NeurIPS 2026 Papers

Large vision–language models (LVLMs) are inherently opaque, making it difficult to determine whether their outputs are grounded in visual evidence or driven by language-model priors. Existing concept-based interpretability methods are limited to classification settings, rely on proxy models, or require predefined tokens, and thus do not extend LVLMs. We propose Text-Guided Concept Learning (TGCL), a weakly supervised framework for extracting multimodal concept vectors in LVLMs without token supervision. TGCL builds concept-to-image mappings from data and extracts patch-level activations via concept-guided probing. It then formulates concept learning as a contrastive disentanglement problem, isolating concept-specific patches from background patches to produce sparse, stable, and semantically aligned concept vectors. We conduct experiments on four datasets—ImageNet, MSCOCO, CIFAR100, and DTD—and three recent LVLMs. TGCL outperforms recent interpretability methods, achieving up to 4\% higher sparsity, 11\% lower instability, 17\% lower overlap, and 20\% improvement on attribution faithfulness compared to state-of-the-art baselines.

Concept vectors aim to enhance model interpretability by linking internal representations with human-understandable semantics, but their practical utility is often limited by noisy and inconsistent activations. In this work, we uncover the SuperActivator Mechanism: a transformer dynamic that amplifies concept activation gaps, concentrating the most reliable concept evidence into a small set of high-activation tokens. To develop a theoretical understanding of this mechanism, we prove that concept-aligned attention heads multiplicatively amplify pairwise activation gaps, with already-extreme activations growing fastest. We find that this amplification is not just theoretical, but also occurs empirically on large-scale models: while in- and out-of-concept activation distributions overlap considerably, the in-concept distribution develops a positive tail clearly separated from the noise. These high-tail tokens, which we call SuperActivators, appear consistently across concept-positive samples, making them reliable indicators of concept presence. Accordingly, SuperActivator-based detection improves F1 by up to 14% over standard concept activation aggregators and prompting baselines across image and text modalities, models, layers, and concept extraction techniques, demonstrating the generality and practicality of our insights. Further empirical analysis demonstrates that the most reliable SuperActivators are sparse, with detection typically peaking when using only 5-10% of in-concept token activations, and capture more faithful localized semantics than global concept vectors.

Recent text-to-image diffusion models (T2I DMs) depend on vast, uncurated datasets that often include copyrighted artworks and personal images, risking the generation of unwanted content. While concept erasure methods have emerged to suppress such outputs, they predominantly rely on text prompts, limiting their effectiveness for visually nuanced or ineffable styles that are difficult to articulate verbally. To bridge this gap, we propose Quiet Prompt (QuP), a reference-based concept erasure method that identifies complex concepts using fewer than ten reference images. QuP operates in two stages: (1) Concept Embedding Generation (CEG), which captures the intricate visual morphology of an unwanted concept into a single latent embedding; and (2) Concept leakage-aware Negative Guidance (CNG), which utilizes this embedding as a negative semantic condition to steer the diffusion process. In addition, to overcome the limitations of conventional fixed-scale negative guidance (negative prompting)—which often suffers from structural distortion and unintended content loss—we introduce concept leakage-aware negative guidance scale $n_g(t)$. This mechanism adaptively modulates the guidance strength at each denoising step based on a semantic measurement of concept leakage between the intermediate generated image and the target concept in a joint embedding space. Extensive evaluations on idiosyncratic art styles and object erasure tasks demonstrate that QuP achieves superior concept suppression and content preservation, outperforming state-of-the-art text-based baselines in both image fidelity and erasure precision. The source code will be publicly provided to ensure reproducibility of our work.

Concept erasure aims to remove a target attribute from a representation while preserving the other information encoded in it. This is difficult beyond the linear setting: a target signal hidden from one probe may remain recoverable by a fresh nonlinear probe, while unconstrained nonlinear updates may remove the target by pushing representations off the manifold of natural hidden states. We propose the *Manifold Representation Hypothesis* (MRH): natural hidden states concentrate near a structured, lower-dimensional manifold, so surgical erasure should act along local manifold degrees of freedom rather than arbitrary ambient directions. We operationalize this hypothesis with the TanGCE family of erasure methods. The core method estimates each representation's local tangent space from nearest-neighbor secants, projects a nonlinear concept-scorer gradient onto that tangent, and applies a per-sample trust-region step; TanGCE+ and TanGCE++ prepend closed-form first- and second-moment erasers before the same manifold-constrained loop. Across 119 settings spanning 13 language models, three NLP concepts, and 40 CelebA-CLIP attributes under two control regimes, TanGCE++ removes more target signal than the strongest published baseline at matched control-damage budgets under one fixed hyperparameter recipe. Applying the same manifold-constrained loop after prior erasers also consistently reduces residual nonlinear leakage without leaving the control budget, empirically supporting MRH's operational prediction that manifold-constrained edits are a useful inductive bias for surgical concept erasure.

Generative recommendation formulates next-item prediction as autoregressive generation over semantic ID (SID) sequences derived from users' historical interactions, making modern recommender systems structurally similar to large language models (LLMs). As privacy and safety concerns grow, these systems increasingly require concept unlearning to remove sensitive or harmful concepts associated with items. However, existing LLM unlearning methods cannot be directly applied to generative recommendation. Unlike word tokens with explicit semantics, SIDs are abstract identifiers that are often shared by both forget and retain items, leading to severe conflicts between concept removal and recommendation utility preservation. To address this challenge, we propose TRACER, an end-to-end concept unlearning framework based on token reassignment. Rather than directly suppressing shared SIDs, TRACER reassigns concept-related items to alternative tokens that better facilitate forgetting while minimizing side effects on retained items. We further introduce a coherence regularizer to preserve semantic consistency among retain items during unlearning. Experiments on real-world recommendation datasets demonstrate that TRACER effectively removes target concepts while substantially better preserving recommendation utility than existing unlearning baselines.

We study equilibrium learning in discretized Bayesian games, focusing on first-price auctions as a testbed for gradient-based learning dynamics under private information. While the symmetric independent private values model admits a unique Bayesian correlated equilibrium (BCE) in the continuous limit, standard convergence guarantees for gradient-based learners do not apply: classical no-regret theory predicts convergence only to Bayesian coarse correlated equilibria (BCCE), which can remain large and far from the competitive equilibrium even under fine discretizations. To bridge this gap, we introduce the notion of Bayesian semicoarse correlated equilibrium (BSCCE), an equilibrium concept that refines BCCE by restricting deviations to probability-preserving transformations consistent with projected gradient dynamics. We provide a linear programming characterization of BSCCE with polynomially many variables and constraints and establish the inclusion hierarchy $BCE \subseteq BSCCE \subseteq BCCE$. Our main finding is that the BSCCE set contracts in many cases as the discretization of the Bayesian game is refined. Even with priors for which BCCE fails to concentrate, the diameter of the BSCCE polytope shrinks toward zero, indicating convergence to the unique continuous-game Bayes-Nash equilibrium. These results provide a principled explanation for the observed robustness of gradient-based learning in discretized Bayesian games, showing that these dynamics select a smaller equilibrium set than standard no-regret theory predicts.

We study the problem of agnostic PAC learning under deterministic labels, where the examples are labeled by an arbitrary deterministic function, which need not belong to the concept class $\mathcal{C}$. This model captures misspecification without stochastic label noise and lies between the realizable and fully agnostic settings. Prior work showed that, when $\mathcal{C}$ has finite diameter $K$ and VC dimension $d$, deterministic labels can improve the sample complexity upper bound from the classical agnostic guarantee $\mathcal{O}(d/\varepsilon^2)$ to $\widetilde{\mathcal{O}}(K/\varepsilon)$, but left a large gap between upper and lower bounds. Our main result sharpens this picture. We prove that every class of VC dimension $d$ and finite diameter $K$ is learnable with sample complexity $ \widetilde{\mathcal{O}}\left( \frac{K^{1/3} d^{2/3}}{\varepsilon} \right), $ and we construct classes for which this bound is tight up to logarithmic factors. We also study the online analogue of this problem and show that the minimax regret is again partly controlled by the diameter of the class. Technically, our upper bounds are obtained via memorization-based learning algorithms that exploit the deterministic label structure.

Many machine-learning quantities are straightforward to compute once a finite data set is fixed, but their population versions can encode much harder computational problems. We study this gap for population risk, attribution, counterfactual explanation, and calibration certification, asking which of these quantities can be approximated in time polynomial in the requested number of bits. The formal basis is the Kawamura-Cook second-order complexity theory, instantiated here with compact-domain real functions carrying explicit evaluation, continuity, and range information. In this framework, expected risk has exactly the difficulty of real integration. Integrated gradients behave differently by dimension: in one dimension they reduce to evaluating endpoints, while in two dimensions even one coordinate can already encode one-dimensional integration when the relevant derivative is provided. The same lens separates the costs in related explanation and certification methods. Interventional SHAP is integration-hard even with two features, succinct coalitional SHAP is #P-hard, nearest-threshold counterfactual search is NP-hard, and continuous calibration certificates combine integration with maximization. These results give a taxonomy of when population-level explanation and certification admit polynomial-time bit approximations, and what additional structure is needed for tractability.

Time-series classifiers deployed in risk-sensitive domains such as healthcare, wearable sensing, and industrial monitoring must be accurate, interpretable, and capable of abstaining when evidence is weak. No existing method delivers all three, and the obstacle is structural, rather than logistical. Meaningful temporal patterns cannot be defined without knowing segment boundaries, yet meaningful boundaries cannot be placed without knowing which patterns to expect. Existing approaches break this circularity through fixed windows or expert-annotated vocabularies, sacrificing at least one of the three properties. We introduce ConceptTime, which closes the loop through a single self-supervised signal. A frozen probabilistic forecaster predicts a distribution over the next observation at every timestep of the sequence. We call the mismatch between its prediction and the realized signal, the predictive surprise. This signal both locates segment boundaries and characterizes each resulting segment through the forecaster's own predictive statistics. Segments are thus grounded by \emph{how} they behaved relative to expectation, not by what they look like. These summaries cluster into a vocabulary of dynamical regimes that we call concepts. The vocabulary distinguishes calm-after-spike from calm-after-calm, and unifies visually distinct segments that violate expectation in the same way. A lightweight head classifies from the concept sequence, and the distance to the nearest concept yields per-input reliability for free. Across seventeen UEA, human-activity, biomedical, and fault-diagnosis benchmarks, ConceptTime matches state-of-the-art black-box accuracy, beats every interpretable baseline, and retains near 98% accuracy on Epilepsy with 1% of labels. It equals or exceeds softmax, prototype-distance, and SHAP/TimeX/TimeX++ on OOD detection and deletion-faithfulness, without a single annotation, language-model call, or domain expert.

Interpreting the internal activations of neural networks can produce more faithful explanations of their behavior, but is difficult due to the complex structure of activation space. Existing approaches to scalable interpretability use hand-designed agents that make and test hypotheses about how internal activations relate to external behavior. We propose to instead turn this task into an end-to-end training objective, by training interpretability assistants to accurately predict model behavior from activations through a communication bottleneck. Specifically, an encoder compresses activations to a sparse list of concepts, and a decoder reads this list and answers a natural language question about the model. We show how to pretrain this assistant on large unstructured data, then finetune it to answer questions. The resulting architecture, which we call a Predictive Concept Decoder, enjoys favorable scaling properties: the auto-interp score of the bottleneck concepts improves with data, as does the performance on downstream applications. Specifically, PCDs can detect jailbreaks, secret hints, and implanted latent concepts, and accurately surface latent user attributes.

Transformer predictions depend on positional embeddings, which are known to produce biases such as first- and last-token dominance. However, feature attribution methods implicitly entangle influence from positions with features into a single opaque importance score per feature. To uncover these positional effects, we propose Position-aware eXplanation (PaX), a model-agnostic framework that transforms standard attribution methods to jointly produce feature and position attributions. PaX additionally produces counterfactual positional explanations: actionable scores quantifying how the prediction changes when a feature is relocated. We use PaX to generalize perturbation-, gradient-, and boundary-based attribution methods across vision, language, and clinical time-series benchmarks with no architectural changes. We demonstrate that separating positional effects improves both feature and position attribution faithfulness by 17.5% / 55% and 16.8% / 43%, respectively (insertion / deletion). In a case study on sepsis forecasting, we find that PaX recovers clinically validated bedside signals that standard attributions miss.

Autoregressive transformers with no positional embeddings (NoPE) can recover absolute position. A common explanation is that causal attention creates a position-dependent variance signal, but this account is incomplete because LayerNorm removes per-token scale, so a viable mechanism must encode position in *direction*, not magnitude. We describe the most compact mechanism with two-layer NoPE transformers that accurately extract position. In Layer1, prefix averaging creates a shared beginning of sequence (BOS)-direction component. In Layer2, the attention's output-value matrix maps the \bos{} and \nbos{} contributions into two distinct directions, and a position-dependent \bos{} attention weight acts as a mixing coefficient that interpolates between them. The resulting directional trajectory is linearly decodable for position. Interestingly, we show two transformer variants that converge to the same theoretical mechanism and instantiate it when trained to predict position.

While While Vision Foundation Models (VFMs) are increasingly used in several computer vision tasks, their internal representations remain substantially opaque. This mostly stems from polysemanticity, i.e., individual neurons encode mixtures of unrelated visual patterns. In stark contrast, human reasoning is inherently monosemantic, i.e., it naturally relies on distinct, isolated features to process information. Therefore, monosemanticity needs to be implemented within the model's activation space to obtain representations that are human-understandable concepts. Existing methods mostly operate on convolutional networks, so when applied to VFMs their "crop-and-resize" approach distorts the model's representations by disrupting global self-attention mechanisms and discarding spatial geometry. Furthermore, prior work requires downstream classification labels and is based on KL-divergence, which requires to propagate the gradients back from the classifier head. Ultimately incurring in excessive concept extraction time, and making it hardly applicable to extract task-agnostic concepts. To overcome these issues, we propose PLACE, a fully unsupervised concept extraction framework that operates directly on the native patch-token activations of a VFM. By introducing a geometric Gram-matrix alignment loss, PLACE mechanistically controls downstream KL divergence, thus enforcing faithfulness -- i.e., the notion that extracted concepts, although obtained in an unsupervised fashion, are relevant to downstream classification tasks. Evaluations on image classification tasks (ImageNet and COCO) on DINOv2, ViT-MAE and ViT-B/16 show that PLACE (i) using Gram-alignment it extracts concepts about $20\times$ faster compared to KL-supervised extraction and $11\times$ faster than Sparse Autoencoders (SAEs). Moreover, PLACE yields highly sparse concepts that (ii) achieve up to 40 percentage points sparser compared to the state of the art (i.e., max 0.933 Gini sparsity score), and (iii) are up to 92.3% similar to those obtained under classifier supervision (i.e., 0.923 cosine similarity).

Contrastive vision-language models such as CLIP are trained to align images and text, not to represent taxonomies. Yet empirical studies show that CLIP embeddings exhibit a hierarchical geometry aligned with the semantic hierarchy of concepts. A theoretical characterization of why such hierarchy emerges from a non-hierarchical objective remains open. In this paper, we formalize this phenomenon as Hierarchical Angular Separation (HAS), a geometric property whereby each non-leaf concept embedding is more similar to its descendants than to off-branch concepts. Empirically, we show that HAS already holds weakly in randomly initialized CLIP image embeddings, suggesting that hierarchical structure may be induced by the raw image geometry, while in text embeddings it emerges only after contrastive training. To explain this asymmetry between text and image embeddings, we analyze an unconstrained features model in which the image embeddings are fixed and only the text embeddings are trained. We show that when the image embeddings satisfy HAS with a positive margin, the global minimizer of the contrastive objective induces HAS in the text embeddings, thereby transferring hierarchical structure across modalities. The key mechanism is the hierarchical co-occurrence structure of image-text pairs, which biases text embeddings toward root-centered image prototypes. For a linear contrastive loss, we derive a closed-form solution that makes this mechanism explicit. We extend the analysis to the InfoNCE loss and establish analogous guarantees under sufficient regularization. Synthetic experiments validate the theory and show that HAS emerges even beyond the regimes covered by our analysis.

Large Language Models (LLMs) can generate plausible free text self-explanations to justify their answers. However, these natural language explanations may not accurately reflect the model's actual reasoning process, indicating a lack of faithfulness. Existing faithfulness evaluation methods rely primarily on behavioral tests or computational block analysis without examining the semantic content of internal neural representations. This paper proposes NeuroFaith, a flexible framework that measures the faithfulness of LLM free text self-explanation by identifying key concepts within explanations and mechanistically testing whether these concepts actually influence the model's predictions. We show the versatility of NeuroFaith across 2-hop reasoning and classification tasks. Additionally, we develop a linear faithfulness probe based on NeuroFaith to detect unfaithful self-explanations from representation space and improve faithfulness through steering. NeuroFaith provides a principled approach to evaluating and enhancing the faithfulness of LLM free text self-explanations, addressing critical needs for trustworthy AI systems.

Discovering explanatory subgraphs is essential for interpreting Graph Neural Networks (GNNs). Many existing and popular explainers optimize by maximizing mutual information (MMI) between selected subgraphs and predicted labels; however, the resulting optimization landscape is often non-smooth and poorly conditioned, leading to suboptimal local optima. To address this issue, we develop a norm-theoretic framework for subgraph analysis. We show that subgraphs containing model-utilized features tend to induce graph embeddings with larger $\ell_2$ norms, as they align more strongly with the supportive subspace shaped by the GNN’s weight matrices, a phenomenon consistently observed across multiple datasets. Based on this insight, we propose NORMA, a parameterized explainer that uses subgraph-embedding $\ell_2$ norms as a stable optimization signal to guide subgraph selection. Experiments on fsix benchmark datasets demonstrate that NORMA effectively alleviates the optimization difficulties of MMI-based methods and achieves superior explanation quality.

Concept bottleneck models predict outcomes from high-level concepts detected in inputs. Although concepts provide a simple way to reap benefits from interpretability, very few datasets include concept labels. This limits researchers' ability to determine which problems are suitable for these models, isolate the factors that drive their performance or lead to failures, or uncover which algorithms perform well. In this paper, we develop synthetic benchmarks for concept-bottleneck models, focusing on their two main use cases: decision support, in which models assist humans in making better decisions, and automation, in which models handle routine tasks without supervision. Our benchmarks can generate labeled datasets while controlling for properties that affect performance, including data modality, concept choice, annotation quality, and completeness. We demonstrate how the benchmarks can be used to evaluate representative classes of concept bottleneck models. Our demonstrations show how the benchmarks can diagnose failure modes and guide follow-up testing.

We study the problem of multiclass classification within the framework of *marginal-nonuniform* learning, where the constants in the learning rate may depend on the unknown marginal distribution generating the unlabeled data, but not on the target concept. This notion naturally lies between classical PAC learning, where the constants are independent of both the marginal distribution and the target concept, and more recent universal learning, where the constants may depend on both. We give a complete characterization of the optimal marginal-nonuniform learning rates, establishing a fundamental trichotomy: exponential, linear, and arbitrarily slow rates. This characterization is witnessed by a new combinatorial dimension, termed the *uncentered DS-eluder dimension*. Technically, a key ingredient in the upper-bound proofs is a set of new *distributional exhaustion* lemmas that exploit the sequential nature of this dimension and may be of independent interest.

Concept Bottleneck Models (CBMs) provide interpretability but typically incur a penalty in accuracy. Recent CBMs rely on LLM-supplied concept vocabularies that are misaligned with the underlying image evidence and obscure dataset-specific discriminative signals. Augmenting these representations with task-specific linear probes recovers little of the lost accuracy. Methods that instead learn concepts directly from data also fall short of the supervised CLIP linear-probe ceiling. We introduce MOSAIC (Mixtures Of Sparse Anchored Interpretable Concepts), a CBM whose concepts are discovered directly from image representations via text-anchored optimal transport, a balanced-marginal pass that softly assigns each image to a small set of learned concepts while keeping them tethered to class-text embeddings, and then named post-hoc by a VLM. Across 5 datasets, MOSAIC matches the supervised CLIP linear-probe ceiling on most dataset-backbone settings and beats raw zero-shot CLIP by +5.2 to +24.4 top-1 and class-level retrieval R@1 by +5.7 to +27.0. Beyond classification, the discrete concept basis enables a sparse per-image decomposition that exposes which features drive each prediction and aggressive bottleneck compression to as little as 2.5% of dense CLIP.