NeurIPS 2026 Papers

Multiple instance learning (MIL) learns from bag-level labels, but a bag’s label is often supported by only a few hidden instances. In heterogeneous bags, these useful instances are rare, diverse, and surrounded by many weak or noisy observations. We propose Sparse Support Grounding, a MIL framework that learns instance influence directly. The key idea is simple: an instance should not matter just because it can be matched somewhere; it should matter when the match provides reliable label evidence. Our method forms a compact, bag-specific evidence support, compares instances with label-consistent supports, and uses unbalanced transport to estimate how much source mass each instance should retain. This retained mass becomes the strength of its grounding signal, so reliable evidence shapes training more than nuisance content, while the bag classifier still uses the full bag. Thus, transport is turned from a matching rule into a way of learning which instances should drive supervision. Experiments on synthetic MIL, whole-slide images, and clinical endomicroscopy show consistent gains, with the largest improvements when sparse diagnostic evidence is embedded in heterogeneous nuisance content.

Inference in nonlinear continuous stochastic processes on trees is challenging, particularly when observations are sparse and the topology is complex. Exact smoothing via Doob's $h$-transform is intractable for general nonlinear dynamics. We propose Neural Backward Filtering Forward Guiding (NBFFG), a unified framework for both discrete transitions and continuous diffusions. Our method constructs a variational posterior by leveraging a proxy linear-Gaussian process. This proxy process yields a closed-form backward filter that serves as a guide, steering the generative path toward high-likelihood regions. We then learn a neural residual to capture the non-linear discrepancies. This formulation allows for an unbiased pathwise subsampling scheme, reducing the training complexity from tree-size dependent to path-length dependent. Empirical results show that NBFFG outperforms baselines on synthetic benchmarks, and we demonstrate the method on a high-dimensional inference task in phylogenetic analysis with reconstruction of ancestral butterfly wing shapes.

Estimating the generative mechanism of large-scale networks is a fundamental challenge in statistical machine learning. It requires the identification of the latent connectivity structure, which is in general an NP-hard combinatorial problem due to the absence of canonical node labels. We address this challenge by allowing for probabilistic couplings, thereby relaxing the assignment problem. Our estimation framework can be formulated as a semi-relaxed Gromov–Wasserstein objective and provides a low-dimensional representation of the generative structure. We solve this via a block-coordinate conditional gradient algorithm. Despite the relaxation, the resulting solution is typically deterministic: in fact, we show that the optimality gap between the relaxed solution and the deterministic assignment vanishes at rate $O(1/n)$, where $n$ is the number of nodes. This allows for tractable recovery of the underlying model and enables rigorous statistical analysis: we establish consistency and minimax-optimal convergence rates for both stochastic block models and smooth graphons. Our implementation scales efficiently with $n$, as demonstrated on both synthetic and real-world datasets.

Tabular Foundation Models have recently established the state of the art in supervised tabular learning, by leveraging pretraining to learn generalizable representations of numerical and categorical structured data. However, they lack native support for unstructured modalities such as text and image, and rely on frozen, pretrained embeddings to process them. On established Multimodal Tabular Learning benchmarks, we show that tuning the embeddings to the task improves performance. Existing benchmarks, however, often focus on the mere co-occurrence of modalities; this leads to high variance across datasets and masks the benefits of task-specific tuning. To address this gap, we introduce MulTaBench, a benchmark of 40 datasets, split equally between image-tabular and text-tabular tasks. We focus on predictive tasks where the modalities provide complementary predictive signal, and where generic embeddings lose critical information, necessitating Target-Aware Representations that are aligned with the task. Our experimental results demonstrate that the gains from target-aware representation tuning generalize across both text and image modalities, several tabular learners, encoder scales, and embedding dimensions. MulTaBench constitutes the largest image-tabular benchmarking effort to date, spanning high-impact domains such as healthcare and e-commerce. It is designed to enable the research of novel architectures which incorporate joint modeling and target-aware representations, paving the way for the development of novel Multimodal Tabular Foundation Models.

The open-source model ecosystem now contains hundreds of thousands of pretrained models, yet picking the best model for a new dataset is increasingly infeasible: new models and unbenchmarked datasets emerge continuously, leaving practitioners with no prior records on either side. Existing approaches handle only fragments of this in-the-wild setting: AutoML and transferability estimation select models from small predefined pools or require expensive per-model forward passes on the target dataset, while model routing presupposes a given candidate pool. We introduce ModelLens, a unified framework for model recommendation in the wild. Our key insight is that public leaderboard interactions, though scattered and noisy, collectively trace out an implicit atlas of model capabilities across heterogeneous evaluation settings, a signal rich enough to learn from directly. By learning a performance-aware latent space over model--dataset--metric tuples, ModelLens ranks unseen models on unseen datasets without running candidates on the target dataset. On a new benchmark of 1.62M evaluation records spanning 47K models and 9.6K datasets, ModelLens surpasses baselines that either rely on metadata alone or require running each candidate on the target dataset. Its recommended Top-K pools further improve multiple representative routing methods by up to 81\% across diverse QA benchmarks. Case studies on recently released benchmarks confirming generalization to both text and vision-language tasks.

Most FDR-controlled feature selection methods are designed for coordinate-wise hypotheses, where each feature has a single weight or importance score. This abstraction fails in sequential and grouped models, where one original feature is represented by a block of sub-features, such as lags, recurrent states, or attention-based interactions. We propose a grouped-feature FDR control framework for such settings. For grouped linear models, we construct null-symmetric block-level mirror statistics with matrix-valued perturbations. For neural sequential models, we combine Permutation SHAP derivatives as model-agnostic block-level importance scores with kernel-based dependence measure. The framework is model-agnostic across network architectures, does not require specifying the covariate distribution, and reduces to Gaussian Mirror or Neural Gaussian Mirror when the block size is one. We prove FDR control for low- and high-dimensional grouped linear models and asymptotic symmetry of smoothed Permutation SHAP derivatives under fixed fitted nonlinear models. Experiments on simulated and real-world datasets show reliable FDR control and improved power under correlated grouped-feature signals.

We provide theoretical guarantees for convergence of discrete-time policy mirror descent with inexact advantage functions updated using temporal difference (TD) learning for entropy regularised MDPs in Polish state and action spaces. We rigorously derive sufficient conditions under which the single-loop actor-critic scheme is stable and convergent. To weaken these conditions, we introduce a variant that performs multiple TD steps per policy update and derive an explicit lower bound on the number of TD steps required to ensure stability. Finally, we establish sub-linear convergence when the number of TD steps grows logarithmically with the number of policy updates, and linear convergence when it grows linearly under a concentrability assumption.

Tree ensembles such as random forests (RFs) and gradient boosting machines (GBMs) are among the most widely used supervised learners, yet their theoretical properties remain incompletely understood. We adopt a spectral perspective on these algorithms, with two main contributions. First, we derive minimax-optimal convergence for RF regression, showing that, under mild regularity conditions on tree growth, the eigenvalue decay of the induced kernel operator governs the statistical rate. Second, we exploit this spectral viewpoint to develop compression schemes for tree ensembles. For RFs, leading eigenfunctions of the kernel operator capture the dominant predictive directions; for GBMs, leading singular vectors of the smoother matrix play an analogous role. Learning nonlinear maps for these spectral representations yields distilled models that are orders of magnitude smaller than the originals while maintaining competitive predictive performance. Our methods compare favorably to state of the art algorithms for forest pruning and rule extraction, with applications to resource constrained computing.

We present the first mini-batch kernel $k$-means algorithm, offering an order of magnitude improvement in running time compared to the full batch algorithm. A single iteration of our algorithm takes $\widetilde{O}(kb^2)$ time, significantly faster than the $O(n^2)$ time required by the full batch kernel $k$-means, where $n$ is the dataset size and $b$ is the batch size. Extensive experiments demonstrate that our algorithm consistently achieves a 10-100x speedup with minimal loss in quality, addressing the slow runtime that has limited kernel $k$-means adoption in practice. We further complement these results with a theoretical analysis under an early stopping condition, proving that with a batch size of $\widetilde{\Omega}(\max \set{\gamma^{4}, \gamma^{2}}\cdot k\epsilon^{-2} )$, the algorithm terminates in $O(\gamma^2/\epsilon)$ iterations with high probability, where $\gamma$ bounds the norm of points in feature space and $\epsilon$ is a termination threshold. Our analysis holds for any reasonable center initialization, and when using $k$-means++ initialization, the algorithm achieves an approximation ratio of $O(\log k)$ in expectation. For normalized kernels, such as Gaussian or Laplacian it holds that $\gamma=1$. Taking $\epsilon = O(1)$ and $b=\Theta(k\log n)$, the algorithm terminates in $O(1)$ iterations, with each iteration running in $\widetilde{O}(k^3)$ time.

Do behavioral styles in large language models correspond to reusable directions in residual space, or do they mainly appear after averaging many noisy contrastive examples? We introduce MinSteer, a two-stage cached-continuation method for estimating a contrastive residual direction (CRD) within one decoding trajectory. Stage-X generates a continuation in one style; a style-flip suffix is then processed using Stage-X's final KV cache; Stage-Y continues from this inherited trajectory. The CRD is computed from generated-token residuals only, excluding prompt and suffix tokens from the average. Across sentiment transfer, politeness control, and a ParaDetox-based bidirectional diagnostic, MinSteer extracts effective steering directions from very small contrasts. On TweetEval sentiment transfer, one MinSteer contrast outperforms a CAA baseline averaged over 50 independent pairs. Japanese keigo CRDs transfer to English and Chinese outputs and outperform trajectory-free TwoSeq controls. In ParaDetox, MinSteer produces bidirectional Detoxify movement under our rewrite setup, while exposing semantic-preservation trade-offs. The directions transfer across tested scenarios and languages and show clear depth dependence. Together, these results show that cached self-rewrite trajectories provide a cleaner signal for residual-space behavior directions than independent-pair averaging.

Neural networks can fit noisy labels, but how their representations change as memorization begins remains unclear. We study this using persistent homology of penultimate-layer representations across training. Across 8 datasets and 8 architectures, we find a consistent pattern: the discriminative topological signal under noise appears in $H_1$ rather than $H_0$, with noisy representations preserving more loop structure than clean representations as training proceeds, supporting a folding view of memorization. When the clean topological signal decreases over training, the noisy signal does not decline as quickly as the clean one, producing an early crossover between clean and noisy trajectories; across the derivation set, crossover occurs when clean trajectories fade and the clean--noisy gap closes, whereas no-crossover cases remain persistent or do not close the gap within training. The same framework also correctly predicts the held-out ResNet-34/CIFAR-10 case, confirmed in all 9 runs. The signal reveals aspects of training dynamics that standard statistics do not represent directly. Activation variance is stronger on average, but the topological measure becomes informative earlier than validation loss and remains significant after controlling for intrinsic dimension. It also distinguishes irreducible label contradictions, including instance-dependent and CIFAR-10N human noise, from learnable confusions such as asymmetric noise. The same measure also falls during grokking in 10 of 12 modular-arithmetic runs and, after per-pipeline calibration from a clean reference plus runs at known noise rates, predicts population-level noise severity without per-sample labels, achieving per-pipeline $R^2 \geq 0.94$ across six tested configurations. These findings position topology as both an explanatory lens on noisy-label memorization and a practical early-stage diagnostic for studying, auditing and improving learning under label noise.

Agent memory systems have demonstrated significant potential in tasks such as long-term dialogue, personalized assistants, and video understanding. However, as inference progresses, continuously accumulated memory imposes substantial storage and retrieval burdens. To address this issue, we propose MemForest, a general memory compression framework adaptable to various agent memory systems. Specifically, MemForest leverages both global semantic similarity and local temporal continuity of memory events to partition the historical memory into a set of event-centric independent units. For each independent unit, the framework constructs a maximum spanning tree structure, referred to as an EventTree, and performs progressive merging by iteratively selecting high-weight edges, thereby effectively compressing redundant memory nodes and reducing storage overhead. In addition, we introduce an anchor-guided propagation retrieval mechanism, which retrieves more relevant memory nodes from the temporal neighborhoods of key memory nodes, thereby enabling more accurate memory retrieval. Extensive experiments demonstrate the effectiveness of MemForest. Under the unimodal Mem0 framework, across three benchmarks (LoCoMo, LongMemEval, and PersonaMem), MemForest preserves 97.1% of the original performance while compressing 50% of historical memory, achieving a 1.89× retrieval speedup. Under the multimodal M3-Agent framework, across two benchmarks (M3-Bench-robot and M3-Bench-web), MemForest retains 99.7% of the original performance under a 50% compression ratio, while achieving a 2.24× retrieval speedup. Our code is available in the supplementary materials, and all data will be released on GitHub.

Multimodal medical AI requires a common visual interface that integrates heterogeneous imaging evidence for patient condition modeling. However, existing medical visual tokenizers are often tied to a specific dimension and imaging modality, forcing multimodal systems to rely on fragmented and poorly reusable representations. We introduce MedVTok, a general-purpose visual tokenizer that maps 2D images and 3D volumes from diverse imaging modalities into a unified token space. Building such a tokenizer poses two key challenges: preserving anatomical structure across dimensions and capturing clinical semantics across modalities. To this end, MedVTok introduces (i) slice differential regularization, which explicitly models inter-slice anatomical coherence missing in previous per-pixel/per-voxel supervision, and (ii) multi-expert representation alignment, which integrates knowledge from multiple medical expert encoders while retaining modality-specific complementary knowledge. Trained on 28M images and 70K volumes across 9 imaging modalities, MedVTok supports classification, retrieval, segmentation, synthesis, visual question answering, and medical report generation through a single tokenizer. Across 24 benchmarks, MedVTok achieves state-of-the-art performance, demonstrating a scalable token interface that unifies medical image dimension, modality, and task usage. Code and weights will be released.

Constantly evolving real-world knowledge necessitates models to be updated continuously. Especially in medicine, as clinical evidence changes over time, outdated knowledge can pose safety risks. Existing evaluations of knowledge integration focus on factual recall, offering limited insight into whether newly integrated knowledge is actually usable. Our benchmark __MedKIT__ (Medical Knowledge Integration and Transfer) provides a granular evaluation of how models integrate and apply knowledge under realistic sequences of clinical updates. Each instance corresponds to a factual update derived from clinical evidence, paired with targeted probes that assess transfer across lexical variation, relational transformations, compositional reasoning, and open-ended operationalization, as well as locality tests for knowledge preservation. Using MedKIT, we conduct a large-scale empirical study of 12 knowledge integration strategies across 5 diverse models, including both general-purpose and medical LLMs. Our results reveal a consistent gap between recall and usable knowledge: while most methods achieve strong gains on the original update task and under lexical variation, relational generalization is limited, and no method yields meaningful improvements on compositional or operational tasks. These findings highlight a fundamental challenge in knowledge integration and position MedKIT as a testbed for developing methods that make newly integrated knowledge more consistently usable across tasks and contexts.

Undesired model behaviors can emerge or re-emerge at any stage of training. Once observed, how effectively can we suppress them? To study this question, we first construct two model organisms---language models trained to exhibit a specific misalignment---with qualitatively distinct structures. In the narrow-trigger organism, undesired behavior is elicited by an identifiable prompt feature. In the broad-trigger organism, misalignment surfaces diffusely across open-ended responses without a separable cue. We then formalize post-hoc behavioral suppression and evaluate methods along three criteria: immediate suppression of the target behavior, durability under further training, and preservation of general capability. We compare standard methods---SFT, GRPO, and gradient-ascent unlearning---along these dimensions while controlling the number of gradient updates. Building on prior work on alignment shallowness, we hypothesize that more durable suppression requires the ability to recover from a wider range of misalignment states. We introduce Prefix-Expanded Adversarial Reinforcement Learning (PEARL), a GRPO variant that augments rollouts with continuations from intermediate states sampled along cached organism trajectories. Across both organism settings and two model families (Qwen3-4B, GPT-OSS-20B), PEARL achieves lower exploit rates while preserving task accuracy, and yields lower reactivation rates under both targeted and benign capability fine-tuning than SFT and GRPO baselines. Together, our framework and method offer a step toward more durable post-hoc removal of undesired behaviors.

Estimating the number of distinct elements in a data stream is well understood when repeated elements are identical. In modern settings, however, observations are high-dimensional and noisy, so repeated instances of the same object are only approximately similar -- for example, different images of the same individual may vary significantly at the pixel level. Classical sketches such as HyperLogLog rely on consistent hash values for identical elements and break down in this regime. Recent work on robust distinct counting in general metric spaces achieves $\tilde\Theta(\sqrt{n})$ memory, which is tight in the worst case. We show that substantially improved memory guarantees are possible under geometric structure common in learned representations. We introduce MaxSketch, a simple max-linear sketch built from random Gaussian projections, and prove that it succeeds in estimating the number of distinct latent objects. Concretely, we show that under this assumption $m = \tilde{O}(\log n / \varepsilon^4)$ random projections (and hence $\tilde{O}(\log n/\varepsilon^4)$ memory) suffice to recover the true distinct count within a $(1+\varepsilon)$ factor. Experiments on image streams confirm that MaxSketch accurately estimates distinct counts and generalizes beyond the training regime. Our results bridge classical streaming algorithms and modern representation learning, showing how geometric structure can fundamentally reduce the complexity of distinct counting.

We present a new paradigm for creating random features to approximate bi-variate functions (in particular, kernels) defined on general manifolds. This new mechanism of $\textit{Manifold Random Features}$ (MRFs) leverages discretization of the manifold and the recently introduced technique of $\textit{Graph Random Features}$ (GRFs) to learn continuous fields on manifolds. Those fields are used to find continuous approximation mechanisms that otherwise, in general scenarios, cannot be derived analytically. MRFs provide positive and bounded features, a key property for accurate, low-variance approximation. We show deep asymptotic connection between GRFs, defined on discrete graph objects, and continuous random features used for regular kernels. As a by-product of our method, we re-discover recently introduced mechanism of Gaussian kernel approximation applied in particular to improve linear-attention Transformers, considering simple random walks on graphs and by-passing original complex mathematical computations. We complement our algorithm with a rigorous theoretical analysis and verify in thorough experimental studies.

Reusable skills have become a core substrate for improving agent capabilities, yet most existing skill packages encode reusable behavior primarily as textual prompts, executable code, or learned routines. For visual agents, however, procedural knowledge is inherently multimodal: reuse depends not only on what operation to perform, but also on recognizing the relevant state, interpreting visual evidence of progress or failure, and deciding what to do next. We formalize this requirement as \emph{multimodal procedural knowledge} and address three practical challenges: (I) \textbf{what} a multimodal skill package should contain; (II) \textbf{where} such packages can be derived from public interaction experience; and (III) \textbf{how} agents can consult multimodal evidence at inference time without excessive image context or over-anchoring to reference screenshots. We introduce \emph{MMSkills}, a framework for representing, generating, and using reusable multimodal procedures for runtime visual decision making. Each MMSkill is a compact, state-conditioned package that couples a textual procedure with runtime state cards and multi-view keyframes. To construct these packages, we develop an agentic trajectory-to-skill Generator that transforms public non-evaluation trajectories into reusable multimodal skills through workflow grouping, procedure induction, visual grounding, and meta-skill-guided auditing. To use them, we introduce a branch-loaded multimodal skill agent: selected state cards and keyframes are inspected in a temporary branch, aligned with the live environment, and distilled into structured guidance for the main agent. Experiments across GUI and game-based visual-agent benchmarks show that MMSkills consistently improve both frontier and smaller multimodal agents, suggesting that external multimodal procedural knowledge complements model-internal priors. The code is accessible in \href{https://anonymous.4open.science/r/MMSkills}{https://anonymous.4open.science/r/MMSkills}.

Transformers dominate video recognition. They split videos into tokens, and processing them has expensive superlinear computational cost. Yet videos are filled with redundancy, so we can question the need for this expense. We introduce LookWhen, a selector–extractor framework that factorizes video recognition into learning when, where, and what to compute. Our shallow selector gets a scaled-down video and quickly scores all tokens across space-time, while our deep extractor gets the top-K selected tokens to approximate full-video representations without actually processing all the tokens. A key challenge is defining effective supervision for selection and extraction. For selection pre-training, we introduce a score on representations that ranks tokens by uniqueness using a simple nearest-neighbor distance. For extraction pre-training, we distill both a video teacher and an image teacher, for which we normalize its frame-wise representations to learn what changes within videos. Through these strategies, our selector-extractor learns general and efficient representations for feature extraction or fine-tuning to a task. Through experiments on Kinetics-400, SSv2, Epic-Kitchens, Diving48, Jester, and Charades, we show that LookWhen achieves a better accuracy-computation trade-off than efficient models and upgraded baselines of similar size. LookWhen Pareto-dominates in accuracy-FLOPs on 9 of 12 cases (6 tasks X 2 settings) and roughly matches on 3. In accuracy-throughput, measuring time in practice, LookWhen is more efficient still at 6.7$\times$ faster than InternVideo2-B at equal accuracy.