NeurIPS 2026 Papers

Recent work has shown that continued fraction inspired neural architectures (CoFrGeNets) can be highly parameter efficient yet performant for language modeling. In this work, we revisit continued fraction-based channel mixing with the goal of improving its performance and simplifying integration into standard pre-training protocols for language models. We achieve this by two simple yet critical contributions: i) we propose a novel non-expanding gated continued fraction inspired architectural component to replace feedforward layers in established language model architectures that may use vanilla multilayer perceptrons or gated linear units or (vanilla/gated) experts in mixture-of-experts (MoE) type architectures. ii) We propose a way to re-implement divisions -- the non-linearity in these architectures -- in terms of logarithms and exponentials. By experimenting on different language modeling architectures such as GPT-2, Llama-3, Mamba-2 and (nano) Llama-MoE we find that our proposed architecture performs similarly or significantly better than the continued fraction architecture proposed in prior work along with having higher throughput. In addition, our re-implementation of divisions makes training of deeper CoFrGeNets stable without the need for incremental training which was proposed in prior work as a means to stabilize training, but added notable complexity to existing large language model training implementations.

Digital entities such as AI agents and humanoid robots increasingly operate alongside real humans, yet their identity infrastructure is based on credentials rather than embodied biometric identity. We introduce Biometric Identity Provisioning (BIP), a new problem and solution framework that addresses: given an enrollment gallery of real human identities, provision virtual identities that are non-colliding with every enrolled identity, maintain sufficient inter-class separability, and are realizable as high-fidelity face images. The key geometric insight is that real face identities occupy a low-dimensional subspace of the embedding hypersphere, leaving no residual subspace for virtual identities. Hence, virtual identities must instead be allocated as unclaimed gaps within the real face manifold itself. BIP is therefore a constrained packing problem: available gaps vastly exceed any foreseeable enrollment scale, and provisioned identities remain non-colliding even as new real identities are subsequently enrolled. Grounded in this geometry, our repulsion-based allocation is not bounded by any fixed provisioning count; we demonstrate 10M non-colliding virtual identity embeddings against a gallery of 360K real identities. Realizing these embeddings as face images requires a generator that operates outside the training distribution of real face images; we introduce GapGen, a gap-aware generator trained with a curriculum that progressively extends synthesis into non-colliding regions, validated at 1M photorealistic virtual face images. We further construct v-LFW, a virtual counterpart to LFW face dataset, with protocols for virtual face verification, cross-reality matching, real-vs-virtual detection, and unified recognition and detection.

In this work we study the Best Policy Identification (BPI) problem in online, tabular Reinforcement Learning. This is an active sequential hypothesis testing problem in which the learner's objective is to identify an optimal policy in a Markov Decision Process (MDP) with high confidence, while minimizing the expected sample complexity to do so. We consider an online setting with deterministic rewards, where the agent must strategically navigate through the MDP in order to effectively explore. Previous works in the literature have provided asymptotically optimal methods for BPI, such as the Navigate and Stop (NaS) algorithm and its variants, however existing analysis remains asymptotic. In this work, we fill that gap by providing the first non-asymptotic sample complexity guarantees for NaS, showing that its sample complexity depends not only on the characteristic time, but also on the connectivity of the underlying MDP, the curvature of the optimal characteristic time, and other instance-dependent quantities. We identify these additional attributes and make explicit their contributions to the overall sample complexity.

Latent signals are often obscured by measurement noise, yet encode the underlying laws and dynamics of complex systems; learning both the signals and their distributions remains a central challenge in scientific inference. The noise is often non-negligible, and the likelihoods for expressive generative models are often intractable. We utilize a convolutional maximum mean discrepancy (convMMD) loss and propose a likelihood-free framework for nonparametric density deconvolution and empirical Bayes denoising under additive measurement error. Our method learns a latent generative model by matching the observed data distribution to the noise-convolved model distribution. This yields a differentiable, simulation-based objective for multivariate homoscedastic or heteroscedastic noise, compatible with expressive sieve classes such as Gaussian mixtures and normalizing flows. The learned density then serves as an empirical prior for posterior denoising of individual latent values. Theoretically, we extend convMMD from parametric to nonparametric estimation, proving finite-sample bounds for empirical sieve minimizers and $L_2$ convergence rates under Sobolev smoothness. These rates recover the classical inverse-problem dependence: polynomial for ordinary-smooth and logarithmic for super-smooth noises. Our method provides a practical, theoretically grounded approach to deconvolution and denoising under generative latent distribution models.

The expressivity of graph neural networks is predominantly characterized in a noiseless setting via binary graph isomorphism tests. This perspective certifies whether two graphs can be distinguished, but does not quantify how robust such distinctions are under structural perturbations. In practice, however, real-world graphs are often corrupted by noise, where even small structural changes might lead to disproportionate distortions in learned representations. In this study, we generalize graph isomorphism to a noisy alignment criterion, requiring that graphs that remain identifiable under noise are embedded closer to their source than to any alternative. This yields a margin-based formulation of expressivity that bridges the gap between combinatorial isomorphism-based theory and the continuous learned representations. Notably, our framework is architecture-agnostic and strictly generalizes classical isomorphism-based expressivity. We further derive certified robustness guarantees for node-level tasks and an explicit robustness radius for graph-level representations. Experiments confirm that our criteria predict and improve robustness under structural perturbations.

Reinforcement learning with verifiable rewards (RLVR) has driven recent capability advances of large language models across domains. Recent studies suggest that improved RLVR algorithms allow models to learn effectively from incorrect annotations, achieving performance comparable to learning from clean data. In this work, we show that these findings are invalid because the claimed 100% noisy training data is "contaminated" with clean data. After rectifying the dataset with a rigorous re-verification pipeline, we demonstrate that noisy data is destructive to RLVR. We show that existing RLVR algorithm improvements fail to mitigate the impact of noisy data, achieving similar performance to that of the basic GRPO. Furthermore, we find that the model trained on truly incorrect annotations performs 8-10% worse than the model trained on clean data across mathematical reasoning benchmarks. Finally, we show that these findings hold for real-world noise in Text2SQL tasks, where training on real-world, human annotation errors cause 5-12% lower accuracy than clean data. Our results show that current RLVR methods cannot yet compensate for poor data quality. High-quality data remains essential.

Pre-trained text-to-image (T2I) diffusion models have shown strong potential for real-world image super-resolution (Real-SR), owing to their noise-started generation process that enables realistic texture synthesis and captures the one-to-many nature of super-resolution. However, diffusion-based Real-SR methods still face a fundamental efficiency-quality trade-off. Multi-step methods generate high-quality results by iteratively denoising random Gaussian noise under LR conditioning, but suffer from slow sampling. Recent one-step methods greatly improve efficiency, yet they typically replace noise-started generation with direct LR-to-HR restoration, which weakens stochasticity and limits realistic detail synthesis. To address this issue, we propose SMFSR, a noise-started one-step Real-SR framework via LR-conditioned SplitMeanFlow and GAN refinement. SMFSR preserves the random-noise starting point of diffusion models and learns a direct noise-to-HR mapping conditioned on the LR image. To this end, Interval Splitting Consistency distills the multi-step generative trajectory into a single average-velocity prediction, enabling efficient one-step generation. To compensate for the reduced opportunity for progressive refinement, we further introduce a GAN refinement stage, where a DINOv3-based discriminator enhances realistic texture synthesis and variational score distillation aligns the generated outputs with the natural image distribution. Extensive experiments demonstrate that SMFSR achieves superior perceptual quality while retaining the high efficiency of one-step diffusion models.

It is now common to use 2D diffusion priors for problems where the underlying object is 3D (e.g., medical imaging). We would pick a noise schedule from the EDM literature, anneal a tilted prior, and then use the formulation as a 3D generative model. It works empirically. But we know very little about what it converges to and how fast. We study this question for a natural target: the multi-marginal Schrödinger bridge, the relative-entropy projection of a 3D prior onto the distribution whose rendered slices match the 2D marginals. But computing it is intractable. The annealed 2D-tilted surrogate is what one actually uses in applications. The main question is how close the latter object is to the former. Our main idea is that the discrepancy between these two laws — a comparison of two high-dimensional, fully coupled distributions — can be reduced to the size of a *single* scalar correction field. This field has a nice two-piece form: residual coupling across slices and tempering mismatch from the annealed discrete schedule. Using a Gaussian-smoothed filtration, Wasserstein screening, and conditional concentration, we obtain an explicit discrepancy bound in which the relevant terms separate into latent and posterior-residual contributions. When the smoothing scale is identified with the EDM noise level $\sigma(\tau)$, our bound yields $\mathrm{KL}(\Pi_\tau\|P^\star)=O\left(N\sigma(\tau)^2\right)$ and $\|\Pi_\tau-P^\star\|_{\mathrm{TV}}=O\left(\sqrt{N}\sigma(\tau)\right),$ if the latent scale remains bounded and the surrogate log-discrepancy is noise-aligned. So, the 2D-tilted construction converges to the Schrödinger-bridge lift at a rate governed by user-checkable quantities like the surrogate Lipschitz constants and the posterior concentration constant.

Robot fine-tuning can improve manipulation success while degrading the semantic structure inherited from large vision-language pretraining. This trade-off limits the reuse of vision-language-action (VLA) policies in settings that require compositional instructions, camera changes, and coordination with higher-level agents. We present \textbf{NoTVLA}, a semantics-preserving robot adaptation framework that replaces dense low-level action supervision with a sparse narrative action interface. NoTVLA converts demonstrations into sparse, semantically meaningful waypoints, grounds each decision with a task-relevant visual anchor and depth value, and reconstructs executable motion through a deterministic detokenizer. The resulting interface keeps the autoregressive vision-language backbone close to its pretrained prediction format while delegating high-frequency control to a transparent motion-rendering stage. We evaluate this design through matched and backbone-family comparisons, semantic retention probes, semantic out-of-distribution manipulation tasks, camera and depth perturbations, and deployment-oriented efficiency analysis. The results suggest that robot adaptation should be judged not only by task success, but also by how much task-relevant semantic competence is preserved during fine-tuning.

Rehearsal-free class-incremental learning (CIL) requires a model to learn new classes sequentially without storing previous data, making catastrophic forgetting a central challenge. Adapter-based methods have become a prevalent solution by storing and accumulating task-specific update matrices, yet the effect of the cumulative update magnitude on CIL performance remains underexplored. In this work, we propose Normalized Tensor Adaptation (NoTA), which addresses these two aspects jointly. MPO adapters provide a compact structured space for storing task-specific updates, while global normalization controls the accumulated update that determines the final model. Our empirical analysis reveals a key observation: global normalization consistently improves both matrix-based and MPO-based adapters by stabilizing the cumulative update magnitude across task sequences. We further provide a theoretical analysis showing that global normalization yields a forgetting bound independent of the task sequence length, in contrast to task-wise normalization whose bound can grow with later tasks. Building on the tensor network structure of MPO, we introduce a variant with hierarchical coefficients (H-NoTA), which uses task-level coefficients and inter-core modulation matrices to refine the contributions of historical adapters. Extensive experiments demonstrate that our methods consistently improve rehearsal-free CIL performance while maintaining strong parameter efficiency.

Delayed hits occur when requests arrive while a previous miss for the same item is still being fetched. They break the equivalence between hit ratio and latency because caching decisions depend on frequency, burst structure, and fetch latency. We develop a regret framework for delayed-hit caching, with static regret against long-term placement and switching regret against piecewise-changing workloads. We show that classical deterministic policies and randomized Marker fail to achieve no-regret guarantees, both suffering linear static regret. For stochastic arrivals, we introduce delayed-hit shadow values and propose UCB-DH, which achieves $O(\log T)$ static regret on fixed instances with positive gaps, where $T$ is the horizon. A matching gap-dependent lower bound shows that the delayed-hit latency-amplification factor is statistically unavoidable, and restarted UCB-DH gives sublinear regret for piecewise-changing workloads. For oblivious adversarial arrivals, we formulate an action-independent shadow loss and propose LA-FTPL with latency-scaled perturbations and completion-time admission. LA-FTPL achieves $O(\sqrt{KT\log N\,(Z^2+C_BZ)})$ shadow static regret, where $K$ is the cache capacity, $N$ the item universe size, $Z$ the fetch latency, and $C_B$ the maximum one-window delayed-hit penalty. With periodic restarts, it achieves $O(ZT^{2/3}B^{1/3}(K\log N)^{1/3})$ switching regret against comparators with at most $B$ cache changes. Experiments on synthetic and real-world traces show that our algorithms achieve sublinear regret and improve latency over the state-of-the-art baselines.

Geospatial foundation models (GFMs) have been proposed as generalizable backbones for disaster response, land-cover mapping, food-security monitoring, and other high-stakes Earth-observation tasks. Yet the published work about these models does not give reviewers or users enough information to tell which model fits a given task. We argue that nobody knows what the current state of the art is in geospatial foundation models. The methods may be useful, but the GFM literature does not standardize evaluations, training and testing protocols, released weights, or pretraining controls well enough for anyone to compare or rank them. In a 152-paper audit, we find 46 cross-paper disagreements of at least 10 points for the same model, benchmark, and protocol; 94/126 papers with extractable pretraining data use a configuration no other paper uses; and 39\% of GFM papers release no model weights. This lack of community standards can be solved. We propose six concrete expectations: named-license weight release, shared core evaluations, copied-versus-rerun baseline annotations, variance reporting, one shared evaluation harness, and data-vs-architecture-vs-algorithm controls. These gaps are a coordination failure, not a fault of any individual lab; the authors of this paper, like many others in the GFM community, have contributed to them. Rather than just critiquing the community, we aim to provide concrete steps toward a shared understanding of how to innovate GFMs.

Iterative fine-tuning on synthetic data causes \emph{model collapse}: output diversity narrows as rare patterns are progressively lost. Existing mitigations either require model log-probabilities, an external oracle, or continued access to real human data. Here we develop a new approach grounded in mathematical information theory: the non-parametric Kontoyiannis entropy rate estimator $H_K$, computed entirely from raw text via match-length statistics, with no model of any kind. We show that this is in fact a superior training-data filter on text-diversity metrics in a fully-synthetic, single-lineage fine-tuning setting. In a six-generation QLoRA collapse experiment on Llama-3.1-8B, logprob-based filtering (the established gold standard) provides no significant text-diversity benefit on any metric ($p > 0.23$), whereas $H_K$-filtering yields +42% unique trigrams, +30% vocabulary, and -19% repetition (all $p < 0.001$). We validate $H_K$ as a cross-domain entropy proxy ($\beta = 0.924$, $R^2 = 0.746$) and collapse detector ($\rho = +0.454$, $p < 0.0001$) across 4 domains, 3 temperatures, 2 generator--scorer model pairs, and 1,680 generated documents. Our results demonstrate that information theoretic approaches to collapse mitigation are efficient, and suggest new approaches for maintaining multi-agent diversity.

Multimodal heterogeneous learning aims to train reliable multimodal models when data sources differ in modality availability, semantic coverage, domain distribution, and missingness patterns. Although cross-modal alignment is widely used to mitigate such heterogeneity, we argue that alignment is not free: enforcing relations that are weakly supported by the observed data can amplify noise, over-transfer missing-modality information, and induce negative transfer. This paper studies when a modality--modality--semantic relation is sufficiently observable to be safely aligned and shared across heterogeneous sources. We propose **ObsAlign**, an observability-aware framework that summarizes visible modality--semantic evidence through compact source-level sketches and estimates relation-level support from bridge evidence, side support, cross-modal consistency, and prototype uncertainty. The resulting observability score gates co-observed alignment, calibrates weak missing-modality transfer, and guides evidence-aware source consolidation so that reliable relations are shared while unsupported variation remains local. Across image--text, audio--visual--text, action-recognition, sensor, and healthcare benchmarks under source-partitioned multimodal heterogeneity, ObsAlign achieves the best overall performance in both encoder-based and VLM-compatible regimes, while sharply reducing negative transfer on low-observability relations. These results suggest that relation-level observability provides a practical principle for deciding when cross-modal alignment should be strengthened, weakened, or avoided.

We investigate the **strategic surplus** obtainable against a **Follow-the-Regularized-Leader (FTRL)** learner with constant step size $\eta$ in $n \times m$ two-player zero-sum games played over $T$ rounds against a clairvoyant optimizer. In contrast with prior analysis, we show that the extraction of such regret-scale surplus is an inherent feature of the FTRL family, rather than an artifact of specific instantiations. First, for a fixed max-min optimizer, we establish a sweeping law of order $\Omega(N/\eta)$, proving that utility surplus scales with the number of the learner's suboptimal actions $N$ and vanishes in their absence. Second, for an alternating optimizer, a surplus of $\Omega(\eta T/\text{poly}(n,m))$ can be guaranteed regardless of the equilibrium structure, with high probability, in random games. Our analysis uncovers a sharp geometric dichotomy: **non-steep** regularizers allow the optimizer to realize the maximal transient surplus via finite-time elimination of suboptimal actions, whereas **steep** regularizers introduce a vanishing tail correction that can delay surplus saturation. Finally, we discuss whether this leverage persists under bilateral payoff uncertainty and propose a susceptibility measure quantifying which regularizers are most vulnerable to learner-aware strategic steering.

Ocean dynamics are inherently chaotic, yet existing machine learning ocean models produce only deterministic forecasts. We introduce Njord, a probabilistic data-driven model for ocean forecasting, applicable to both global and regional domains. Njord combines a deep latent variable framework with a graph neural network architecture, enabling sampling each forecast step in a single forward pass. We apply Njord globally at 0.25° resolution and regionally to the Baltic Sea at 2 km resolution. To scale to these large ocean grids we introduce K-means cluster meshes that adapt to irregular sea surface geometry. Experiments demonstrate strong performance on both domains compared to deterministic machine learning baselines, while also providing uncertainty estimates from the sampled ensemble forecasts. On the global OceanBench benchmark, Njord achieves the lowest errors on average across upper-ocean variables when evaluated against real-world observations, with the largest improvements in surface temperature prediction.

Recent work on agentic training efficiency has mainly focused on improving learning algorithms, rollout strategies, or model-hardware co-design. However, these approaches overlook an important source of training latency: the sandbox that runs the environments. Recent studies show that creating, executing, and resetting containers accounts for an important portion of the rollout process in RL training, substantially slowing down overall training. This overhead largely stems from the centralized management mechanisms (e.g., DAEMON) of existing sandboxes, along with additional features (e.g., multi-tenancy) that are unnecessary for reinforcement learning (RL). Moreover, existing sandboxes are designed for general-purpose isolated execution, making them a poor fit for the requirements of RL rollout and training. We propose NitroBox, a novel sandbox system purpose-built for agentic RL that significantly accelerates the rollout and training process. Technically speaking, we first comprehensively redesign NitroBox's lifecycle to eliminate the dependency on DAEMON, so that every sandbox instance is fully independent during creation, startup, and reset (e.g., via rootless full-namespace isolation). Furthermore, we introduce a copy-on-write mechanism to enable fast startup across multiple instances. Second, we knock off features that are time-consuming but unnecessary for RL — such as multi-tenancy — while preserving the core functionality and security guarantees needed for RL training. Finally, we introduce some RL-nature features to further improve efficiency, such as optional CRIU-based process checkpoint saving and restoring for mid-trajectory branching. Our experiments show that NitroBox is faster than Docker across the entire lifecycle: 10.6× faster in sandbox creation, 24.6× in reset, and 19.1× in deletion. When training Qwen2.5-Coder-32B-Instruct with GRPO at c=32 concurrency, end-to-end wall-clock training time drops from 40.6 hours to 28.4 hours when replacing Docker with NitroBox, while preserving the same task pass rate.

The foundational capabilities of large language models are acquired during pretraining on internet-scale, highly heterogeneous data mixtures. In this work, we investigate a geometric question regarding the converged state of this pretraining process: Does the model converge to a common minimizer across all data sources (e.g., \cref{fig:cwa_illustration:close}), or merely a minimizer of the averaged loss (e.g., \cref{fig:cwa_illustration:distant})? We hypothesize that the geometric ``closeness'' of task-specific minima is intrinsically linked to downstream generalization. However, we reveal that standard optimizers (e.g., AdamW) often converge to points where task-specific minima are distant from each other. To address this, we propose the \textit{Nexus} optimizer, which encourages the closeness of these minima by maximizing gradient similarity during optimization. Extensive experiments across models ranging from 130M to 3B parameters demonstrate that Nexus significantly boosts downstream performance, despite achieving nearly \textit{the same pretraining loss} (see \cref{fig:demo:benchmark}). Notably, on the 3B model, Nexus reduces the out-of-distribution loss by 0.012 and yields up to a 15.0\% accuracy improvement on complex reasoning tasks (e.g., GSM8k). Our findings challenge the reliance on pretraining loss as the sole proxy for model evaluation and highlight the critical role of optimizer implicit bias in unlocking downstream generalization.

Transformers replace recurrence with a memory that grows with sequence length and self-attention that enables ad-hoc look ups over past tokens. Consequently, they lack an incentive to compress history into compact latent states with consistent transition rules. This often leads to learning solutions that generalize poorly. We introduce **Next-Latent Prediction** (NextLat), which extends standard next-token training with *self-supervised* predictions in the latent space. Specifically, NextLat trains a transformer to learn latent representations that are predictive of its next latent state given the next output token. Theoretically, we show that these latents provably converge to *belief states*, compressed information of the history necessary to predict the future. This simple auxiliary objective injects a recurrent inductive bias into transformers, while leaving their architecture, parallel training, and inference unchanged. NextLat effectively encourages the transformer to form compact internal world models with its own belief states and transition dynamics—a crucial property absent in standard next-token prediction transformers. Empirically, across benchmarks in world modeling, reasoning, planning, and language modeling, NextLat demonstrates significant gains over standard next-token training in downstream accuracy, representation compression, and lookahead planning. Furthermore, NextLat enables *variable-length self-speculative decoding*, accelerating inference by up to $3.3\times$ in the language domain. NextLat stands as a simple and efficient paradigm for shaping transformer representations toward stronger generalization.