NeurIPS 2026 Papers

Heterogeneous AI systems composed of multiple models, architectures, harnesses, or inference-time settings can improve quality and efficiency by routing queries to the \emph{specialist} who can answer most effectively at the lowest cost. Routing requires estimating each specialist's expected return, but this value estimation has a cost. Cheap estimators (e.g., embedding-based predictors) are fast but noisy, while accurate estimators (e.g., fine-tuned models with access to retrieval results or partial reasoning traces) are expensive. We formalize this tradeoff as an instance of Pandora's Box, the classical problem of optimal search with costly inspection. Under a Gaussian signal model, the resulting policies have closed-form value-of-information expressions that determine, for each specialist and input, whether refining the value estimate is worth its cost. We call the centralized policy Pandora's Router. We extend this to a decentralized setting, Pandora's Bidder, where specialists independently decide whether to invest in self-assessment before accepting an offered price to claim a query. Experiments across three domains---a standard multi-LLM benchmark, retrieval-augmented specialists, and LLMs with variable inference-time reasoning---show that Pandora's Router matches the routing quality of exhaustive estimation, while querying the expensive estimator far less often. In the decentralized setting, value-of-information reasoning improves allocative efficiency when competing estimates are accurate; when competing estimates are noisy, however, it can increase the strategic specialist's utility at the expense of

Large language models (LLMs) have recently improved sequential recommendation, yet the task remains a retrieval problem without explicit user queries: the system must infer the next item from user histories where diverse intents and preferences are intertwined. Existing methods typically compress this heterogeneous evidence into a single embedding or query, while LLM-based recommenders can produce overly general queries that are semantically plausible but weakly aligned with retrieval. We propose PRISMIC (Preference Reconstruction via Intent Synthesis and Multi-signal Inference Consolidation), which reformulates sequential recommendation as multi-signal intent consolidation. PRISMIC first generates multiple candidate queries from different historical signals, then trains an LLM-based consolidator with GRPO using an NDCG-based retrieval reward to merge relevant signals into a single query. The resulting consolidated query is not directly used at inference time; instead, it serves as a semantic supervision target for training a user encoder, enabling LLM-free inference. Across four real-world datasets, PRISMIC consistently outperforms strong baselines. Ablations further show that the gains come not from GRPO alone, but from consolidating multiple intent-signals and distilling the consolidated intent into an encoder.

Standard single-cell differential expression (DE) analysis identifies genes that change across conditions, but it usually overlooks two key sources of heterogeneity that single-cell data are uniquely positioned to reveal: which cells within a donor are truly disease-affected, and how disease effects depend on cell state or subject-level context. Assigning every cell the diagnosis of its donor can contaminate DE signal when disease penetrance is partial, while modeling each gene with a single disease effect cannot capture heterogeneous responses across cell populations. We present PRISM (Phenotype-Resolved Inference in Single-cell Mixed models), a negative-binomial mixed-effects framework that augments standard DE analysis with three outputs: a context-DE vector $\theta_g$ that tests whether disease effects vary along a biological axis $z_{ij}$ (such as cell state, sex, or age); a cell-level disease posterior $q_{ij}$ that provides unsupervised disease annotation; and a subject-level disease burden $\rho_i$. Naive likelihood-only inference of $q_{ij}$ is not identified onto the disease axis when nuisance variation (batch, cell-cycle, dominant subtypes) competes with disease; we resolve this with a closed-form 1-D Wasserstein-2 projection of the disease-arm posterior onto a bimodal reference marginal that enforces the \{healthy, affected\} structure implied by the disease label.

In continual learning, most parameter-efficient fine-tuning methods rely on task-specific low-rank adapters to incrementally adapt pre-trained models to new tasks. However, they overlook two critical challenges: one is the accumulation of numerous adapters that leads to linear growth in parameters and memory, and the other is the lack of an explicit mechanism to preserve previously learned knowledge during continual adaptation. In this work, we propose a simple yet scalable framework, termed Progressive Merging of Low-Rank Adapters (PM-LoRA), to address both issues simultaneously. Specifically, PM-LoRA trains a small task-specific spectral orthogonal adapter for each new task and progressively integrates it into a single evolving LoRA adapter that encodes all prior knowledge. This merging process enables continuous knowledge integration while directly mitigating catastrophic forgetting at the parameter level. Benefiting from this design, PM-LoRA achieves true scalability by maintaining constant parameter overhead throughout continual training. Extensive experiments on four benchmark datasets demonstrate that PM-LoRA achieves state-of-the-art performance with superior memory and computation efficiency.

Large pretrained models have achieved remarkable success across diverse domains, yet adapting them to multiple downstream tasks remains challenging: full fine-tuning is costly, while maintaining separate task-specific models is storage-inefficient and fails to produce a unified multi-task model. Low-Rank Adaptation (LoRA) offers an efficient alternative for Model Merging by representing each task adaptation as a low-rank update over frozen pretrained weights. However, existing LoRA merging methods often treat task-specific updates as directly comparable objects and merge them in their original low-rank parameter spaces. This overlooks two key factors: which update directions provide reliable shared support across tasks, and how these directions should be re-parameterized with respect to the pretrained weights before merging. To address this limitation, we propose Pretrained-Guided Shared Basis (PGSB), a two-stage framework for LoRA model merging. PGSB first identifies shared directions among task-specific LoRA updates, and then re-parameterizes these directions using their interaction with the pretrained weights, producing a more suitable basis for merging. Extensive experiments demonstrate that PGSB consistently outperforms state-of-the-art baselines across multiple benchmarks.

AdaBoost sequentially fits so-called weak learners to minimize an exponential loss, which penalizes misclassified data points more severely than other loss functions like cross-entropy. Paradoxically, AdaBoost generalizes well in practice as the number of weak learners grows. In the present work, we introduce Penalized Exponential Loss (PENEX), a new formulation of the multi-class exponential loss that is theoretically grounded and, in contrast to the existing formulation, amenable to optimization via first-order methods, making it a practical objective for training neural networks. We demonstrate that PENEX effectively increases margins of data points, which can be translated into a generalization bound. Empirically, across computer vision and language tasks, PENEX improves neural network generalization in low-data regimes, matching and in some settings outperforming established regularizers at comparable computational cost. Our results highlight the potential of the exponential loss beyond its application in AdaBoost.

We study the problem of multiclass PAC learning with bandit feedback in the realizable setting. In this framework, there is an unknown data distribution over an instance space $\mathcal{X}$ and a label space $\mathcal{Y}$, as in classical multiclass PAC learning, but the learner does not observe the labels of the i.i.d. training examples. Instead, in each round, it receives an unlabeled instance, predicts its label, and receives bandit feedback indicating only whether the prediction is correct. Despite this restriction, the goal remains the same as in classical PAC learning. We provide a general characterization of the optimal sample complexity of this problem, sharp for every concept class up to logarithmic factors. Our characterization is based on a new combinatorial dimension, termed the *bandit $\mathrm{DS}$ dimension*, defined via generalized combinatorial structures we call *pseudo-boxes*. These extend the pseudo-cubes underlying the $\mathrm{DS}$ dimension by allowing a different number of neighbors in each coordinate. In contrast to the $\mathrm{DS}$ dimension, which governs the full-information setting by counting the number of coordinates in the pseudo-cube, the bandit $\mathrm{DS}$ dimension aggregates the number of neighbors across coordinates, leading to a characterization in which the sample complexity scales with the total number of neighbors. We also propose a general learning algorithm achieving the upper bound, based on an algorithmic principle called *ListCascade*, which connects bandit learning to list learning and may be of independent interest.

We study the contextual combinatorial semi-bandit (CCSB) problem with general reward function approximation. At each round, the learner observes a context, selects a combinatorial action consisting of a subset of basic arms, and receives the reward of each selected arm; the goal is to maximize the cumulative reward over time. We propose SquareCB.Comb, a computationally efficient algorithm that, at each round, solves a convex optimization problem to sample a combinatorial action that balances exploration and exploitation. SquareCB.Comb scales to large arm sets and imposes no structural assumptions on the action set beyond a cardinality bound of $m$ on each combinatorial action. We prove that SquareCB.Comb achieves a minimax optimal regret bound of $\tilde{O}{\sqrt{m A T \log |\mathcal{F}|}}$, where $A$ is the number of arms, $m$ is the maximum number of arms in a combinatorial action, $T$ is the time horizon, and $\mathcal{F}$ is the reward function class. In the realizable setting, this bound matches the state-of-the-art regret guarantees achieved by policy search-based algorithms in the more restricted slate recommendation settings, while simultaneously generalizing to arbitrary combinatorial action structures and general reward function approximation.

Online inventory optimization (OIO) is online convex optimization with physical memory: inventory can be raised immediately, but it can be reduced only by demand. A natural principle, used in stochastic inventory learning and recently in OIO with linear capacity constraints, is to maintain a hidden target chosen by an online learner and implement its projection onto the currently feasible order-up-to set. We prove that this simple principle is optimal for OIO on arbitrary bounded convex capacity sets. With online gradient descent as the hidden learner, the method improves the best known regret guarantee for OIO on general convex sets from inverse to inverse-square-root dependence on the common-demand probability, and we prove a matching lower bound. The analysis identifies the right geometric state variable: the Euclidean distance between the hidden target and the implementable set. This distance evolves pathwise as a scalar queue, with target movement as arrivals and common demand as service, reducing the state-dependent feasibility cost to the control of a one-dimensional queue. The same reduction gives the first logarithmic regret guarantee for strongly convex losses and the first dynamic regret guarantee adapting to Euclidean path variation on general convex capacity sets.

Score-based diffusion models have achieved remarkable empirical success in generating high-quality samples from target data distributions. Among them, the Denoising Diffusion Probabilistic Model (DDPM) is one of the most widely used samplers, generating samples via estimated score functions. Despite its empirical success, a tight theoretical understanding of DDPM --- especially its convergence properties --- remains limited. In this paper, we provide a refined convergence analysis for the DDPM sampler and establish near-optimal convergence rates under general distributional assumptions. Specifically, we consider a relaxed smoothness condition parameterized by $L$, which is small for many practical distributions (e.g., Gaussian mixture models). We prove that, to approximate a target distribution on $\mathbb{R}^d$ to accuracy $\varepsilon$ in total variation distance and $\varepsilon^2$ in KL-divergence, the DDPM sampler with accurate score estimates requires at most $$T = \widetilde{O}\left(\frac{\sqrt{d}\min\{\sqrt{d},L\}}{\varepsilon}\right)$$ iterations, where $\widetilde{O}$ hides polylogarithmic factors in $d$ and $1/\varepsilon$. This result substantially improves upon the best-known $\widetilde{O}(d/\varepsilon)$ iteration complexity when $L < \sqrt{d}$. By establishing a matching lower bound, we show that our convergence analysis is tight for a wide array of target distributions. Moreover, it reveals that DDPM and Denoising Diffusion Implicit Models (DDIM) share the same dependence on $d$, raising an interesting question of why DDIM often appears empirically faster.

Recent advances in bioacoustics have been driven by large-scale corpora and standardized benchmarks, yet existing resources are overwhelmingly bird-centric and shallow per species, limiting their use for studying the structure of a single species' communication system. This gap is particularly acute for cetaceans: despite bottlenose dolphins (\textit{Tursiops truncatus}) being a compelling case of complex vocal communication among non-human mammals, existing dolphin datasets are small, fragmented, and largely closed. We introduce OpenWhistle, the largest publicly available dataset of dolphin vocalizations. It comprises approximately 180,000 whistles (114 hours) recorded over five years from a stable pod of five individuals in a semi-natural environment, paired with a curated subset of 8,354 expert-annotated whistles and reproducible evaluation protocols for whistle-type detection and classification. We further release the full processing pipeline for whistle detection, segmentation, and categorization. To demonstrate its utility, we pretrain a Wav2Vec2.0 model adapted to dolphin acoustics on the OpenWhistle corpus and show that it learns effective representations, outperforming general-purpose bioacoustic models such as AVES and BioLingual on both tasks while leaving meaningful headroom for future work. By releasing the dataset, pipeline, and evaluation protocol, we provide the first open dolphin whistle dataset tailored for training self-supervised models, laying the groundwork for advancing dolphin communication research and developing models that capture fine-grained acoustic structure within species.

We study sequential learning for classification metrics that are non-decomposable across instances and are general functions of the confusion matrix, such as the Jaccard index and the $F$-measure. The learner observes instances sequentially, makes irrevocable predictions, and is evaluated by the final empirical performance metric, leading to an online-regret objective. After observing the sequence, the learner also returns a classifier for future population use, leading to a population-regret objective. This single protocol connects two central frameworks for optimizing non-decomposable metrics: Expected Test Utility (ETU) and Population Utility (PU). For smooth concave utilities of the confusion matrix, we show that the ETU benchmark is controlled, in expectation, by the population PU comparator, and that an algorithm with small online regret also yields PU guarantees by an online-to-batch conversion. We then turn to non-concave linear-fractional metrics. We devise a general stochastic Dinkelbach-root method that tracks the relevant metric parameter online and gives online and population regret of order $n^{-1/2}$, up to logarithmic factors, with explicit dependence on conditional-probability estimation error. Our empirical studies evaluate these methods on benchmark datasets.

The Representation Autoencoder (RAE) shows that a frozen high-dimensional semantic ViT representation enables both image reconstruction and DiT-based generation. However, RAE's semantic-only representation lacks appearance details, yielding suboptimal reconstruction. Full fine-tuning the encoder with reconstruction loss induces catastrophic semantic collapse: the encoder overfits to appearance, loses semantic discriminability, and cripples generation. This failure arises because the reconstruction loss imposes an appearance prior that erodes semantic structure. To resolve this conflict, we propose Semantic Structure Regularization (SSR). SSR builds on self-distillation, using a learnable student and a frozen teacher that retains the original semantics. In addition to standard latent alignment, which forces the student's representation to match the teacher's, we introduce two complementary regularizers that prevent appearance overfitting from corrupting semantic structure. First, spatial self-similarity distillation forces the student to reproduce the teacher's patch-wise relational patterns. Invariant to appearance changes, this relational signature anchors high-order semantics and prevents the sacrifice of structure for pixel fidelity. Second, structure-anchored alignment decouples structure from appearance. Using a frequency-domain transform, we extract a high-frequency structural map (edges, boundaries) from each image while discarding low-frequency appearance. The teacher's representation of this structural map then regularizes the student's representation of the original image. Consequently, the student learns appearance details from the reconstruction loss without penalty, where the regularization enforces only structural consistency and enables pixel-level encoding without semantic drift. The result is a unified representation that excels at high-fidelity reconstruction, discriminative perception, and efficient DiT-based generation. Our work provides a principled path toward a single semantic space bridging perception and generation, challenging the classic dichotomy. Code and models will be released.

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

We study issues which may arise when using transformer models to formally reason about problems stated informally using natural language. In particular, we focus on transfer learning between tasks that are semantically equivalent but phrased differently. We assume that examples with a particular target label are only available for a subset of tasks, and aim to transfer zero-shot the ability to predict this label to other tasks. Using controlled experiments on propositional logic problems, we show that transformers are capable of such transfer, but that it can be impeded by the presence of task-specific tokens, due to the model picking up on correlations between these tokens and the label early in the training. As a simple solution, we show that introducing the target label later in the training dramatically improves transfer.

We study offline reinforcement learning under $Q^\star$-approximation and partial coverage, a setting that motivates practical algorithms such as Conservative $Q$-Learning (CQL) [KZTL20] but has received limited theoretical attention. Our work is inspired by the following open question: \emph{Are $Q^\star$-realizability and Bellman completeness sufficient for sample-efficient offline RL under partial coverage?} We answer this question in the negative through an information-theoretic lower bound. To identify additional structure that enables sample-efficient offline RL under partial coverage, we introduce a general decision-estimation framework, inspired by model-free decision-estimation coefficients (DEC) for online RL [FGO$^+$23, LWZ25b]. Our framework decomposes the complexity of offline RL into two parts: the \emph{decision complexity} and the \emph{value estimation error}. This decomposition allows us to study the two sub-problems in a modular way. Our result not only unifies existing results in the $Q^\star$-approximation and partial coverage regime [CJ22, UKLS23], but further improves and generalizes them. On the decision complexity side, our improvement includes: the first $\epsilon^{-2}$ sample complexity bound for soft $Q$-learning under partial coverage that improves [UKLS23]'s $\epsilon^{-4}$ bound, the removal of the need for additional online interaction in the gap-dependent setting of [CJ22], and new learnable settings beyond the above two cases. On the value estimation side, we provide the first characterization of offline learnability for general low-Bellman-rank MDPs [JKA$^+$17, DKL$^+$21, JLM21], a canonical online RL setting that has remained unexplored in offline RL outside special cases. As a side contribution, our techniques give the first analysis of CQL in the function approximation setting.

Omni-modal Large Language Models (OLLMs) greatly expand LLMs' multimodal capabilities but also introduce cross-modal safety risks. However, a systematic understanding of vulnerabilities in omni-modal interactions remains lacking. To bridge this gap, we establish a modality-semantics decoupling principle and construct the AdvBench-Omni dataset, which reveals a significant vulnerability in OLLMs. Mechanistic analysis uncovers a Mid-layer Dissolution phenomenon driven by refusal vector magnitude shrinkage, alongside the existence of a modal-invariant pure refusal direction. Inspired by these insights, we extract a golden refusal vector using Singular Value Decomposition and propose OmniSteer, which utilizes lightweight adapters to modulate intervention intensity adaptively. Extensive experiments show that our method not only increases the Refusal Success Rate against harmful inputs from 69.9% to 91.2%, but also effectively preserves the general capabilities across all modalities.

Multimodal representation learning has been shifting from traditional two-tower architectures to large language model (LLM)-based embedders due to their strong instruction-following capabilities. Despite this progress, existing approaches primarily focus on language and image modalities, which also remain the dominant modalities for human-AI interaction in current embedders. In this paper, we propose the first Omni-Interactive Universal Embedder (OmniUE), which not only learns a unified embedding space across text, video, and audio by leveraging intermediate-layer representations from dedicated learnable tokens, but also supports omni-interactive querying, enabling users to provide inputs in the form of text, visual regions of interest, and audio spans. Within OmniUE, visual and audio segmenters process diverse user interactions and integrate them with an omni-LLM to produce user-conditioned any-to-any embeddings via context aggregation. To evaluate OmniUE’s omni-interactive capabilities, we introduce OmniCHOIR, benchmarking models for omni-interactive compositional audio retrieval based on the given text, video, and audio as well as unimodal or multimodal interaction prompts. OmniUE consistently surpasses state-of-the-art baselines across diverse modalities, with average improvements of 10.5% on textual-interactive video benchmarks (MMEB-v2-video), 1.1% on audio tasks (MAEB), 83.7% on visual-interactive benchmarks (SCaR), and 24.1% on our omni-interactive OmniCHOIR benchmark We believe that jointly advancing omni-modal representation learning and omni-interactive querying paves the way toward universal embedders.

We study offline constrained reinforcement learning with general function approximation in discounted constrained Markov decision processes. Existing methods either require full data coverage for evaluating unsupported intermediate policies, are not oracle efficient, or requires the knowledge of data-generating distribution for policy extraction. We propose PDOCRL, an oracle-efficient primal-dual algorithm based on a decomposed linear-programming formulation. The decomposition makes the policy an explicit optimization variable, avoiding policy extraction through the unknown distribution $\mu_D$. We show that naive restricted saddle-point formulations may have spurious saddle points, so realizability of an optimal solution alone is insufficient. We then show that, under a stronger but explicit realizability assumption, every restricted saddle point is optimal, avoiding the regularization and auxiliary function classes used in prior LP-based analyses. These ingredients yield PDOCRL, an oracle-efficient primal-dual algorithm that computes a near saddle point of the empirical decomposed LP Lagrangian and returns a near-optimal, near-feasible policy with a $\widetilde{\mathcal O}(\epsilon^{-2})$ sample guarantee under partial coverage, without access to $\mu_D$ or a reference policy. Empirically, PDOCRL is competitive with strong baselines on standard offline constrained RL benchmarks.