NeurIPS 2026 Papers

Building performant Vision-Language Models (VLMs) requires carefully curating large-scale training datasets, yet the community lacks systematic benchmarks for evaluating such curation strategies. We introduce DataComp for VLMs (DCVLM), a benchmark for controlled data-centric experiments to improve VLM training. As part of DCVLM, we standardize 160 datasets spanning four data types—image-caption pairs, multimodal interleaved documents, text-only, and instruction-tuning data—into a corpus of 6T multimodal tokens. DCVLM allows participants to test curation strategies (filtering, mixing, formatting, sampling) across 1B–8B models and 6.25B–200B token budgets. Models are then evaluated on a carefully selected suite of up to 52 downstream benchmarks across 9 domains. We conduct extensive experiments on DCVLM and find that data mixing, not filtering, is key to a high-quality training dataset: instruction-heavy mixtures scale better than caption-heavy ones, with gains widening at larger scales. The resulting dataset, DCVLM-BASELINE, enables training an 8B VLM to 63.6% accuracy on our 33-task core suite with 200B training tokens. Compared to FINEVISION, the state-of-the-art open VLM training dataset, this represents an improvement of +5.4pp. DCVLM and all accompanying artifacts will be made publicly available.

Lightweight vision-language models perform competitively on standard benchmarks yet fail systematically in dense-scene reasoning, where multiple objects, attributes, and relations must be jointly grounded and resolved through multi-step inference. Such capability is critical for real-world applications where models must reliably interpret cluttered environments. Yet existing training signals provide no explicit grounding between reasoning steps and the underlying visual entities and relations, leaving lightweight models free to generate fluent but visually unanchored reasoning chains. To address this gap, we first introduce DRBench, a benchmark of 14,573 questions across 2,943 images, organized into five task categories spanning three progressive reasoning layers. Building on DRBench, we propose DRScaffold, a supervised fine-tuning framework that decomposes the supervision target into four causally ordered stages, enforcing grounded reasoning without architectural modification. Experiments on three lightweight VLMs demonstrate substantial gains on DRBench while preserving or improving performance on general-purpose benchmarks. Notably, Qwen2.5-VL-3B trained with DRScaffold surpasses the frozen Qwen2.5-VL-32B on DRBench, demonstrating that structured supervision can substitute for a significant portion of model scale in dense-scene reasoning.

World models are increasingly used to reduce environment interaction when post-training vision-language-action (VLA) policies with reinforcement learning. Yet they are typically trained to predict pixels one step ahead, even though their role here is not merely to render visually faithful videos but to produce learning signals from imagined rollouts that improve the policy. We show that this mismatch is substantial: lower pixel MSE can produce worse policy gradients, and a world model trained for policy utility can yield better downstream policies despite 17\% higher pixel MSE on the next frame. We propose the Downstream-Return Imagination Learning Loop (DRILL), a bilevel framework that updates the world model to maximize the policy's return after an inner GRPO update on imagined rollouts. DRILL has two instantiations. DRILL-VWI is a closed-form surrogate, requiring no additional simulator rollouts, that reweights prediction errors by policy visitation, the group-relative GRPO advantage, and model confidence. DRILL-IMG is a full meta-gradient method that differentiates through the inner update using truncated Hessian-vector products. We show that DRILL-VWI recovers the local one-step component of the DRILL-IMG meta-gradient under score-compatibility and small-residual assumptions. Across five manipulation simulators and two VLA backbones, DRILL-IMG raises mean success rate by 13.1 points over WMPO and 7.9 points over RLVR-World, and complements RLVR pretraining rather than competing with it. More broadly, we believe video world models for VLA reinforcement learning can be productively trained by the policy gradients they produce, not only by pixel-level prediction.

Diagram question answering (DQA) requires models to interpret structured visual representations such as charts, maps, infographics, circuit schematics, and scientific diagrams. Recent vision-language models (VLMs) often achieve high answer accuracy on these tasks, yet correct answers do not guarantee that models ground their reasoning in the diagram regions that support the prediction. Models may instead rely on textual correlations or dataset artifacts without identifying the visual evidence required to verify the answer. This limitation prevents reliable evaluation of diagram reasoning and reduces interpretability. We introduce **DRAGON**, a benchmark for evaluating evidence-grounded visual reasoning in diagrams. Given a diagram, a question, and the correct answer, a model must predict bounding boxes that correspond to the visual elements required to justify the answer. These evidence regions may include answer-bearing components, textual labels, legends, axes, connectors, and other supporting structures involved in the reasoning process. The DRAGON dataset contains 11,664 annotated question instances from six diagram QA datasets: ChartQA, Circuit-VQA, InfographicsVQA, MapIQ, MapWise, and AI2D, with a 2,445-instance test set carrying human-verified evidence annotations and a standardized evaluation framework. Our evaluation on recent VLMs reveals that even Claude Opus 4.6 achieves a Grounding IoU of only 23.8% on AI2D, showing that faithful visual grounding remains an open challenge. DRAGON supports future research on models that ground their predictions in visual evidence.

Differential privacy (DP) has a wide range of applications for protecting data privacy, but designing and verifying DP algorithms requires expert-level reasoning, creating a high barrier for non-expert practitioners. Prior works either rely on specialized verification languages that demand substantial domain expertise or remain semi-automated and require human-in-the-loop guidance. In this work, we investigate whether large language models (LLMs) can automate DP reasoning. We introduce DPrivBench, a benchmark in which each instance asks whether a function or algorithm satisfies a stated DP guarantee under specified assumptions. The benchmark is carefully designed to cover a broad range of DP topics, span diverse difficulty levels, and resist shortcut reasoning through trivial pattern matching. Experiments show that while the strongest models handle textbook mechanisms well, all models struggle with advanced algorithms, revealing substantial gaps in current DP reasoning capabilities. Through further analytic study and failure-mode analysis, we identify several promising directions for improving automated DP reasoning. Our benchmark provides a solid foundation for developing and evaluating such methods, and complements existing benchmarks for mathematical reasoning.

Decentralized training of deep learning models is widely used to enable data privacy and on-device learning over networks. In realistic scenarios, heterogeneity across clients' local data poses an optimization challenge and can severely degrade test accuracy, especially on sparse communication graphs. Building on recent advances in centralized primal averaging, we propose Decentralized Primal Averaging (DPA), a decentralized learning algorithm that combines primal averaging with a local quasi-global momentum estimate. DPA maintains three coupled sequences for optimization, gradient evaluation, and output averaging, while using consecutive mixed iterates to construct a network-aware momentum direction without transmitting an auxiliary tracking variable. We prove a nonconvex stationarity guarantee for DPA's averaged output model and establish a linear-speedup result up to network- and heterogeneity-dependent residual terms. We also introduce DPA-1G, a one-gossip-per-iteration variant that matches the communication budget of DSGD-style baselines. Experiments following the GUT decentralized image-classification benchmark show that DPA outperforms DSGD, gradient tracking, QG-DSGDm, momentum tracking, and QG-GUTm across datasets, architectures, topologies, and heterogeneity levels, with the largest gains under high heterogeneity. DPA-1G matches or exceeds the one-gossip baselines at the same per-iteration communication budget.

We present DMax, a new paradigm for efficient Diffusion Language Models (dLLMs). It mitigates error accumulation in parallel decoding, enabling aggressive decoding parallelism while preserving generation quality. Unlike conventional masked dLLMs that decode through a binary mask-to-token transition, DMax reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings. At the core of our approach is On-Policy Uniform Training, a novel training strategy that efficiently unifies masked and uniform dLLMs, equipping the model to recover clean tokens from both masked inputs and its own erroneous predictions. Building on this foundation, we further propose Soft Parallel Decoding. We represent each intermediate decoding state as an interpolation between the predicted token embedding and the mask embedding, enabling iterative self-revising in embedding space. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of DMax. Compared with the original LLaDA-2.0-mini, our method improves TPF on GSM8K from 2.04 to 5.47 while preserving accuracy. On MBPP, it increases TPF from 2.71 to 5.86 while maintaining comparable performance. On one single H200 GPU using the SGLang framework, our 16B model achieves an average of 955 TPS at a batch size of 1.

Recent advances in reinforcement learning have made verifiable rewards a central post-training paradigm for improving multimodal reasoning. However, methods such as GRPO remain data- and compute-inefficient, often spending rollout and annotation budget on prompts that provide little learning signal. We propose DISCOVER, an online data-discovery framework for budgeted multimodal GRPO that enables vision-language models to dynamically identify prompts that are most useful for the current policy. DISCOVER first organizes the multimodal sample pool into semantic clusters and queries a small set of density-diverse representative samples to obtain initial reward observations. During training, it estimates the utility of unannotated samples by transferring sparse reward feedback from queried samples. Since sample utility changes as the policy evolves, DISCOVER dynamically updates these estimates with recency-weighted rollout observations. It then selects only high-utility prompts for annotation using a policy-conditioned score that accounts for reward variance and difficulty, and continues utility-based replay once the discovery budget is exhausted. Experiments on ViRL39K and VLAA-Thinking with Qwen3-VL-2B show that DISCOVER substantially improves multimodal mathematical reasoning under a strict 10\% unique-sample budget, outperforming budget-matched random selection, hard-example sampling, full-pool GRPO under the same training-step budget, and online selection baselines. Ablations further validate the importance of semantic coverage, online reward transfer, and stable discovery cadence, establishing DISCOVER as a practical framework for data-efficient multimodal RL post-training.

Video multimodal large language models have made strong progress on open-ended video understanding, but they still lack precise local spatiotemporal perception. When two videos share almost the same global semantics and differ only in a short time span or a small region, current models often fail to find the change and provide reliable evidence. We propose **DeltaVid**, a verifiable proxy-task framework that enhances fine-grained spatiotemporal perception with cross-video differences. The key idea is to turn cross-video spot-the-difference into a trainable perception signal, where a model identifies local changes, judges temporal boundaries, and organizes spatial evidence by comparing similar videos. To make this signal scalable to train and reliable to evaluate, we further introduce DeltaVid-10K and DeltaVid-Bench, which convert controllable local differences in real videos into evidence-labeled training and test samples. Experiments show that DeltaVid substantially improves performance on cross-video difference understanding and transfers the learned local evidence ability to general video understanding benchmarks, including MMVU, MLVU, Video-MME, VideoHolmes, VideoMMMU, LVBench, TempCompass, and LongVideoBench. These results show that cross-video differences are not only an effective way to diagnose fine-grained perception failures, but also a scalable proxy supervision that moves Video MLLMs from coarse semantic understanding toward fine-grained spatiotemporal evidence reasoning.

EEG foundation models (EFMs) have improved EEG representation learning by transferring knowledge from large-scale multi-subject pretraining to downstream tasks. However, their pretrained embeddings often retain substantial subject-related variation, placing an additional burden on the downstream classifier, especially in cross-subject settings. To address this issue, we propose DISENTANGLED EFM FINE-TUNING (DEFT), a plug-and-play adapter for EFM fine-tuning. DEFT encourages branch disentanglement of pretrained tokens into task-oriented content and subject-consistent style, then performs downstream prediction from the content branch. It instantiates this factorize-before-classify design with recursive Gaussian refinement, temporal-variance regularization, subject-aware contrastive learning, and reconstruction. Experiments across multiple EEG datasets and four EFM backbones show that DEFT improves matched fine-tuning baselines. Further ablations and analyses support the intended content/style branch organization.

Diffusion large language models (dLLMs) have shown advantages in text generation, particularly due to their inherent ability for parallel decoding. However, constrained by the quality--speed trade-off, existing inference solutions adopt conservative parallel strategies, leaving substantial efficiency potential underexplored. A core challenge is that parallel decoding assumes each position can be filled independently, but tokens are often semantically coupled. Thus, the correct choice at one position constrains valid choices at others. Without modeling these inter-token dependencies, parallel strategies produce deteriorated outputs. Motivated by this insight, we propose DAWN, a training-free, dependency-aware decoding method for fast dLLM inference. DAWN extracts token dependencies and leverages two key observations: (1) masked positions dependent on unmasked certain positions become more reliable; (2) coupled masked positions strongly influence each other's predictions. Given those findings, DAWN leverages a dependency graph to select more reliable unmasking positions at each iteration, achieving high parallelism with negligible loss in generation quality. Extensive experiments across multiple models and datasets demonstrate that DAWN speedups the inference by **1.80–8.06**× over baselines while preserving the generation quality.

Pathological whole slide images (WSIs) from different hospitals often exhibit severe domain shifts due to variations in scanning devices, staining protocols, and tissue preparation procedures. This leads to significant performance degradation when a trained model is applied to data from different sources. Although unsupervised domain adaptation (UDA) has achieved strong performance in transferring a model trained on a source domain to an unlabeled target domain in natural images, it is much less studied in the computational pathology field. This is due to the unique properties of WSIs, including ultra-high resolution and strong inter-slide and intra-slide heterogeneity. As a result, existing UDA methods are often unstable and may even cause negative transfer in WSI classification. To address this issue, we propose Dual-level Adversarial learning with domain-aware REgularization (DARE) for WSI Classification. Specifically, we design an adaptive pseudo-labeling framework with a teacher-student architecture to provide stable supervision for the target domain and perform adversarial alignment on both patch-level and slide-level embeddings. This enables joint optimization of patch-level and slide-level semantic embeddings for dual-level alignment across domains. We also introduce domain-aware attention masking to regularize attention and improve cross-domain generalization. We evaluate our method on four public WSI datasets through extensive experiments. The results show that our method consistently outperforms state-of-the-art approaches across various domain shifts.

Self-supervised Vision Transformers in the DINO family encode rich semantic structure within their self-attention maps, yet existing parameter-efficient fine-tuning (PEFT) methods apply static, content-agnostic transformations that ignore this internal signal. We present DAGA (Dynamic Attention-Guided Adaptation), a PEFT module that repurposes a frozen backbone's emergent attention maps as instance-specific guidance: per-image attention is converted into channel and spatial gates that condition a lightweight bottleneck adapter. With only 1.4\% additional parameters, DAGA reaches 85.9\% top-1 on ImageNet-1K with DINOv3-ViT-B and surpasses recent PEFT baselines, with strong gains on dense prediction (+12.1\% mIoU on ADE20K) and fine-grained classification (+1.6 / +3.5 on CUB-200 / FGVC-Aircraft). DAGA's effectiveness is tied to the semantic quality of the backbone's attention: gains are large on DINOv2/v3 and iBOT but small on MAE, CLIP, and DeiT, positioning it as the first PEFT module to operationalize emergent SSL attention as in-loop guidance and complementing self-distillation methods that refine attention itself. Anonymized code: https://anonymous.4open.science/r/ID1680.

Existing deep learning architectures for survival prediction from electronic health records (EHRs) largely assume unimodality, failing to capture the complex relational structure across diverse clinical modalities. Graph learning can allow us to robustly adapt to the underlying clinical structure of the data to have more predictive and interpretable representations for time-to-event analysis. However, unconstrained graph learning produces dense, bidirectional adjacencies that overfit. We introduce M-CADENCE, a multimodal survival model that integrates structured acyclic priors with attention-based propagation. Each modality has its own learnable adjacency, regularised toward acyclicity through a log-determinant characterisation and toward sparsity through soft thresholding. A cross-modal meta-adjacency keeps the cross-modal parameter count quadratic in the number of modalities rather than in their feature dimensions. Propagation is then performed by multi-head attention biased by the learnt structure. The propagation step degrades gracefully under sparse or misspecified structure, addressing a known weakness of hard-routing graph variants. We evaluate M-CADENCE on five EHR benchmarks covering intensive-care mortality, circulatory failure, and emergency-department competing risks, against multiple survival, graph, and structure learning baselines. M-CADENCE achieves the best concordance on all datasets, the best early-warning AUC on four of five, and competitive precision-recall against the strongest deep survival baselines. Removing the DAG bias lowers concordance, as does substitution with a density-matched random adjacency, which confirms that the specific learnt structure is required. Discrimination is stable across four orders of magnitude in regularisation strength, and evaluation by a panel of ten language models confirms M-CADENCE learns clinical edges that are more physiologically plausible than baseline structure-learning methods.

The rapid evolution of Embodied AI has enabled Vision-Language-Action (VLA) models to excel in multimodal perception and task execution. However, applying Reinforcement Learning (RL) to these massive models in large-scale distributed environments faces severe systemic bottlenecks, primarily due to the resource conflict between high-fidelity physical simulation and the intensive VRAM/bandwidth demands of deep learning. This conflict often leaves overall throughput constrained by execution-phase inefficiencies. To address these challenges, we propose D-VLA, a high-concurrency, low-latency distributed RL framework for large-scale embodied foundation models. D-VLA introduces "Plane Decoupling," physically isolating high-frequency training data from low-frequency weight control to eliminate interference between simulation and optimization. We further design a four-thread asynchronous "Swimlane" pipeline, enabling full parallel overlap of sampling, inference, gradient computation, and parameter distribution. Additionally, a dual-pool VRAM management model and topology-aware replication resolve memory fragmentation and optimize communication efficiency. Experiments on benchmarks like LIBERO show that D-VLA significantly outperforms mainstream RL frameworks in throughput and sampling efficiency for billion-parameter VLA models. In trillion-parameter scalability tests, our framework maintains exceptional stability and linear speedup, providing a robust system for high-performance general-purpose VLA Models.

The landscape of high-performance image generation models is currently shifting from the inefficient multi-step ones to the efficient few-step counterparts (e.g, Z-Image-Turbo and FLUX.2-klein). However, these models present significant challenges for direct continuous supervised fine-tuning. For example, applying the commonly used fine-tuning technique would compromise their inherent few-step inference capability. To address this, we propose D-OPSD, a novel training paradigm for step-distilled diffusion models that enables on-policy learning during supervised fine-tuning. We first find that the modern diffusion models, where the LLM/VLM serves as the encoder, can inherit its encoder's in-context capabilities. This enables us to formulate the training as an on-policy self-distillation process. Specifically, during training, we make the model act as both the teacher and the student with different contexts, where the student is conditioned only on the text feature, while the teacher is conditioned on the multimodal feature of both the text prompt and the target image. Training minimizes the two predicted distributions over the student's own roll-outs. By optimizing on the model's own trajectory and under its own supervision, D-OPSD enables the model to learn new concepts, styles, etc., without sacrificing the original few-step capacity.

Large language models (LLMs) deliver strong performance, but their high compute and memory costs make deployment difficult in resource-constrained scenarios. Weight-only post-training quantization (PTQ) is appealing, as it reduces memory usage and enables practical speedup without low-bit operators or specialized hardware. However, accuracy often degrades significantly in weight-only PTQ at sub-4-bit precision, and our analysis identifies two main causes: (1) down-projection matrices are a well-known quantization bottleneck, but maintaining their fidelity often requires extra bit-width; (2) weight quantization induces activation deviations, but effective correction strategies remain underexplored. To address these issues, we propose D$^2$Quant, a novel weight-only PTQ framework that improves quantization from both the weight and activation perspectives. On the weight side, we design a Dual-Scale Quantizer (DSQ) tailored to down-projection matrices, with an absorbable scaling factor that significantly improves accuracy without increasing the bit budget. On the activation side, we propose Deviation-Aware Correction (DAC), which incorporates a mean-shift correction within LayerNorm to mitigate quantization-induced activation distribution shifts. Extensive experiments show that D$^2$Quant achieves superior weight-only PTQ performance at sub-4-bit precision. We will release the code and quantized models of D$^2$Quant.

We introduce CurveBench, a benchmark for hierarchical topological reasoning from visual input. CurveBench consists of $\textbf{756 images}$ of pairwise non-intersecting Jordan curves across easy, polygonal, topographic-inspired, maze-like, and dense counting configurations. Each image is annotated with a rooted tree encoding the containment relations between planar regions. We formulate the task as structured prediction: given an image, a model must recover the full rooted containment tree induced by the curves. Despite the visual simplicity of the task, the strongest evaluated model, $\texttt{Gemini-3.1-Pro-Preview}$, achieves only $\textbf{71.1\\%}$ tree-generation accuracy on CurveBench-Easy and $\textbf{19.1\\%}$ on CurveBench-Hard. We further demonstrate benchmark utility through RLVR-style fine-tuning of open-weight vision-language models. Our trained $\texttt{Qwen3-VL-8B}$ model improves over $\texttt{Qwen3-VL-8B-Thinking}$ from $\textbf{2.8\\%}$ to $\textbf{33.3\\%}$ tree-generation accuracy on CurveBench-Easy, exceeding $\texttt{GPT-5.4}$ and $\texttt{Claude-Opus-4.5}$ under our evaluation protocol. The remaining gap, especially on CurveBench-Hard, shows that exact topology-aware visual reasoning remains far from solved.

Submodular functions---functions exhibiting diminishing returns---are central to machine learning. When the objective is monotone and non-negative, the greedy algorithm achieves a tight $63\%$ approximation. But many practical objectives incorporate costs that make them negative on some inputs, and all existing multiplicative guarantees require non-negativity. Prior work handles negativity through additive bounds for the special class of decomposable functions and non-monotonicity through partial-monotonicity parameters, but these address each difficulty in isolation and neither extends the classical structural theory. We extend curvature---a parameter measuring how far a function deviates from linearity---to all submodular functions, handling both non-monotonicity and negativity through a single classical concept. A greedy algorithm with pruning achieves a curvature-controlled multiplicative ratio for any submodular function, including those taking negative values---the first such guarantee beyond monotonicity and non-negativity. In the non-monotone regime $1 \le c_g < 2.2$, the bound strictly beats the best known uniform ratio of $0.401$ (for non-negative $f$), and it recovers the classical $(1-e^{-c_g})/c_g$ guarantee for monotone functions. A multilinear-extension variant extends the framework to general combinatorial constraints via multilinear relaxation. Experiments on cost-penalized experimental design, coverage, feature selection, and a curvature sweep on Multi-News passage selection support the theory.