NeurIPS 2026 Papers

Federating a LGN is challenging because each gate is parameterised by a distribution over 16 Boolean operations, so averaging parameters across clients does not correspond to averaging the underlying functions. Under non-IID data, such parameter-space merging collapses to near-chance accuracy. Boolean Feature Selection (BFS) avoids this issue by selecting discrete Boolean functions rather than averaging parameters. We propose Federated Boolean Feature Selection (FBFS), which extends BFS to LGNs by treating hardened last-layer gates as candidate Boolean features and aggregating per-class, per-gate sufficient statistics across clients. Clients share only firing aggregates computed on a held-out split, allowing the server to reconstruct discriminative scores, defined as the gap between a gate's mean firing for a class and for the remaining classes, without accessing raw data. This procedure is equivalent to centralized BFS on the union of client data. We further introduce a diversity-regularized variant of FBFS that encourages complementary feature selection across gates. Empirically, FBFS is the only data-free method we evaluate whose performance remains stable under extreme non-IID settings, outperforming parameter-space merging and other baselines by large margins. It maintains high accuracy across a wide range of heterogeneity levels and scales to large architectures. In addition, FBFS enables efficient deployment: the resulting models compile to hardware-efficient logic representations with substantially fewer post-synthesis cells than ensemble-based alternatives while remaining bit-exact with their software counterparts.

Diffusion models generate data through a continuum of denoising problems, and are widely observed to reveal coarse structure before fine detail. Yet, this intuition is mostly empirical and qualitative. We introduce \emph{feature information dynamics}, an information-theoretic framework for localizing when a feature is generated during diffusion. Using the I-MMSE identity, we connect the rate of feature mutual information change to a gap between optimal unconditional and feature-conditional denoising losses, yielding practical estimators for feature information density. We further develop a chained decomposition that separates shared from incremental information in a feature hierarchy. We use this framework first to quantitatively confirm spectral autoregression in pixel diffusion, and then to extend the analysis beyond frequency: under a class$\to$mask$\to$Canny conditioning chain, the per-feature information densities differ across pixel, SDVAE, VAVAE, and RAE, exposing fundamental differences between these representations.

Gradient-based attribution methods are model-faithful and scalable, but Integrated Gradients (IG) can be brittle because explanations depend on heuristic baselines, straight-line paths, discretization, and saturation. We propose Fisher--Rao Integrated Gradients (FRInGe), which defines both the reference and interpolation schedule in predictive distribution space. FRInGe replaces input baselines with a maximum-entropy predictive reference and follows a Fisher--Rao geodesic on the probability simplex. The corresponding input-space trajectory is realized through the pullback Fisher metric and stabilized by KL and Euclidean trust regions; attributions are obtained by integrating input gradients along this trajectory. Across six ImageNet architectures, FRInGe most clearly improves calibration-oriented attribution metrics, especially MAS scores, while remaining competitive on perturbation AUC and infidelity.

We present FM-ChangeNet, a pathwise-supervised framework for change detection that reformulates bi-temporal reasoning as continuous transport in feature space rather than static endpoint comparison. Given encoded pre and post-temporal representations, we construct intermediate latent states and learn a time-conditioned velocity field $\hat{v}_\theta(z_t,t)$ along the transformation trajectory. This pathwise formulation constrains the predictor over a continuum of intermediate states, providing a denser and less ambiguous supervision signal than conventional endpoint-only segmentation and enabling the model to capture temporal evolution explicitly. The learned velocity field is not only a transport mechanism but also an interpretable representation of change: its magnitude serves as a spatially localized change cue that helps distinguish true structural variation from nuisance effects such as illumination shifts and spatial misalignment. We develop a hierarchical multi-scale architecture with cross-temporal alignment, time-conditioned coarse-to-fine flow decoding, and a unified objective that couples flow supervision, trajectory consistency, spatial regularization, and segmentation loss. Experiments on remote sensing benchmarks show that the proposed framework produces more structured and robust change representations while achieving state-of-the-art performance.

Efficient benchmarking techniques aim to lower the computational cost of evaluating LLMs by predicting full benchmark scores using only a subset of a benchmark's questions. By reframing this problem as an instance of multiple regression with feature selection, we find that existing efficient benchmarking methods can be greatly improved by simply using kernel ridge regression at the prediction stage. Additionally, using an information-theoretic feature-selection algorithm called minimum redundancy maximum relevance (mRMR), we can further improve upon these methods by selecting question subsets that will be maximally useful for prediction. Except in very data-poor settings, these approaches consistently achieve smaller prediction errors (in both MAE and RMSE), and greater ranking correlation between predicted and true scores (in both Spearman $\rho$ and Kendall $\tau$) across a range of benchmarks using both binary and continuous metrics. Furthermore, mRMR subsampling is much faster than competitor methods (which often involve fitting probabilistic models or running clustering algorithms), and is more likely to select the same questions under different random seeds or training data splits.

We propose ESENSC, a computationally efficient and theoretically grounded alternative to SHAP. ESENSC is a polynomial-time attribution rule with a deterministic, closed-form representation that satisfies the null-player property and admits an axiomatic characterization based on efficiency, a restricted form of differential marginality, and explicit computational constraints. Empirically, ESENSC closely replicates the attribution patterns of exact SHAP by preserving its essential structural properties, while maintaining a fraction of the cost. Across neural network and XGBoost models, it achieves lower deviation and higher rank agreement than widely used sampling-based SHAP approximations. While the computational cost of exact SHAP grows exponentially with the number of features, ESENSC scales linearly, enabling reliable feature attribution in regimes where exact SHAP is intractable. These results demonstrate that ESENSC achieves near-SHAP attribution accuracy while reducing computational complexity from exponential to linear in the number of features, providing a practical and theoretically rigorous alternative for feature attribution in high-dimensional applications.

Disentangling writing style from semantic content is a fundamental challenge in literary text modeling. Style-content entanglement causes models to rely on semantic content for authorship attribution (AA) tasks and memorize author-specific content for style imitation (SI) tasks. We propose EDITORS (Editing LoRA Subspaces), a diagnose-then-deflate framework that adapts pretrained large language models and prioritizes stylistic features, making it effective for both AA and SI tasks. EDITORS first trains a diagnostic LoRA adapter using style-neutral statements derived from the training corpus to identify a content-orientedsubspace. It then trains the final style adapter under activation-space deflation that projects input activations away from the identified content-oriented directions. We demonstrate that EDITORS is empirically effective: On three AA benchmarks spanning literature, social media, and news, EDITORS achieves state-of-the-art accuracy, outperforming the strongest existing baseline by up to 29%. On SI, EDITORS generates stylistically faithful and diverse passages, outperforming baselines in achieving style similarity while adhering to content specifications.

While Gaussian Splatting-based Feature Fields (GSFFs) have shown promise for visual localization, this paper highlights that photometrically optimized GSFFs are inherently ill-suited for 2D-3D matching. The volumetric extent of each Gaussian induces many-to-one pixel-to-point mappings that destabilize PnP-based pose estimation, while photometric optimization gives rise to superfluous Gaussians devoid of multi-view consistency. To address these issues, we propose SplitGS-Loc, a localization-specialized GSFFs construction framework that disambiguates 2D-3D correspondences by exploiting Gaussian attributes. Our key design, Mixture-of-Gaussians-based splitting, decomposes each Gaussian into smaller Gaussians, replacing ambiguous many-to-one with precise one-to-one correspondences. In parallel, we exploit composition weights from GS rasterization to select Gaussians that significantly and consistently contribute across multiple views and aggregate discriminative features through strong pixel-Gaussian associations, enforcing multi-view consistency. The resulting compact yet discriminative feature fields enable stable PnP convergence, achieving state-of-the-art performance on localization benchmarks. Extensive experiments validate that SplitGS-Loc extends the utility of photometric GSFFs to accurate and efficient localization by exploiting Gaussian attributes, without per-scene training or iterative pose refinement.

Mechanistic interpretability seeks to explain neural network behavior by decomposing model computations into interpretable features and circuits. While transcoder-based circuit tracing has recently enabled detailed causal analyses of large language models, multimodal diffusion transformers for image generation remain comparatively opaque. We still lack tools for understanding how semantic information propagates across denoising steps and how text and image representations interact within double-stream MM-DiT architectures. Existing methods provide only partial insight: attention maps expose a limited view of token interactions, while sparse autoencoders can discover interpretable features but do not directly reveal how these features are transformed and composed through nonlinear MLP layers. In this work, we extend transcoder-based circuit tracing to multimodal diffusion transformers. We train timestep-conditioned transcoders that faithfully approximate the input–output behavior of MLP sublayers in FLUX.1[schnell]. By replacing MLPs with transcoders and linearizing the remaining computation, we obtain exact feature-to-feature attribution and recover compact, interpretable circuits. Empirically, our transcoders match or slightly outperform sparse autoencoders on the sparsity–faithfulness tradeoff. The resulting circuits reveal mechanisms underlying attribute binding and cross-stream semantic propagation, and provide causal explanations for systematic generation errors. Moreover, circuit-guided interventions are substantially more precise and effective than standard SAE-based steering. Our results demonstrate that transcoder-based circuit analysis is feasible for state-of-the-art diffusion transformers and provides a powerful framework for understanding and controlling multimodal generative models.

We study cross-sample modality attribution in multimodal learning: how a modality in one training sample affects learning from other samples. This problem is central to interpretation because a modality may be redundant for some samples, uniquely informative for others, or useful only through synergy with other modalities. However, existing data attribution methods are insufficient for this setting because they typically estimate sample-level contributions using additive or linear approximations, and therefore cannot capture nonlinear interactions across samples and modalities. We propose a structured kernel surrogate (SKS) framework for multimodal attribution. SKS learns a mapping from sample-and-modality selections to the final loss. To approximate this mapping, SKS uses kernel ridge regression with a structured radial basis function kernel that captures nonlinear effects across samples and modalities. The kernel also preserves the natural hierarchy of multimodal perturbations by modeling sample participation separately from within-sample modality composition. Across diverse environments and benchmarks, SKS achieves $60.3$% higher linear datamodeling score than the strongest baseline. For downstream data selection, SKS improves the average accuracy by $2.24$% over the strongest baseline. Sharpness analysis further shows that models trained on data selected by SKS have lower Hessian trace values, indicating a flatter loss surface and better generalization. Beyond these quantitative gains, SKS also provides interpretable modality-level selections in spatiotemporal traffic accident prediction, preferentially retaining satellite tiles with visible road infrastructure over vegetation-dominated tiles.

In high-stakes settings, decision-makers using conformal prediction (CP) read the predictive interval's width as their operational uncertainty signal. Local feature attribution methods, applied to that width, silently conflate two semantically distinct contributions: a calibration-driven baseline that depends on the held-out calibration set, and an instance-driven term that depends on the test point. We propose a two-game Shapley construction---one game over feature coalitions at the test point, one over feature coalitions across the calibration set---that decomposes the total attribution into calibration-driven and instance-driven components. The construction is method-agnostic on two axes: any local attribution method that can be applied to a scalar uncertainty target can be used, and any split CP family whose uncertainty size admits a known separation into a calibration-only and an instance-only summary satisfies the decomposition. Empirically, the calibration-driven and instance-driven attributions frequently carry opposite signs on synthetic and real data, with representative cases where a feature's sign reverses between the two views; the total attribution then becomes a misleading sum. On a clinical-decision task with semi-synthetic data, this conflation degrades decision quality, whereas jointly considering the calibration-driven and instance-driven attributions recovers high-uncertainty cases missed by total attribution. We further empirically verify that the decomposition holds across multiple CP families that satisfy the separability conditions.

This paper aims at analyzing the regularization effect that data augmentation induces on supervised regression methods in the proportional regime, where the number of covariates grows proportionally to the number of samples. We provide a tight characterization of the test error, measured in mean squared error, in terms only of the population quantities of the true data, as well as first and second order statistics of the augmentation scheme. Our results are valid under misspecified feature maps, and for any network architecture where only the last readout layer is trained, and the rest of the network is either frozen or randomly initialized. We specify our results in the case of Gaussian data, and show that our asymptotic characterization is tight in this setting.

Generative models can produce individually plausible samples while deviating substantially from a reference set in the distribution of key features. For example, a model pretrained on broad drug-like chemical space may generate molecules whose molecular features differ from those of a therapeutic class of interest, such as known antibiotics. Correcting such distributional _miscalibration_ is challenging: direct finetuning on the reference set can overfit and does not control which features are matched. To fill this gap, we introduce kernel Calibrating Generative Models (kCGM). kCGM minimizes a maximum mean discrepancy (MMD) between generated and reference feature distributions using an unbiased score-function estimator, with KL regularization to remain close to the pretrained model. On a reference set of 174 antibiotics, direct finetuning sacrifices chemical validity for feature-distribution matching, whereas kCGM improves target feature matching while increasing validity. We further demonstrate kCGM in protein and DNA generation tasks, showing it can adapt autoregressive, continuous-space diffusion, and discrete diffusion models using only feature-level supervision.

The design of modern neural architectures has converged through incremental empirical choices, yet the mechanisms governing their training dynamics remain only partially understood. We identify and analyze a negative weight drift induced by the interaction between standard losses and positively biased activation functions. We prove that under MSE or cross-entropy loss, the gradient with respect to positive pre-activations is non-negative in expectation at initialization, driving downstream weights toward negative values during early training. The drift is intrinsic to optimization rather than data, and persists across architectures (MLP, ResNet, ViT, GPT-nano, MP-SENe) and asymmetric activation functions (ReLU, GELU, SiLU). Coupled with ReLU, weight drift produces activation sparsity reaching up to 90\% in GPT-nano. We characterize the sparsity–accuracy tradeoff across 79 configurations and identify a sharp accuracy cliff above $\sim$70\% activation sparsity. While ReLU$^2$ achieves a good sparsity–accuracy tradeoff, it pathologically amplifies identified activation spikes in intermediate transformer layers. Clipping resolves this while preserving the representational benefits of squaring: clipped ReLU$^2$ outperforms its unclipped version, and GELU$^2$ achieves the lowest validation loss on GPT-nano.

Low-label transfer is often decided by a practical question: after a representation looks useful, how many labels are still needed to learn the downstream boundary? Existing transferability and probe-quality scores rank global alignment, likelihood, margins, or calibration, but they do not measure how much target data remains close to the fitted decision boundary, where scarce labels are most costly. We propose **effective boundary complexity**, a label-budget diagnostic for frozen representations and linear probes. It measures mass in a thin strip around the learned boundary and normalizes it by class balance, so dense ambiguous regions and rare-class difficulty are both counted. We prove that, under a smooth representation map, this boundary mass does not change because of generic volume expansion: density change and tangential surface change cancel at first order, leaving only stretch or compression in the boundary-normal direction. This yields a cached-feature estimator that fits a probe on one split, counts held-out features near its boundary, and avoids high-dimensional density estimation. Experiments support the diagnostic at three levels. Controlled transformations match the predicted law across 74 maps with 0.38% median relative error. On real frozen features, normal-direction changes move both measured boundary mass and label demand, while tangent changes do not. A differentiable surrogate reduces measured complexity by 16–36% at matched top-1 utility. Finally, CIFAR-100 and Tiny-ImageNet audits show boundary mass acting as a second-stage diagnostic for near-tied representations, with external multiclass results showing how label-grid saturation limits policy-level gains on label grids where many tasks tie.

Graph neural ordinary differential equations provide continuous-time propagation on graphs, but their long-time behavior is largely determined by the asymptotic structure of the mixing operator. We show that diffusion-style Graph ODEs with positive irreducible mixing face a *monostability trap*, where node features converge to a rank-one consensus subspace. This reveals a structural limitation of positive diffusion-style continuous graph propagation: within this consensus class, exact long-time avoidance requires changing the effective support; otherwise, smooth suppression should be interpreted as finite-horizon or metastable mitigation. We address this by introducing a *hysteretic mechanism* for topology evolution, where candidate edges carry latent bistable potentials in a double-well landscape. A node-feature-conditioned force field evolves these potentials, and a smooth sigmoid gate converts them into differentiable propagation weights. We prove the consensus trap theoretically, show that idealized support separation yields block-wise rather than global consensus, and evaluate the framework on real-world benchmarks.

Early-exit neural networks improve inference efficiency by enabling predictions at intermediate layers, but their effectiveness critically depends on the exit policy that decides whether to stop or continue computation. Existing policies, often based on confidence thresholds, enable early termination but offer limited insight into the reliability of such decisions. To address this, we introduce an explainability-aware framework that defines an attribution-grounded reliability signal for early-exit decisions. We propose Progressive Feature Attribution Maps (PFAM), which capture how feature importance evolves across exits, and the Interpretability-Based Early-Exit Score (IEES), a unified decision metric that integrates prediction confidence with attribution strength and cross-exit stability. To enable efficient deployment, we further introduce a lightweight proxy predictor that estimates IEES using inexpensive, attribution-free features, allowing attribution-informed decisions without incurring inference-time overhead. Experiments on convolutional architectures (ResNet, MobileNet, MSDNet) show that our approach improves accuracy–efficiency trade-offs, achieving up to 2.5\% accuracy gains while reducing inference time by up to 24\% compared to full-depth inference. Similar trends are observed on a vision transformer and a BERT-based language model, suggesting applicability beyond convolutional architectures.

Attribute-Missing Graph Clustering (AMGC) is a critical yet challenging task in real-world applications, in which only a subset of nodes hold complete attributes information while others are partially missing. Though existing feature propagation-based methods can effectively recover missing node attributes, they commonly assume uniform contributions across all nodes, ignoring unreliable nodes that may degrade imputation quality. Furthermore, adopting low-pass graph filters often suffers from the over-smoothing problem and inevitably sacrifices discriminative information. %Besides, traditional contrastive learning pulls all the node embeddings evenly, which might conflict with the rule that intra-cluster nodes should be closer to each other. To address these limitations, we propose \underline{\textbf{A}}daptive \underline{\textbf{F}}eature \underline{\textbf{P}}ropagation (AFP) for attribute-missing graph clustering. Specifically, we first design a reliability-aware feature propagation mechanism that adaptively weights edges based on node importance. Then, we introduce a multi-scale embedding module to capture both local and global structural information. Finally, we develop a topology-aware contrastive loss to enhance clustering consistency. Extensive experiments on benchmark datasets demonstrate the superiority of our method.

Understanding what features are encoded by learned directions in LLM activation space requires identifying inputs that strongly activate them. Feature visualization, which optimizes inputs to maximally activate a target direction, offers an alternative to costly dataset search approaches, but remains underexplored for LLMs due to the discrete nature of text. Furthermore, existing prompt optimization techniques are poorly suited to this domain, which is highly prone to local minima. To overcome these limitations, we introduce ADAPT, a hybrid method combining beam search initialization with adaptive gradient-guided mutation, designed around these failure modes. We evaluate on Sparse Autoencoder latents from Gemma 2 2B, proposing metrics grounded in dataset activation statistics to enable rigorous comparison, and show that ADAPT consistently outperforms prior methods across layers and latent types. ADAPT also produces significantly more interpretable prompts in human evaluation. Our results establish that feature visualization for LLMs is tractable, but requires design assumptions tailored to the domain.