NeurIPS 2026 Papers

Vision-language models map images and text into a joint embedding space. However, these embeddings often entangle multiple semantic features, which limits their interpretability and controllability. While sparse autoencoders have emerged as a useful tool for decomposing these embeddings into monosemantic features, their application to joint embedding spaces has largely relied on an implicit, untested assumption that semantically corresponding features share the same directions across modalities. In this paper, we challenge this assumption by identifying discrepancies in feature directions for the same concept across image and text modalities, a phenomenon we term *cross-modal feature heterogeneity*. We demonstrate that this heterogeneity is a key driver of the *modality split*, where a shared concept activates different latents depending on the modality. This finding further reveals why aligning latent activations alone is insufficient to resolve the underlying feature mismatch. To address this misalignment, we propose an approach that trains sparse autoencoders to preserve the unique feature geometry of each modality and aligns corresponding features post hoc. Our method improves reconstruction fidelity and enhances performance in cross-modal retrieval and concept steering.

Guard models are the last line of defense between a language model and a harmful output, yet their training objective is surprisingly narrow. Existing guards learn to predict a single verdict token from a conversational context, concentrating supervision on a single target. The consequences are structural: models latch onto shortcut features, are overconfident, and remain sensitive to where safety evidence appears in the sequence rather than its role in the full context. We propose a different framing. Rather than predicting a label from text, our $\textbf{LLaDA Guard}$ asks which label better explains the text: scoring the prompt or response under each label hypothesis and classifying based on their difference. This shifts supervision to every token in the moderated region, forcing the model to account for full content rather than its most discriminative fragments. We instantiate this idea with a masked diffusion language model, fine-tuning LLaDA-8B-Instruct with a class-conditional reconstruction objective using LoRA and requiring no architectural changes beyond the base model. LLaDA Guard leads on average rank against discriminative baselines trained on stronger backbones across seven held-out safety benchmarks, while exhibiting substantially better confidence calibration (ECE 0.0875 vs. 0.1384 for Qwen3Guard), less over-defense on benign prompts with $\texttt{unsafe}$-looking cues, and less prompt leakage when moderating responses. Its generative nature further enables token-level risk localization as a natural byproduct, yielding a pipeline for rewriting $\texttt{unsafe}$ prompts into $\texttt{safe}$ equivalents without additional training and achieving a 60.7% average conversion-to-$\texttt{safe}$ rate. The weights and code for reproduction are available at the following [anonymized repository](https://anonymous.4open.science/r/LLaDA-Guard2026-B056/).

Zero-shot 3D point cloud classifiers built on large-scale pretrained models (OpenShape, Uni3D, ReCon++, DuoduoCLIP) excel on clean synthetic benchmarks such as ModelNet40 but degrade sharply on real-world scans (ScanObjectNN, ScanNet) and corrupted point clouds (ModelNet-C)–a persistent real-world generalization gap. We propose SVG-3D, the first 3D extension of Support Vector Generation (SVG), to close this gap without retraining a more capable encoder or accessing target-domain test data. SVG-3D mines a discriminative kernel-machine support set from a frozen encoder paired with DiffSplat, a text-conditioned 3D Gaussian splatting generator. It rests on two design choices that preserve strict zero-shot rigour: (i) a dynamic pair-selection rule that picks the top-$K$ most confusable class pairs from a pseudo-confusion matrix computed entirely on generated samples, never on the test set; and (ii) a Metropolis–Hastings sampler operating on the CLIP text-embedding hypersphere with a Slerp-plus-Gaussian proposal whose density admits a closed-form ratio, yielding the same acceptance rule as SVG. On ScanObjectNN (OBJ\_ONLY/OBJ\_BG/PB\_T50\_RS) and ScanNet, SVG-3D surpasses prior zero-shot 3D classifiers built on large-scale pretrained models without consulting any test data; on ModelNet-C, a complementary corruption-aware setting in which the corruption family is assumed known further improves mean robustness over the same baselines.

In this paper, we introduce SP$^3$, a novel Plug-and-Play algorithm that accelerates Maximum \emph{a posteriori} image restoration by replacing denoisers with Spherical Encoders (SE) as generative priors. SP$^3$ approximates the intractable proximal prior step by utilizing the SE tightly structured latent space as a robust projection onto the natural image manifold. Alternating this projection with a closed-form data-consistency step, via Half-Quadratic Splitting, guarantees stable convergence without requiring gradient computation during inference. This unique formulation unlocks ``anytime'' restoration capabilities, producing sharp, plausible images from the first iteration. Evaluations across a variety of image restoration tasks demonstrate that SP$^3$ achieves perceptual quality comparable to state-of-the-art zero-shot diffusion and flow methods while being $3$-$630\times$ faster.

How fast *could* a deep-learning model run on target hardware, and how far is today’s implementation from that limit? These questions are central to software, hardware, and algorithm optimizations. Speed-of-Light (SOL) analysis answers them by computing a workload’s theoretical minimum execution time on a given architecture. Yet deriving SOL bounds remains manual, error-prone, and disconnected from rapid model development. To close this gap, we introduce SOLAR, a framework that automatically derives validated SOL bounds from Pytorch and JAX source code. SOLAR leverages both *generative* and *deterministic* components in its flow: an LLM frontend translates any source programs into an executable Affine Loop IR, validated by output comparison; a deterministic flow lifts the IR into an einsum graph; and an analytical backend computes unfused, fused, and cache-aware SOL bounds. SOLAR provides comprehensive operator and language coverage, produces validated bounds with zero observed SOL violations, and offers multi-fidelity analysis that tightens bounds and surfaces optimization insights. We evaluate SOLAR across KernelBench, JAX/Flax models, and robotics workloads. These experiments demonstrate four use cases: headroom analysis at multiple fidelity levels, identifying optimization opportunities, cross-platform exploration, and inverse-roofline hardware provisioning.

Diffusion models achieve strong performance across image and audio synthesis, but their inference remains computationally expensive, requiring tens to hundreds of sequential evaluations of a large denoising network. We observe that the effective rank of the denoiser output varies sharply across timesteps and architectures, motivating step- and block-wise adaptation of computation. We propose Step-Adaptive Rank Adjustment (SARA), a training-free framework that treats the SVD rank of pretrained linear layers as an inference-time control variable, adapted per step and block by SARA-DP and per step by SARA-Online. SARA introduces the Low-Rank Approximation Error (LRAE) — the discrepancy a rank-r truncation induces against the full-rank network — and instantiates it in two algorithms: SARA-DP measures LRAE on a few warmup prompts and solves the globally optimal rank assignment under a MACs budget by dynamic programming; SARA-Online computes a closed-form, spectrum-based LRAE from the model output at inference time, requiring no calibration and no trajectory lookahead. SARA applies uniformly to UNet (SDXL), MMDiT (SD3), DiT (Stable Audio Open), and hybrid MMDiT/DiT (TangoFlux) backbones, spanning ϵ, v, and rectified-flow parameterizations. SARA combines with existing solver- and cache-axis acceleration to yield compounded wall-clock speedups in our experiments. Together, these results introduce rank as a new axis of diffusion inference acceleration.

Disentangled representation learning (DRL) aims to identify and decompose the interpretable underlying factors of observations. However, most DRL methods rely on independence-oriented regularization to pursue statistical factorization, leaving learned factors without explicit semantic grounding. Such independence-oriented regularization induces an intrinsic trade-off between disentanglement and reconstruction. To address these issues, we establish a spectral learnability threshold for independence-oriented DRL, theoretically revealing a mismatch between the factors favored by the learning objective and human-perceivable semantics. We further introduce SAGE, a semantic-guided DRL framework that leverages MLLMs to automatically extract semantic signals from unlabeled data and uses a triangular latent-geometry constraint to anchor each semantic factor to a designated latent dimension, enabling factor-wise control over human-perceivable semantic variations in generated images. Extensive experiments support our theoretical analysis and demonstrate that SAGE achieves state-of-the-art performance across multiple benchmarks. In addition, our framework exhibits high flexibility and can be integrated with most advanced generative models.

A two-branch routed regressor realizes exactly the function class of ordinary quadratic regression. Because routing does not enlarge the observable function class here, the routed model differs from quadratic regression only through the prior induced by routed coordinates. We compute this induced prior in closed form and show that fiber integration along the routing map creates a logarithmically singular density on the linear submodel. The induced posterior over quadratic coefficients then coincides exactly with the quadratic-regression posterior under the pushed-forward prior, isolating the role of prior geometry in this proxy model. For compactly supported priors on linear truth ($q_2=0$), the support-relative global real log canonical threshold (RLCT) is $(3/2,2)$ when the support reaches the collapsed stratum and $(3/2,1)$ when it avoids it; under the full-support Gaussian-coordinate prior, the global RLCT is $(3/2,2)$. Against smooth coefficient-space baselines with the same known noise variance on linear truth, this singular-versus-smooth prior comparison yields a rigorous $\log\log n$ Bayes-factor advantage. The same multiplicity-$2$ product singularity transfers locally to sigmoid gates on linear truth and extends locally to $d$-dimensional inputs.

Diffusion models have emerged as a powerful tool across diverse domains. However, their purely data-driven nature can produce samples that violate domain-governing constraints. We introduce a plug-and-play Reinforcement Learning framework that optimizes initial noise samples in the latent space of frozen, pre-trained diffusion models. Leveraging the near-spherical geometry of high-dimensional Gaussian distributions, we introduce a novel rotation-matrix-based scheme for efficient latent space exploration. This steers the model toward more feature-preserving outputs, guided by task-specific rewards. We evaluate our method on three diffusion models: one trained on solutions of the Darcy Flow PDE, one on a synthetic dataset with complex structural features, and a text-conditioned one. Across all three settings, our framework yields significant improvements in sample quality, achieving a ${\sim}25\\%$ relative reduction in PDE residual, up to a ${\sim}44\\%$ relative improvement on the synthetic dataset's feature-alignment metric, and up to a ${\sim}80\\%$ relative improvement on human preference, compared to the vanilla diffusion models. Finally, we show that rotation-matrix-based exploration significantly outperforms unconstrained exploration, validating our geometry-aware approach and establishing a more effective method for latent space control.

Flow Matching (FM) is a strong framework for generative modeling due to its stable training and efficient sampling. However, FM assumes clean target data and labels, an assumption often violated in practice. We formulate FM training as a time-dependent regression problem over conditional vector fields and analyze how the two corruption types affect the objective: i) target corruption adds a time-dependent residual term that grows with time, ii) label noise causes updates to the wrong class-conditional vector field. Based on this analysis, we propose to improve FM robustness by mapping corruption-specific residual scores to sample weights. For target corruption, we propose the \emph{Temporal Residual Score} (TRS), which emphasizes residual magnitudes at intermediate-to-late FM times, where clean and corrupted samples are more separable. For label noise, we propose the \emph{Comparative Conditional Score} (CCS), which compares the residual of the observed class-conditional vector field against alternative class-conditional vector fields for the same trajectory and velocity target. Experiments across controlled corruption settings and real-world noisy labels show that the proposed methods improve robustness over standard FM and robust-training baselines, supporting corruption-specific residual structure as a practical basis for robust FM.

Flow matching has recently emerged as a powerful framework for continuous-time generative modeling. However, when applied to long-tailed distributions, standard flow matching frameworks are susceptible to *majority bias*, over-representing dominant modes at the expense of tail-class fidelity and distributional accuracy. In this work, we propose ***Unbalanced Optimal Transport Reweighted Flow Matching (UOT-RFM)***, a novel framework for generative modeling under class-imbalanced (long-tailed) distributions that operates without any class label information. Our method constructs the conditional vector field using mini-batch Unbalanced Optimal Transport (UOT) and mitigates majority bias through a principled inverse reweighting strategy. The reweighting relies on a *label-free majority score*, defined as the density ratio between the target distribution and the UOT marginal. This score quantifies the degree of majority based on the geometric structure of the data, without requiring class labels. By incorporating this score into the training objective, UOT-RFM theoretically recovers the target distribution with first-order correction ($k=1$) and empirically improves tail-class generation through higher-order corrections ($k > 1$). Our model outperforms existing flow matching baselines on long-tailed benchmarks, while maintaining competitive performance on balanced datasets.

A reward is not a universal RL interface: it becomes a valid update only through the probability object exposed by the policy branch. Autoregressive post-training works cleanly because token ratios are exact, but parallel discrete policies hide within-step dependence and flow/diffusion policies often lack cheap density ratios. We introduce MindRL, a reward-to-update interface controller that translates a shared reward through branch-native score objects while turning factorization, tractability, drift, and smoothness barriers into budgets over block size, serialization, anchors, clipping, reranking, and branch weights. As a closed-loop controller, MindRL makes the shared reward actionable for AR, parallel discrete, flow/diffusion, and AR+flow policies by selecting branch-native update objects and adapting their control budgets. Across language-side AR/parallel-discrete tests and matched flow/diffusion and AR+flow evaluations, MindRL improves task-quality or reward-risk tradeoffs while preserving each branch's native structure.

While Value Iteration (VI) is one of the most fundamental algorithms in Reinforcement Learning, its theoretical convergence guarantees still exhibit a persistent mismatch with empirical behavior. In the discounted-reward case, classical theory guarantees geometric convergence with rate $\gamma$, while in the average-reward case recent work suggests that only sublinear convergence can be expected. In practice, however, VI is often observed to converge significantly faster. In this work, we show through a unified geometry-based analysis that, under an assumption of a unique and unichain optimal policy, (i) convergence is geometric in *both* the discounted- and average-reward settings and (ii) the convergence rate is faster than previous analyses suggest.

We present a robust deep clustering framework for unsupervised visual discovery that bridges the gap between statistical robustness and mathematical tractability. Current generative approaches typically rely on either noise-prone Gaussian priors or computationally intensive diffusion models; however, neither paradigm simultaneously offers explicit latent structures and exact analytic optimization. While diffusion models generate strong priors, their reliance on iterative sampling and intractable likelihoods limits their scalability and interpretability in clustering tasks. Our approach addresses these limitations by constructing a latent space governed by a Student's $t$ distribution via a Bayesian scale-mixture formulation. This yields the first fully closed-form augmented evidence lower bound (ELBO) for this model class, effectively eliminating the biased variational approximations and high computational costs inherent in prior heavy-tailed or diffusion-based frameworks. By treating precision as a latent variable, our objective automatically modulates the influence of each data point; it downweights outliers to learn a data-dependent measure of trust. This mechanism leads to superior robustness and more stable manifold learning compared to state-of-the-art approaches. We demonstrate the resilience of our framework on standard vision benchmarks and complex neuroimaging tasks. In real-world brain magnetic resonance imaging (MRI)-driven neurodegenerative disease analysis, our model successfully recovers clean and anatomically coherent clusters where clinical noise typically collapses rare subtypes. These results underscore the ability of our framework to preserve clinically crucial morphological subtypes, allowing precise discovery within clinical interventions.

Peer review at ML conferences increasingly relies on reviewer scores as the primary decision instrument. As submissions have scaled from thousands to tens of thousands per year, no systematic audit has examined whether this instrument functions uniformly across research areas, or whether acceptance outcomes are in practice shaped by forces that reviewer scores neither capture nor control. This position paper argues that acceptance outcomes are shaped by forces beyond reviewer scores, and that the underlying cause is a measurement design failure, not individual bias. When a fixed numerical scale aggregates quality judgments across communities with structurally non-uniform reviewer pools, absolute scores become incomparable across areas, and area chairs must substitute community priors for score-based decisions. Using ICLR 2021--2026 data covering 50,289 papers across 219 research topics, we show that at any given reviewer score, a paper's acceptance probability varies by up to 8$\times$ depending on its topic. We rule out scoring culture, expert reviewer standards, rational area chair reweighting, and quality dilution as alternative explanations. We call on program committees to adopt inherently calibrated review signals and publish topic-stratified, score-conditional acceptance rates as a first-class fairness metric.

Taming diffusion models for generative segmentation has attracted increasing attention. While existing approaches primarily focus on architectural tweaks or training heuristics, there remains a limited understanding of the intrinsic mismatch between continuous flow matching objectives and discrete perception tasks. In this work, we revisit diffusion segmentation from the perspective of vector field learning. We identify two key limitations of the commonly used flow matching objective: gradient vanishing and trajectory traversing, which result in slow convergence and poor class separation. To tackle these issues, we propose a principled vector field reshaping strategy that augments the learned velocity field with a detached distance-aware correction term. This correction introduces both attractive and repulsive interactions, enhancing gradient magnitudes near centroids while preserving the original diffusion training framework. Furthermore, we design a computationally efficient, quasi-random centroid encoding scheme inspired by Kronecker sequences, which integrates seamlessly with an end-to-end pixel neural field framework for pixel-level semantic alignment. Extensive experiments consistently demonstrate significant improvements over vanilla flow matching approaches, substantially narrowing the performance gap between generative segmentation and strong discriminative specialists.

Semantic ID-based generative recommendation represents each item as a sequence of discrete tokens, enabling structured modeling of item semantics. A critical challenge is constructing semantic IDs that are both semantically expressive and computationally efficient. While recent approaches favor complex learned quantization, simple hashing-based methods such as SimHash are widely regarded as fundamentally inferior. In this work, we challenge this consensus by showing that the apparent performance gap does not stem from inherent limitations of hashing, but rather from a structural mismatch with autoregressive decoding, coupled with the inevitable information loss during rigid discretization. Based on this insight, we propose FLASH, a two-stage framework that revitalizes training-free SimHash tokenization through parallel decoding and explicit semantic alignment. Despite its simplicity, FLASH achieves state-of-the-art performance across multiple datasets without requiring any tokenizer training, while exhibiting stronger generalization in cold-start scenarios. Notably, we demonstrate that semantic alignment acts as a universally effective mechanism across diverse paradigms. Our findings suggest that, with compatible decoding and semantic grounding, simple and efficient tokenizers can outperform complex learned counterparts. Our code and data are available: https://anonymous.4open.science/r/Flash-014A/.

Sparse autoencoders (SAEs) were originally motivated by discovering monosemantic features, a property that characterizes individual features but does not specify what the learned feature set should look like as a whole. As SAEs have increasingly been adopted for attribution, circuit discovery, and intervention, researchers have begun evaluating metrics that describe global properties of the feature set, such as sparsity, dead feature rates, reconstruction fidelity, and disentanglement, a diverse set that extends far beyond monosemanticity. However, this implicit shift beyond monosemanticity has not been formalized as a clear, well-defined objective that comes before these evaluation metrics. Without such an objective, existing metrics remain ad hoc, with their necessity unjustified. In this work, we formalize this ultimate objective by introducing a new paradigm for SAE evaluation: requiring that SAE features, and more generally dictionary-based decompositions of model representations, form an atomic interpretable unit basis, a coordinate system for semantically meaningful explanation. Under this paradigm, we organize existing and new measurements such as DAME (Dominance-Aware Monosemanticity Evaluation) as proxy metrics and introduce a comprehensive benchmarking framework. Evaluating hundreds of SAEs across multiple architectures and model families, we uncover failure modes in which some SAEs appear monosemantic under existing evaluation metrics while remaining functionally polysemantic: the most dominant features, those contributing most to model computation, are systematically less interpretable and more redundant. These failure modes warrant attention from SAE researchers as targets for improved training and architectures, and motivate dominance-aware evaluation as a standard check for practitioners before SAE features are deployed for attribution, circuit tracing, or intervention.

Standard generative models face a memorization--generalization trade-off, that is avoiding memorization is considered necessary for generalization. In supervised learning, however, recent studies show that highly overparametrized models can memorize (interpolate) training data while still generalizing well, a phenomenon known as benign overfitting. Motivated by this, we investigate whether generative models can similarly bypass the trade-off and generalize while interpolating. Generative models should minimize the distance between the true data distribution and the distribution induced by mapping the full latent distribution through the generator. But since the true distribution is inaccessible, existing models instead minimize empirical risk with respect to the training distribution. Under this standard formulation, exact empirical-risk minimization forces the generator to produce only training samples. To address this issue, we consider an alternative empirical risk based on presampled latent variables. Our experiments demonstrate that the resulting presampled-latent version of the flow matching model exhibits benign overfitting on standard image benchmarks, such as MNIST and CIFAR-10. As a theoretical proof of concept, we recast generative modeling as a regression problem and extend existing benign overfitting theory to our setting. Together, these results establish, for the first time, that benign overfitting can occur in generative models.