates tend to activate different circuits in the model. This leads us to argue that it may be challenging to find a comprehensive circuit that explains and controls the model's behavior on a task, which can be expressed in countless templates, suggesting that LLMs may be inherently hard to steer. We show that sparsity, which has been claimed to form more compact and interpretable task circuits, fails to solve this problem. Regarding sample-wise variance, we argue that it is largely benign: extremely poor unfaithfulness scores often stem from how unfaithfulness is defined, rather than from defects in the measured circuits. We show that the magnitude of unfaithfulness is affected by selective contribution scaling, a neural mechanism that accounts for the extremely poor scores sometimes observed.

Evaluating representation similarity is fundamental to representation learning. However, existing metrics suffer from significant limitations: they lack interpretability due to shifting baselines, lack robustness to outliers, and are computationally intractable for large datasets, forcing reliance on heuristic approximations. To address this, we develop an ordinal-similarity framework, instantiated by the Triplet (TSI) and Quadruplet (QSI) Similarity Indices, which measure alignment by quantifying the consistency of ordinal relationships. We theoretically demonstrate this formulation is inherently interpretable, robust to outliers, and computationally efficient. Finally, we establish a formal equivalence between TSI and local neighborhood alignment, measured by Mutual Nearest Neighbors. Empirically, we validate these properties and show that ordinal similarity offers a scalable approach to measuring alignment, enabling practitioners to better understand and design representations.

Predictive modeling for clinical decision support requires both strong predictive performance and transparent, auditable, and human-reviewable decision logic. Although deep learning and tree-based ensemble methods can achieve high accuracy, their black-box nature remains a major obstacle to trustworthy clinical deployment. Moreover, clinical prediction often operates under practical constraints, including limited sample sizes, severe class imbalance, and feature evolution arising from changes in diagnostic criteria or clinical documentation practices. We propose Medical Heuristic Learning (MHL), a constrained paradigm for LLM-assisted rule learning. Rather than relying on updates to implicit model weights, MHL integrates statistical probes, medical knowledge probes, initial rule synthesis, and iterative rule optimization to construct an executable rule-based expert system. The resulting rule system is expressed entirely using the native logical and control-flow constructs of a programming language. Valid rule versions are recorded and retained along the search trajectory, making the decision logic explicit, interpretable, and auditable. MHL also supports continual learning by using previously validated rules as a starting point and iteratively revising them in response to updated feature information under data drift or feature evolution. MHL is not tied to any specific programming language. Comprehensive experiments on medical datasets show that MHL achieves predictive performance comparable to that of state-of-the-art methods, performs favorably in small-sample and highly imbalanced settings, and supports the transfer and adaptive revision of validated rules under feature evolution. Overall, these findings suggest that non-gradient-based heuristic systems offer an approach to balancing predictive performance and transparency in clinical decision support.

Protein language models (pLMs) can generate novel protein sequences with properties beyond those observed in nature, yet the mechanisms underlying protein generation remain poorly understood. Existing mechanistic interpretability methods based on sparse autoencoders and transcoders primarily focus on protein representation learning models and do not capture the computation required for autoregressive generation. Here, we introduce ProGenMech, a mechanistic interpretability framework for generative protein language models that extends cross-layer transcoders (CLTs) to ProGen3, a sparse Mixture-of-Experts model trained for both causal generation and span infilling. Unlike per-layer approaches, CLTs reconstruct each layer using sparse latent variables from all preceding layers, enabling faithful recovery of inter-layer generative computation. We further develop a zero-shot circuit discovery framework to identify sparse latent circuits responsible for protein generation and fitness prediction. In causal generation and zero-shot fitness estimation tasks, ProGenMech outperforms local transcoder baselines in recovering ProGen3's probability distribution and functional scoring behavior, while matching the original model's generative distribution in span infilling tasks. Moreover, the recovered circuits reveal biologically meaningful motifs and functional regions associated with conserved sequence patterns and protein fitness landscapes, establishing a foundation for interpretable and steerable protein generation.

Lung cancer is the leading cause of cancer-related mortality, with approximately 2.5 million new cases and 1.8 million deaths annually, making reliable diagnosis a clinical priority. Although deep learning models have achieved strong performance in lung cancer classification, evaluation has largely focused on predictive accuracy, leaving their decision-making processes insufficiently examined. This study compares three architecturally distinct models: a Convolutional Neural Network (CNN), a pretrained ResNet50, and a Vision Transformer (ViT), trained on the IQ-OTH/NCCD lung cancer CT dataset. Local Interpretable Model-Agnostic Explanations (LIME) were applied to investigate model reasoning. In addition to standard performance metrics, a dual-correlation framework was introduced to measure both prediction agreement and explanation agreement across model pairs. All three models achieved strong classification performance, with ResNet50 attaining 98.61% accuracy, CNN 97.91%, and ViT 93.75%, while all achieved ROC-AUC scores of 0.99. Prediction correlations exceeded 0.99 across all model pairs, indicating highly consistent outputs. However, LIME explanation correlations remained below 0.26, revealing substantial differences in the image regions used to reach those predictions. Analysis of misclassified samples further identified a consistent spatial pattern: incorrect predictions were associated with attention outside the lung parenchyma, whereas correct predictions focused primarily within lung regions. These findings demonstrate that prediction agreement is a poor proxy for reasoning consistency, and that interpretability evaluation must be treated as an independent validation criterion alongside predictive performance in clinical AI systems.

Interpreting machine-learning models has attracted increasing attention, particularly in the physical sciences, where one often seeks to understand the underlying mechanisms rather than merely make predictions. Multiple linear regression is often regarded as an interpretable alternative to more complex models, such as deep neural networks, because its predictions are expressed as explicit weighted sums of input features. However, when input features are strongly correlated, namely in the presence of multicollinearity, the learned weights can exhibit large dataset-to-dataset fluctuations and oscillatory behavior across physically similar features, making their interpretation difficult or even impossible. Although the instability of the weights under multicollinearity is well known in statistics, its consequences for physical interpretation, in particular its connection to oscillatory weights across physically similar features, have not been systematically clarified. Here, we theoretically discuss the mechanism behind this loss of interpretability by analyzing the eigenmodes of the feature correlation matrix. We show that small-eigenvalue modes associated with multicollinearity amplify fluctuations in the weights and generate oscillatory patterns that do not necessarily reflect meaningful contributions. We test this theoretical picture numerically on physics datasets and show that Ridge regularization suppresses these unstable modes, although the resulting weights must still be interpreted with caution. We further confirm the generality of our findings beyond physics by analyzing a diverse collection of publicly available datasets. Our results clarify why, in the presence of multicollinearity, physical interpretation can remain difficult even for linear regression models.

Interpretability is essential for trustworthy medical image diagnosis. However, existing concept-driven interpretable methods have key limitations: Concept Bottleneck Models (CBMs) require scoring all predefined concepts at inference time and for manual intervention, imposing a substantial burden on clinicians, while rationale-based generative approaches often select concepts by class discriminability, which can drift from diagnostic ontologies. To address these issues, we propose Neuro-Symbolic Rule Distillation (NeRD), a framework that produces efficient, ontology-grounded reasoning chains that are sufficient yet non-redundant, without manually crafting diagnostic rules. Experiments on two skin datasets demonstrate strong diagnostic performance and interpretability, and blinded expert evaluation confirms the clinical plausibility of NeRD rationales. Our method further enables a first expert-in-the-loop study for Multimodal Chain-of-Thought-based diagnosis, achieving efficient and effective concept-level intervention.

Behavioral audits of Large Language Models on moral prompts measure what the model says, not the internal computation producing it. We use Transluce, an AI-driven mechanistic-interpretability platform, to examine LLaMA 3.1-8B-Instruct on 54 moral prompts in four batteries: 17 dilemmas, policy, and meta-ethical questions (B1); 6 role-playing scenarios (B3); and a controlled trolley contrast varying the switching mechanism with people fixed (B4, 15 prompts) or identity attributes with mechanism fixed (B5, 16 prompts). Two complementary metric families, five cluster-level metrics and a six-metric neuron-level panel, converge on a Situational Anchor Effect: domain-specific representations dominate the top of the activation list across every battery. The model's ethics-labeled capacity stays essentially constant; its salience (rank, priority, top-of-list presence) is highly sensitive to the interpretive frame the prompt selects. The B4-vs-B5 contrast confirms the model attends to whichever surface feature varies: aggregate ethics metrics are indistinguishable, but the dominant non-ethics distractor mirrors the design. A multi-temperature audit identifies a candidate ethics neuron (L16/N3837) stable across temperatures; a cross-model behavioral proxy on two frontier models yields preliminary evidence of divergence in self-reported moral focus, consistent with an Alignment Wrapper in which RLHF re-orders surface text without removing underlying domain-first frames. We unify these as Frame-Conditioned Moral Computation: the prompt's surface vocabulary selects a feature manifold, and the moral conclusion is downstream of that selection. Behavioral alignment must be supplemented by Mechanistic Alignment: a research program asking whether ethics-related features can be shown causally privileged under controlled frame variation, not merely loud in the explanation.

This study proposes an overall deep learning architecture for multi-class classification of plant diseases from high-resolution leaf imagery, with a particular interest in investigating the behavior of ResNet-50 and a hybrid ResNet + Vision Transformer (ViT) design. A specially gathered image database with 15,200 training images and 3,800 validation images spanning 38 classes across multiple crops, including tomato, apple, grape etc. were subjected to preprocessing steps such as resizing, normalization, and data augmentation to enhance model robustness. Multiple architectures, including ResNet-50, MobileNetV2, and EfficientNet-B0, were trained and compared with the hybrid ResNet + ViT model. All models were fine-tuned using the AdamW optimizer and cross-entropy loss, with early stopping applied to prevent overfitting and ensure generalization. Furthermore, interpretability techniques such as Grad-CAM and saliency maps were implemented to indicate disease-relevant regions, while segmentation-based analysis was performed to identify the affected parts of a leaf. For every one of the considered architectures, ResNet-50 led to the highest accuracy of 98.74%, whereas the hybrid ResNet + ViT model achieved a competitive accuracy of 98.58%, showing that the hybrid architectures were effective in capturing both local and overall information. The experimental results showcase the promise of transformer-based models to achieve highly accurate, interpretable, and computationally efficient computer-based multi-class multi-disease classification systems, providing helpful assistance for cultivation management practices as well as for precision farming.

Mechanistic interpretability methods summarize a transformer component by a single importance score, conflating two distinct roles: a component may matter because it transports task-relevant content, or because the forward computation degrades when its contribution is removed. We introduce \emph{Interchange-Group Sobol Decomposition} (IGSD), a paired-intervention framework that compares matched activation replacement with zero ablation on the same component, estimates two Sobol-style variance indices, and uses their signed difference to separate the two roles, with intervention validity monitored by a symmetric off-manifold diagnostic $\widehat{\mathrm{ST}}>1$. In factual recall, IGSD identifies an early-layer content channel in both GPT-2 small and Qwen2.5-1.5B that standard importance methods underestimate. A controlled subject and relation donor design shows that the early channel transports relation-frame content while late attention transports subject-retrieval content, refining at head granularity to the known $\mathrm{Attn}_{L9H8}$ head. Late-layer clamping confirms that the early signal is expressed through downstream transformations rather than residual pass-through. These results show that replacement and deletion are not interchangeable controls and their divergence provides a practical statistical diagnostic for content transport in transformer components.