NeurIPS 2026 Papers

Sparse Autoencoders (SAEs) are increasingly used to interpret foundation models, but their role as an actionable intervention space remains less understood, especially in vision. We study whether sparse visual features can be used not only for post-hoc analysis, but also to steer frozen vision-language models. We introduce \textbf{Visual Sparse Steering (\ours)}, a label-free method that trains a top-$k$ SAE on unlabeled activations from a frozen CLIP image encoder and, at test time, constructs an interpretable steering vector by amplifying the input's active sparse features and decoding the induced change. We show that this procedure admits a closed-form decomposition as \emph{centroid-deviation steering}: each input is moved along its deviation from the SAE-learned centroid. The residual term is controlled exactly by the SAE's per-sample reconstruction error, measured by FVU, yielding an FVU-based residual bound and motivating a reliability gate that falls back to zero-shot CLIP when SAE reconstruction is unreliable. With target-domain SAEs trained on unlabeled CLIP image-encoder activations, \ours improves zero-shot accuracy across nine image-classification datasets, achieving gains up to $+4.12\%$ with less than $0.1\%$ additional inference compute. Finally, a controlled upper-bound study, \oursplus, shows that selective amplification of sparse features can yield gains up to $+21.44\%$, exposing a \emph{reconstruction-vs-task saliency gap}: features salient for reconstruction need not align with features useful for downstream prediction.

Conformal prediction (CP) provides a distribution-free framework for uncertainty quantification with finite-sample marginal coverage guarantees. Yet, missing values introduce distributional heterogeneity, where marginal coverage can hide systematic undercoverage for certain missingness patterns. At the same time, existing methods for conditional guarantees often fail when missing values are present, as they rely on geometric continuity of the covariate space, which is broken by missingness. Although imputation restores continuity, it obscures the missingness patterns and introduces task-dependent overhead. We develop a new CP framework that is \emph{adaptive to missingness patterns}, without imputing covariates. We rely on data augmentation by feature embedding that accounts for missingness. Our framework represents each input with missing values by a tree embedding together with its raw missingness mask, and then learns a regularized quantile score function. Varying the choices for the penalty function and the augmentation enables us to characterize three practically relevant validity levels: exact coordinate-wise validity, exact validity on observed masks, and approximate missingness-conditional guarantees for general mechanisms (MCAR, MAR, MNAR). Experiments on synthetic and real datasets support these guarantees and show that the proposed method reduces miscoverage risk while maintaining informative intervals, especially in non-MCAR settings. These results highlight the practical value of uncertainty quantification that adapts to missingness.

Offline evaluation of ad ranking policies in multi stage delivery systems is fundamentally challenging because action propensities are often unavailable, limiting inverse propensity scoring (IPS) estimators. While reward model based estimators such as Direct Method (DM) do not rely on propensity estimates, they can suffer from bias when candidate policies induce impression distributions that differ from the logging mixture. In this work, we formalize DM based counterfactual replay under mixture logging and propose a mixture targeted shift correction based on density ratio weighting through domain classifier. We further derive a per policy value estimation error and characterize an asymptotic error ceiling governed by the cold start mass. Experiments on both ads delivery simulator and one week of production traffic from a large scale industrial CTR ranking system demonstrate that the proposed mixture weighted reward model consistently outperforms its unweighted counterpart under distribution shift.

Foundation models for predictive machine learning on tabular data have recently gained significant traction in academia and industry. Research communities across disciplines are increasingly evaluating tabular foundation models on diverse datasets and tasks. However, these task- and discipline-specific evaluations remain largely inaccessible to model researchers because benchmark software and evaluation protocols are fragmented. As a result, model researchers rely on standard benchmarks, which are mostly defined for tasks where tabular foundation models already excel. The most challenging scenarios are excluded, limiting meaningful progress in the field by focusing on marginal improvements on IID data rather than on broader, more demanding challenges. To overcome this, we introduce BeyondArena, the first unified holistic benchmark for tabular data that supports diverse task types (IID, temporal, grouped), across sample size and feature dimensionality scales, with diverse feature types (with text, with high cardinality) from a broad range of disciplines. To enable unified benchmarking beyond standard benchmarks, we introduce DataFoundry, a Python framework and metadata schema for curating tabular datasets for predictive machine learning. Our results across 11 models and 142 curated datasets show that existing tabular foundation models excel on tiny- to medium-sized IID data, while traditional tree-based and deep learning models still dominate on non-IID, large, and high-dimensional datasets. BeyondArena guides model research for the most demanding challenges in tabular data, enabling progress towards truly foundational tabular models.

We prove the identifiability of deep generative models (DGMs) with Piecewise-Affine (PWA) decoders and Gaussian Mixture Model (GMM) priors, in a purely unsupervised setting. We introduce three algebraic contrast principles for symmetry breaking: *domain contrast*, which trivializes the mixture symmetry group; *mechanism contrast*, which ensures every decoder branch is witnessed by a unique boundary; and *interaction contrast*, which forbids parameter conspiracies between latent components and decoder branches. Together they exploit the interplay between the discrete combinatorics of the PWA map and the continuous symmetry structure of the latent GMM. Continuity is replaced by algebraic symmetry conditions; injectivity is decoupled from structural identification and required only for pointwise inversion. Our results form a hierarchy: from *law identifiability* (LID; latent distribution up to a global affine map) through *map identifiability* (MID; decoder up to the same map) to *posterior* and *pointwise identifiability*. The ICA-form ambiguity emerges under diagonal covariance conditions; classical ICA is a further specialization under independence. Assumptions are only on the data-generating process, not on learning methods, except for the interaction contrast.

Large language model agents are increasingly used as synthetic participants in social, behavioural, and decision-making research, yet it remains unclear whether they can express graded cognitive diversity rather than collapsing toward generic competence or stylised persona descriptions. We introduce a cognition-suite evaluation for LLM persona simulation, testing whether models reproduce population-level variation across structured reasoning tasks, self-report cognitive/personality measures, cross-construct associations, and downstream behavioural reasoning. Agents are conditioned on synthetic or human-derived trait profiles spanning cognitive reflection, Need for Cognition, and Big Five personality, then evaluated on CRT-style reasoning items, trait questionnaires, and Behavioural Reflection Task scenarios requiring evidence weighting, risk evaluation, and socially embedded justification. Across four matched simulation runs, agents partially recover questionnaire-style Need for Cognition and Big Five structure, especially under numeric prompting, but fail to preserve cognitive-reflection fidelity: CRT-style scores collapse toward correctness, intuitive-lure errors are almost eliminated, and reflection-related association structure weakens. Downstream BRT responses show the same pattern of flattening: open-ended decisions compress into over-regularised behavioural profiles rather than preserving human-like variation. The evaluation reframes persona simulation as a problem of cognitive-diversity fidelity: useful synthetic populations must preserve not only demographic or personality descriptors, but the structured variation in reasoning style, error type, and behaviour that those descriptors are meant to organise.

As LLMs increasingly serve in advisory and deliberative roles, users rely on them for non-verifiable reasoning in domains lacking objective ground truths. However, traditional evaluations of LLM reasoning focus almost exclusively on fact-based domains, such as mathematics and science, leaving uncertainty over whether and to what degree models can handle ambiguous, subjective, or value-laden problems over time. To address this concern, we propose moral reasoning as a paradigmatic subdomain of non-verifiable reasoning. We define moral robustness as a model’s capacity to exhibit sound moral reasoning across time and contexts, and we introduce a scalable, adversarial, multi-turn evaluation framework to empirically measure this capability. We simulate 48,000 user-agent moral deliberations across four frontier LLMs, varying premise relevance, premise order, conversation duration, and the user's stated moral view. We find that models successfully ignore morally-irrelevant distractors, but shift their reasoning by up to 6.5\%, on average, towards the user's stated preferred moral view, and varying their reasoning depending on factors such as order (altering moral judgments by order in 13-22\% of the cases) and duration (altering moral judgments between single-turn and multi-turn in 10-24\% of the cases). Our analysis indicates that models tailor not just their final verdicts but their underlying justifications to align with a user’s moral viewpoint — a failure mode we characterize as moral deliberative sycophancy.

Despite remarkable progress in image-text understanding, vision-language models (VLMs) still struggle with compositional reasoning. In particular, they often fail to distinguish relational direction and attribute-object binding, leading to similar representations for semantically different image-text pairs. This problem mainly stems from the reliance on global image-text alignment, which captures coarse correspondence but overlooks fine-grained compositional structures. Toward this end, we propose an evidence-aware framework, termed Directional Relation Bucketing with Binding Localization (DELTA), to improve compositional understanding in VLMs. The key idea of our DELTA is to improve compositionality of VLMs from two complementary perspectives: direction modeling and binding localization. More specifically, we first leverage a learnable gate to model the cumulative contextual distance between anchor terms, thereby assigning relation concepts to direction-aware discrete buckets. To enrich textual representations with fine-grained relational semantics, we calibrate global semantic representations by incorporating intermediate-layer hidden states. Furthermore, DELTA leverages textual cues to ground visual evidence for object attributes, pulling image patches closer to their matched attribute descriptions while pushing them away from incorrect augmented ones. Extensive experiments on four compositional datasets demonstrate the effectiveness of the proposed DELTA. Our implementation is available at https://anonymous.4open.science/r/DELTA-2D60.

Recent advancements in generative models have increasingly leveraged self-supervised visual representations to guide the synthesis process, significantly improving generation quality and semantic consistency. However, the reverse paradigm of harnessing the generative process to enhance visual representations remains largely underexplored. In this paper, we investigate the semantic properties of class tokens synthesized by generative models. We observe that generative models equipped with representation entanglement can generate class tokens that exhibit stronger discriminative capabilities than those extracted from the original pretrained visual models. Motivated by this observation, we propose a completely new Generation-to-Perception Knowledge Distillation (GPKD) framework, where our method generates class tokens as teacher signals to instill global semantics into the student model. To prevent the degradation of local details, we further incorporate a new masked patch-level distillation objective. This dual distillation strategy enhances global representations while mitigating the forgetting of local details, thereby producing more robust representations. Extensive experiments demonstrate that GPKD obtains consistent improvements across various downstream tasks compared with the base visual encoder, achieving a gain of more than 2\% in $k$-NN accuracy on ImageNet-1K.

Autoregressive graph generation is commonly formulated by linearizing a graph into a sequence via a flat walk over its nodes and edges. While effective for small molecular graphs, such representations do not scale gracefully to large, structurally complex biological graphs, where higher-order organization is both dense and functionally important. To address this limitation, we introduce a hierarchical graph abstraction together with a coarse-to-fine autoregressive pipeline that separates coarsening, tokenization, and decoding, and we study sequence constructions that expose hierarchy in the tokenization. Concretely, we compare a flat representation with two hierarchical alternatives: one that explicitly factorizes communities and their interactions, and another that induces hierarchy through a depth-first traversal. Across eight benchmarks, differences are modest on simpler domains but become pronounced on larger, more structurally complex settings such as cyclic peptides, proteins, and hard conditional-generation tasks. In these regimes, flat representations degrade in validity or distributional fidelity, while the traversal-based hierarchical representation is consistently more reliable; the explicitly factorized variant is generally weaker, suggesting that how hierarchy is exposed in the sequence matters. Finally, the same coarse-to-fine representation naturally enables motif-conditioned generation without architectural changes: on larger and more structurally complex graphs, conditioning on higher-level prefixes yields strong motif retention together with high validity, uniqueness, and novelty, whereas flat representations can exhibit severe mode collapse under comparable constraints.

As collaborative machine learning scales to thousands of edge devices, two failure modes emerge: centralized federated learning creates server bottlenecks and single points of failure, while flat peer-to-peer systems force every node to increase its communication degree as the network grows. We propose Tiered Gossip Learning (TGL), a two-layer push–gossip–pull protocol that decouples data-holding leaf nodes from a decentralized relay backbone. In each round, leaves push local models to a subset of relays, relays gossip among themselves to mix models globally, and leaves pull updates from another random relay subset. This asymmetric design keeps per-leaf communication fixed as the network scales, offloading the mixing burden to a thin layer of higher-capacity relays without relying on any central coordinator for aggregation. Across CIFAR-10, FEMNIST, and AG News, TGL matches or exceeds baseline accuracy with up to 80\% fewer model exchanges. We provide convergence guarantees under standard smoothness, bounded variance, and heterogeneity assumptions, with explicit stage-wise consensus bounds characterizing how relay-layer connectivity governs the global mixing rate.

Recent studies have shown that leveraging the reasoning and analytical capabilities of large language models (LLMs) for long video understanding has become a promising approach. However, these methods are constrained by a structural limitation: representing video as a flat frame sequence makes it difficult to model the temporal structure and logical dependencies of events, which in turn hampers long-range reasoning and risks overlooking important information. To address this limitation, we propose LVGraph, a framework that models video content using a coarse-to-fine hierarchical graph. LVGraph constructs a semantic hierarchy where high-level graphs characterize the global event structures within the video, while low-level graphs capture fine-grained, query-relevant information. Instead of linear scanning, reasoning is performed via a query-aware traversal of this graph, adaptively identifying and retrieving the most salient keyframes by navigating the nodes and edges most relevant to the query. Finally, we leverage a Vision Language Model (VLM) to augment the keyframe information, and subsequently answer the question by reasoning over the graph and the enriched keyframe captions. Extensive experiments on long video understanding benchmarks confirm the effectiveness of our method. LVGraph significantly outperforms existing LLM-based state-of-the-art methods on EgoSchema, NExT-QA, and long-duration segments (averaging 44 minutes) of the Video-MME benchmarks.

Evaluating large language models (LLMs) today rests on *fixed benchmarks* that apply the same set of items to any model, producing ceiling and floor effects that mask capability gaps. We argue that the most informative evaluation signal lies at the *boundary*, where the per-prompt pass probability is near 0.5, and propose **Dynamic Boundary Evaluation (DBE)**, which actively locates each model's boundary and places it on a globally comparable difficulty scale. DBE delivers three artifacts: (i) a calibrated item bank covering safety, capability, and truthfulness, with per-item difficulty labels validated across 9 reference LLMs; (ii) **Skill-Guided Boundary Search (SGBS)**, a search algorithm that finds boundary items for a given target LLM using only API-level query access; and (iii) an evaluation protocol that places a new LLM on a unified ability scale and grows the evaluation set adaptively when the target falls outside the bank's coverage. We instantiate DBE on four categories spanning safety (harmful refusal, over-refusal), capability (constrained instruction following), and truthfulness (multi-turn sycophancy resistance). The resulting evaluation covers a broader model spectrum without saturation while remaining compatible with existing datasets.

Unlearnable Examples (UEs) have emerged as a promising data protection strategy against unauthorized model training, which adds imperceptible perturbations into data to degrade generalization. Existing studies predominantly assume a fully trainable target model, where unlearnability is achieved by disrupting clean feature representations. However, this assumption grows increasingly unrealistic with the prevalence of pretrained models, where unauthorized users can freeze the backbone and perform linear probing, preserving discriminative representations and thereby invalidating the core mechanism of previous UEs. In this paper, we investigate this practical yet challenging linear probing setting and reveal a fundamental vulnerability. To address this, we propose $\textbf{DIVERT}$ ($\textbf{D}$ecision-boundary d$\textbf{I}$version $\textbf{V}$ia s$\textbf{E}$mantic-decoupled o$\textbf{R}$thogonal $\textbf{T}$argets), a novel framework that shifts the design principle from feature disruption to explicit decision boundary diversion. Our intuition is that projecting perturbed samples into a subspace orthogonal to the semantic manifold induces spurious linear separability, thereby steering the decision boundary away from original prototypes. Building on this principle, DIVERT first constructs synthetic class anchors within the semantic-decoupled subspace, then employs a cross-attention mechanism to jointly optimize these anchors while pushing perturbed samples toward them to establish spurious separability. Subsequently, it incorporates a surrogate linear head to simulate the linear probing procedure, further refining perturbations to actively divert the decision boundary. Extensive experiments demonstrate that DIVERT establishes effective unlearnability across diverse datasets and backbones.

Early-exit networks promise inference savings by terminating computation at intermediate classifiers, but FLOPs-based gains rarely translate into on-device latency reduction under static-graph compilation. On NVIDIA Jetson Orin Nano, the per-sample exit policy of Shallow-Deep Networks (SDN) runs approximately 2× slower than vanilla ResNet-56 and misses the 60 fps deadline despite using only 40% of the FLOPs, because dynamic branching precludes layer fusion. We propose the Selective Lightweight IC Ensemble Network (SLIENet), a train-full, deploy-partial framework compiled as a single static FP16 engine. SLIENet trains the full backbone with light self-distillation, using the final classifier as a stop-gradient teacher to preserve error diversity across internal classifiers (ICs). At deployment, a sub-second calibration search selects an IC subset for softmax averaging, while overthinking late blocks are truncated at depth k. On CIFAR-100 across ResNet-56, VGG-16, and MobileNetV1, SLIENet outperforms SDN-based variants and remains competitive with ZTW. On Jetson Orin Nano, the recommended SLIE5 configuration improves accuracy by +2.46 pp while reducing p50 latency by 14% and energy by 15% over vanilla, with zero deadline misses across 10,000 test samples. Thus, multi-exit deployment should be a pre-compiled deterministic inference plan, not a sample-wise routing policy.

Monte-Carlo Tree Search (MCTS) typically stores only scalar mean estimates, which makes it poorly suited for tail-risk objectives such as conditional value-at-risk (CVaR). We introduce **Categorical Thompson Sampling with Optimism (CATSO)** and **Particle Thompson Sampling with Optimism (PATSO)**, two distributional MCTS algorithms that maintain finite-support empirical backup laws at Q-edges. CATSO represents Q-edge laws with categorical atoms and Dirichlet Thompson sampling, while PATSO represents them with adaptive particles. Both algorithms select actions using the CVaR of a Thompson-sampled Q-edge law plus a polynomial optimism bonus, and propagate scalar continuation values through visit-weighted averages of child Q-edge CVaR scores. For the finite-depth nested CVaR planning target, we prove an $O(n^{-1/2})$ bound on the root value error after $n$ rollouts, which upper-bounds the risk-sensitive simple regret. CATSO adds a fixed-grid discretization bias, while capped PATSO adds a tunable Wasserstein compression bias. Experiments on stochastic planning benchmarks show that distributional Q-edges improve lower-tail return and reduce catastrophic outcomes when mean-optimal and risk-sensitive behavior differ.

Replay is a standard strategy for mitigating catastrophic forgetting in continual learning, where a bounded memory buffer stores exemplars from previously seen classes. Existing replay methods have largely focused on instance-level decisions, such as which samples to store, replay, or replace, while the class-wise allocation of memory slots is often treated as a uniform heuristic. Here, we study class-wise memory allocation as a complementary design axis and propose VAMA, a value-aware memory allocation framework. VAMA keeps cross-task budgets balanced, estimates task-local class values from Shapley-inspired sample utility scores, and converts them into class-wise quotas through a regularized allocation rule that conservatively deviates from class-balanced replay. This design changes only the class-wise budgeting rule and leaves within-class exemplar selection unchanged, making VAMA easy to integrate into replay pipelines. Experiments on CIFAR-100 and ImageNet-100 show that VAMA consistently improves ER, iCaRL, and FOSTER across different memory budgets and task protocols. Further analyses show that the improvements are not due to non-uniform allocation alone, and that allocation-relevant class-value rankings stabilize early in training, enabling efficient value-aware allocation. These results suggest that class-wise memory allocation is a consequential design choice in replay-based continual learning.

As foundation models scale toward fusing more heterogeneous visual streams, understanding how diverse encoders interact under joint training becomes a prerequisite for principled design. Yet large vision-language models (LVLMs) currently lack the tools to do so, and parameter-efficient encoder configurations remain hard to identify before training. To re-examine encoder roles under joint training, on the 16-benchmark Cambrian-1 suite we retrain and evaluate all 31 non-empty subsets of five common vision encoders under a unified pipeline ($\approx$20k GPU-hours total), and report three findings. First, retraining each subset from scratch reveals encoder rankings that differ from those obtained by masking encoders on a fixed checkpoint, including which encoder ranks first overall. Second, we decompose each encoder's contribution into two axes, \emph{Capacity}, the score an encoder reaches on its own, and \emph{Necessity}, the drop when it is removed from the full pool. The two axes are not interchangeable. Pairing the two highest-Capacity encoders is suboptimal, while pairing a high-Capacity anchor with an adaptive complement matches the full five-encoder model. Adding further encoders beyond this pair yields only marginal gains. Third, at fixed parameter count, per-encoder pre-projector effective rank explains the residual score variation. The strongest pairs combine an anchor whose rank survives joint training with a complement whose rank \emph{expands} under it, suggesting that higher-rank, less-collapsed projector inputs correspond to a more favorable optimization regime at the encoder–projector interface. Together, the Capacity–Necessity decomposition and the pre-projector rank analysis, along with comprehensive evaluation through retraining, expose a methodological gap in multi-encoder LVLM design, and offer concrete primitives for closing it.

Outliers are essential for evaluating and improving the robustness of machine learning systems, especially when future distributions may differ significantly from historical training data. In high-stakes applications, robustness often depends on rare cases that finite datasets fail to capture, making simple resampling or perturbation insufficient for stress scenario generation. Existing outlier synthesis methods typically rely on sparse neighborhoods, low support latent regions, or classifier boundary crossings, which can be heuristic, unstable, and tied to specific modalities or architectures. We therefore propose Sinkhorn Boundary Outlier Generation (SBOG), a structured framework for latent-space outlier generation that couples Sinkhorn optimal transport geometry with distributionally robust boundary modeling. The resulting Sinkhorn-induced support cost guides the sampler toward weakly supported boundary regions, while semantic constraints prevent uncontrolled drift from the intended context, yielding controlled deviations from the in-distribution reference measure rather than arbitrary sparse-region samples. Experiments on time series anomaly generation and image outlier synthesis show that our framework produces informative, semantically controlled outliers and improves downstream robustness evaluation across modalities, providing a foundation for stress scenario generation beyond empirical support.