NeurIPS 2026 Papers

Video Diffusion Transformers (DiTs) generate high-quality videos but demand substantial compute due to wide blocks, deep architectures, and iterative sampling. Recent methods reduce cost by compressing width, depth, or sampling steps, but typically commit to a fixed architecture that cannot adapt to individual inputs or denoising stages. We propose PARE (Pruning and Adaptive Routing for Efficient video generation), which jointly compresses width and depth with structure-aware pruning and input-adaptive routing. For width, we observe that attention heads specialize into spatial and temporal roles, and design importance scoring that accounts for this distinction to prevent motion-critical temporal heads from being pruned prematurely. For depth, we train a lightweight router conditioned on denoising timestep and visual content to dynamically select which blocks to execute at each step, enabling per-input compute adaptation rather than static block removal. A progressive pipeline first recovers width-pruned quality via distillation, then jointly optimizes the student and router to decouple the two learning objectives. Experiments on Wan2.1-14B for both image-to-video and text-to-video generation show that PARE substantially reduces per-step computation while preserving quality across VBench dimensions, and composes with step distillation for further acceleration.

Vision-Language-Action (VLA) models show promising ability in language-guided robotic tasks. However, making VLA policies reliable remains challenging, because a manipulation task is completed through closed-loop interaction, where each action affects subsequent execution. To analyze this problem, we revisit VLA policy during execution and argue that a VLA policy acts both as a planner, which makes task-oriented decisions that change the direction of execution, and as an executor, which realizes these decisions through dense continuous actions. This view suggests that improving VLA reliability requires particular attention to planning actions. Existing optimization methods can imitate actions or improve complete trajectories, but they usually do not explicitly identify planning actions or measure their importance for task success. To address this issue, we propose Planning-Aware Policy Optimization for VLA models (PAPO-VLA). PAPO-VLA first identifies planning actions by jointly considering action variation and trajectory outcome, then estimates their importance through causal sufficiency and causal necessity, and finally incorporates this importance into GRPO advantage estimation. In this way, more important planning actions receive stronger optimization emphasis, while the whole trajectory is still optimized by trajectory-level feedback. Experiments on multiple benchmarks demonstrate the effectiveness of PAPO-VLA.

We present PAI-Actor, a cinematic multi-character animation framework for character replacement in dynamic movie scenes. Unlike conventional animation systems that mainly drive a single static image or a single subject, our goal is to replace and animate multiple characters within real video clips while preserving the original scene dynamics, camera motion, and background content. This setting is particularly challenging because the generated characters must remain consistent with the source performance in motion and interaction, while also matching the surrounding background in lighting, shadow, composition, and overall cinematic appearance. To address this, we formulate multi-character animation as a structure-guided human recovery problem and build a movie-driven training pipeline from high-quality film data. Furthermore, to support practical cinematic production, we introduce a bidirectional-to-autoregressive distillation framework: we first train a bidirectional diffusion transformer for high-quality short-clip generation at 1080P resolution, and then distill it into an autoregressive video-to-video model for efficient inference and longer video generation. Experiments show that PAI-Actor enables high-fidelity multi-character animation with strong scene consistency, cinematic visual quality, and efficient long-form generation.

Large vision-language models have significantly advanced GUI agents, enabling executable interaction across web, mobile, and desktop interfaces. Yet these gains largely rely on a forgiving \emph{region-tolerant} paradigm, where many nearby pixels inside the same component remain valid. Precise geometric construction breaks this assumption: actions must land on points in continuous canvas space rather than tolerant regions. Because geometric primitives carry ontological dependencies, a local coordinate error can induce cascading topological failures that distort downstream objects and invalidate the final construction. We identify this regime as \emph{precision-sensitive GUI tasks}, requiring point-level accuracy, geometry-aware verification, and robustness to dependency-driven error propagation. To benchmark it, we introduce \textsc{PAGE} Bench, with 4{,}906 problems and over 224K process-supervised, pixel-level GUI actions. We further propose \textsc{PAGER}, a topology-aware agent that decomposes construction into dependency-structured planning and pixel-level execution. Pixel-grounded supervised tuning establishes executable action grammar, while precision-aligned reinforcement learning mitigates rollout-induced exposure bias through state-conditioned geometric feedback. Experiments reveal a pronounced ``Semantic-Execution Gap'': general multimodal models can exceed 88\% Action Type Accuracy yet remain below 6\% Task Success. \textsc{PAGER} closes this gap, delivering $4.1\times$ higher Task Success than the strongest evaluated general baseline and raising Step Success Rate from below 9\% for GUI-specialized agents to over 62\%, establishing a new state of the art for point-precise GUI control.

Vision-language-action (VLA) models have made rapid progress in robotic manipulation, yet the policy acts with no sense of the *pace* at which its own execution unfolds—whether it is advancing, stalling, or has quietly gone wrong—especially for long-horizon tasks. Recent methods supply progress through external reward models or as an auxiliary prediction inside the policy, but the signal runs alongside action generation rather than actively shaping it. Moreover, progress is inherently tied to the terminal state, making estimation fragile when the policy lacks an internal awareness of where the task should end. We propose **PACE**, which actively internalizes task progress as a predictive interface bridging commitment and execution, equipping the policy with a calibrated sense of the *pace* at which it advances toward task completion. Specifically, we first equip the VLA with a *foresight anchor*, a predicted terminal representation from observation and instruction, from which progress emerges as a geometric projection. Building on this anchor, the model perceives a compact past–present–future progress state capturing *how much the last step advanced*, *how far execution has come*, and *how far the next step should go*. These predicted progress further actively conditions the action decoder to guide the prediction, while the gap between committed and realized increments provides a built-in consistency check for lightweight failure detection. Experiments across LIBERO, SimplerEnv, and real-world single-arm and bimanual platforms show that PACE achieves 98.7% on LIBERO and 74% on real-world manipulation, with consistent gains especially on long-horizon tasks. Robustness and stable generalization across diverse scenarios further confirm the advantage of internalizing progress within the policy, validating a calibrated sense of execution pace as an effective paradigm for long-horizon robot control.

Native multimodal large language models (native MLLMs) jointly encode visual and linguistic sequences within a unified autoregressive Transformer, emerging as a promising paradigm for multimodal understanding. Unlike conventional ViT-based MLLMs that rely on an external vision encoder, ViT-free native MLLMs tokenize every pixel patch directly, which is intuitively expected to suffer from severe visual token redundancy. To examine this hypothesis, we conduct a pilot study and find that, unlike ViT-based models which degrade under compression, ViT-free MLLMs tolerate aggressive compression with near-lossless or even improved per9 formance. Existing compression methods are built exclusively for the ViT-based setting, whereas native MLLMs, which stand to gain the most from visual token compression, remain entirely unexplored. To close this gap, we introduce an analysis pipeline for native backbones, centered on two metrics, i.e., token discriminability and cross-dimensional heterogeneous redundancy, which together reveal the following two findings: (i) redundancy is distributed independently and heterogeneously along the height and width axes, and (ii) in early layers, visual tokens are indiscriminable and uniformly redundant. Guided by these findings, we propose PACE (Pareto Adaptive Compression for Efficient native MLLM), the first visual token compression framework tailored for native MLLMs with Multimodal-RoPE. Leveraging the dimensional decoupling of position encoding, PACE decomposes attention into two orthogonal spatial saliency estimates along height and width, gated by a positional attention signal, yielding a direction-aware importance measure with negligible overhead. It further casts token budget allocation as a Pareto optimization over three objectives, i.e., performance preservation, inference efficiency, and spatial coverage, which jointly identify the optimal compression depths and retention ratios. Experiments show that PACE maintains near-lossless performance at only 25%–50% visual token retention, and even achieves performance gains on several benchmarks, while delivering over 2× FLOPs reduction.

Diffusion large language models (dLLMs) such as LLaDA and Dream now rival autoregressive LLMs in quality while retaining native parallel decoding. To deploy them efficiently, block-wise decoding partitions generation into blocks of size $B$ and commits a fraction of each block before moving on, but $B$ conflates two roles: \emph{how far the model can look ahead} and \emph{how many tokens get committed per step}. Recent accelerators relax this trade-off with indirect heuristics, yet the underlying difficulty is already exposed at every forward pass by the model's per-step confidence. Specifically, the in-window confidence profile exhibits a context-dependent \emph{cliff} (a high plateau, a sigmoidal transition, and a residual floor) that directly encodes how far ahead is safe to look. We propose \textbf{PACE-dLLM}, a decoder that handles the two roles independently. At each step, PACE-dLLM fits the cliff in closed form on the current window's confidences and reads off the next prediction horizon at the cliff's saturation point; a separate confidence threshold decides which tokens are committed. Both rules read the same signal, add a single hyperparameter, and admit a strict-dominance guarantee over fixed-block decoding. Extensive experiments across four reasoning and code benchmarks on two open-source dLLM backbones demonstrate that PACE-dLLM achieves the best average accuracy at a roughly $\mathbf{4.5\times}$ wall-clock speedup over the unaccelerated semi-AR baseline, pushing the Pareto frontier outward.

Reasoning LLMs produce longer outputs, requiring speculative decoding drafters trained on extended sequences. Parallel drafting—predicting multiple tokens per forward pass—offers latency benefits over sequential generation, but training complexity scales quadratically with the product of sequence length and parallel positions, rendering long-context training impractical. We present P(arallel)-EAGLE, which transforms EAGLE from autoregressive to parallel multi-token prediction via a learnable shared hidden state. To scale training to long contexts, we develop a framework featuring attention mask pre-computation and sequence partitioning techniques, enabling gradient accumulation \textit{within} individual sequences for parallel-prediction training. We implement P-EAGLE in vLLM and demonstrate speedups of 1.10×–1.36× over autoregressive EAGLE-3 across GPT-OSS 120B, 20B, and Qwen3-Coder 30B.

Examining existing benchmarks for grounding in 3D VLMs reveals a pervasive "semantic leakage" in the evaluation and prompting protocols: a heavy reliance on bounding boxes as spatial prompts or representation primitives. Because bounding boxes are highly informative of object categories, they introduce a shortcut for the models to neglect visual information during 3D VLM training. We demonstrate this with a "blind" LLM that, prompted only with the box coordinates, can perform surprisingly well. To further analyze this shortcut, we propose point-based grounding prompting - querying models with a single 3D point instead of a box. Under this protocol, the accuracy of the blind LLM drops from $\textbf{36.4}$% to $\textbf{18.0}$% on ScanNet's Nr3D, and similarly degrades for 3D VLMs (e.g. LL3DA). Building on this insight, we introduce $P^3$-VLM, a novel tri-pathway 3D VLM that supports point-based prompts and achieves $\textbf{52.1}$% state-of-the-art accuracy on location-grounded captioning under this setting. To achieve this, we redesign the architecture while keeping it detector-free, scene-centric, and suitable for 3D GS inputs. $P^3$-VLM utilizes three complementary pathways (location-, task-, and context-aware), enabling strong reasoning about scenes without any bounding box dependencies. The key benefit of point-based learning is a consistent improvement across settings: $P^3$-VLM achieves state-of-the-art performance under point-based grounding while also improving robustness, OOD generalization, and performance on ungrounded scene-centric 3D VQA benchmarks, reflecting its reduced reliance on bounding-box shortcuts and resulting stronger visual-spatial reasoning. Our source code and models will be made publicly available.

Embodied AI agents increasingly require parallel execution of multiple tasks, such as manipulation, conversation, and memory construction, from shared observations under distinct time constraints. Recent Mixture-of-Transformers (MoT) Vision-Language-Action Models (VLAs) architecturally support such heterogeneous outputs, yet existing inference systems fail to achieve efficient multi-task parallelism for on-device deployment due to redundant computation and resource contention. We identify isolated KV cache management as the root cause. To address this, we propose *unified KV cache management*, an inference design that treats KV cache as a first-class shared resource across tasks and over time. This abstraction enables two key optimizations: *cross-task KV sharing* eliminates redundant prefill of shared observations, while *cross-frame continuous batching* decouples variable-length language decoding from fixed-rate action generation across control cycles. We implement this design for $\pi_{0.5}$, the most popular MoT VLA, and evaluate on both NVIDIA GeForce RTX 4090 and Jetson AGX Thor, two representative platforms for on-device VLA inference. \nickname achieves up to 3.7$\times$ speedup over isolated execution, delivering over 200 tokens/s language throughput and 70 Hz action frequency simultaneously without action quality degradation, and we further validate the gains on a real humanoid robot with on-board Jetson AGX Thor.

Differentiable logic gate networks, which operate using only logic gates, have recently attracted attention as an efficient alternative to conventional neural networks. However, despite their efficiency, the scaling behavior of logic gate networks remains underexplored. By contrast, scaling model capacity is a central design principle in deep neural networks and typically leads to improved performance. This discrepancy raises a key question: Can similar scaling benefits also be achieved in logic gate networks? In this work, we focus on width as a primary scaling axis and conduct a systematic analysis of its behavior in logic gate networks. We observe that naive width scaling often introduces redundancy among logic kernels, limiting the effective use of additional kernels and leading to performance saturation. To address this limitation, we propose a dynamic logic kernel framework that reorganizes kernel utilization by promoting specialization across kernel groups. This enables the network to better utilize increased width via input-dependent kernel routing, while ensuring that both routing and computation are implemented entirely with gate-level Boolean operations at inference time. We further find that kernel redundancy is most pronounced at the first gate level, motivating an early-stage dynamic logic kernel strategy that concentrates adaptation at this level. Experimental results demonstrate that our approach improves kernel utilization and increases kernel diversity, leading to higher accuracy with improved parameter efficiency.

Precipitation downscaling over complex terrain is essential for resolving water availability, flood risk, and extremes, but current evaluations rarely test whether models learn orographic structure rather than only pixel-wise accuracy. We introduce OroPrecipBench, a km-scale benchmark for orographic precipitation downscaling across five U.S.\ terrain regimes. The benchmark provides a shared-domain dual-track protocol: a perfect-model track that coarsens 3\,km HRRR analyses to isolate controlled super-resolution skill, and an observation-grounded track that downscales 25\,km ERA5 to 4\,km PRISM to test practical robustness under source-target shift. We also introduce PEP-D, a terrain-conditioned diagnostic that measures errors in elevation--precipitation profiles, complementing pointwise, spatial, and extreme-precipitation metrics. Evaluating ten statistical, deterministic deep-learning, and generative baselines shows that strong controlled super-resolution does not reliably translate to real-case or cross-regime downscaling. Models that reduce pixel-wise error can still distort terrain-dependent precipitation and upper-tail rainfall, and terrain-conditioned structure is especially fragile under transfer. OroPrecipBench exposes these failure modes and motivates source-shift-robust, terrain-constrained, and tail-aware downscaling methods. Benchmark is available \href{https://anonymous.4open.science/r/OroPrecipBench-612D}{here}.

General-purpose embodied AI systems are increasingly expected to interpret multimodal instructions and perform fine-grained physical tasks. However, many manipulation benchmarks provide limited support for jointly evaluating procedural correctness, local geometric sensitivity, and human-guided correction. We propose origami as a benchmark based on real images for procedural state-transition reasoning in embodied AI. Origami requires precise operations such as tucking, inserting, reversing, and folding overlapping layers, making it a focused domain for studying physical state transitions. We introduce a dataset of step-by-step folding images, textual entries, and difficulty annotations, and define two tasks: autonomous state transitions from multimodal instructions and human-guided collaborative transitions. Unlike crease-pattern- or diagram-based origami benchmarks, our setting exposes agents to actual intermediate folding states. For an initial simulator-based evaluation, we use a multimodal LLM controller to evaluate procedural understanding, action-interface grounding, and retry-based correction.

DETR-style detectors use one-to-one bipartite matching during training to assign object queries to ground-truth objects, enabling end-to-end set prediction without non-maximum suppression (NMS). However, without an explicit de-duplication procedure, multiple queries can still produce highly similar hypotheses for the same object, making training unstable and predictions less decisive. Inspired by the sequential ordering of NMS, we propose DETRNN, a plug-and-play module that turns unordered object queries into a competition-aware sequence for recurrent refinement. DETRNN builds an explicit confidence-and-similarity based order from prior predictions, then refines queries with an RNN along this order to model competition inside the decoder. This ordered recurrent refinement reduces redundant predictions, stabilizes optimization, and improves final detection accuracy. Experiments on multiple DETR-style detectors show consistent gains with comparable efficiency. Code will be released on GitHub.

Pretrained vision transformers provide strong semantic representations, but their token embeddings are not designed to transform predictably under image rotations. This limits their direct adaptation to domains such as satellite imagery, histopathology, microscopy, and other scientific imaging settings, where in-plane orientation is often arbitrary and no canonical pose exists. Existing equivariant adaptation methods for pretrained models typically rely on frame averaging or canonicalization, which can be effective on synthetically rotated natural images but do not teach the adapted model an internal rotation-aware representation. We introduce OrbitLoRA, a lightweight orbit-aware adapter for frozen visual foundation models. OrbitLoRA combines standard LoRA with a typed residual branch that maintains scalar and vector fields over ViT patch tokens. The typed branch is initialized from local steerable image features and harmonic coordinate fields, updated at selected transformer depths, and injected into the frozen token stream through zero-initialized residual gates. During finetuning, paired rotated views supervise the typed state through a transport-and-rotation consistency loss, while the semantic output is trained to remain invariant for the downstream task. This design preserves the rich features of pretrained ViTs while adding a learnable rotation-symmetry prior. Across classification and segmentation tasks, and across DINOv2-, CLIP-, SAM-, and LLaVA-style backbones, OrbitLoRA matches strong LoRA-with-augmentation baselines in high-data regimes and substantially improves adaptation in low-data regimes, where learning the symmetry prior is most valuable.

Squared Wasserstein distance is a frequently used tool to compare probability distributions. This distance is typically computed between empirical measures of size $n$ from two underlying random samples. Unfortunately, even in lower dimensional Euclidean space problems $\left( d \in \{2,3\} \right)$, Wasserstein distance algorithms with approximate or exact precision guarantees scale poorly in the runtime as a function of $n$ and the desired precision. In response, we consider the computational-statistical runtime, where the goal to estimate from samples the Wasserstein distance up to the $\varepsilon$-additive error which is achievable under a sample; we allow $O(1)$ time per sample. Towards this, we develop a Sample-Sketch-Solve paradigm where we introduce a regular cartesian grid sketch of the samples. We show that (especially under $\alpha$-H\"older smooth distributions) this can compress the data without increasing asymptotic error, and also regularizes the structure which enables faster exact algorithms. Ultimately, we approximate $W_2^2(P,Q)$ within $\varepsilon$ error in $\varepsilon^{-\max(2,\frac{d+1+o(1)}{2})}$ time for Lipschitz $P,Q$ on $(0,1)^d$; nearly an optimal $\Theta(\varepsilon^{-2})$ for $d \leq 3$.

Retrieval-augmented generation (RAG) is bottlenecked by the token budget of large language models: feeding long documents into the context window is expensive, while aggressive textual compression discards fine-grained information that is irreversibly lost in a one-dimensional token sequence. We argue that this bottleneck is not fundamental to information density but to the choice of compression \emph{modality}. Modern visual encoders can pack the contents of a rendered text page into roughly a hundred visual tokens with little semantic loss, opening a new compression-fidelity tradeoff that purely textual methods cannot reach. We present \textbf{OpticalRAG}, a framework that renders documents as page images, encodes them with a frozen vision encoder, and distills the retrieved visual tokens into as few as $16$ injected tokens before passing them to the LLM. Two ingredients make this work without expensive cross-modal training: (i) \emph{Encoder-Bridged Retrieval}, which reuses the shared representation space of a pretrained vision-language model to retrieve visual chunks from a text query, with no contrastive training; and (ii) \emph{Query-Driven Distillation}, a lightweight cross-attention compressor that condenses thousands of visual tokens into a small set of query-conditioned prefix tokens. On LongBench-v2, BAMBOO, and LooGLE-v2, OpticalRAG achieves the best compression-aware accuracy (CAP) on all three benchmarks while injecting only $16$ tokens, and improves the closed-book Qwen2.5-7B-Instruct backbone by $5.17$ points on LongBench-v2.

Detecting LLM reasoning failures at inference time without ground-truth labels motivates a family of confidence baselines, including self-consistency, semantic entropy, and $P(\text{True})$, built on within-question sampling and self-evaluation. Operad theory, the formalism for systems built by iterated substitution, suggests a complementary diagnostic: a model's direct answer to a compositional query should agree with the answer it produces by composing a stated decomposition of the same query. We instantiate this idea as *operadic consistency* (OC), a per-question signal. Across twelve instruction-tuned LLMs (4B--671B parameters, open-weights and closed-source) on four multi-hop QA datasets, OC is strongly correlated with accuracy on every dataset (Pearson $r \in [{+}0.86, {+}0.94]$, all $p \leq 0.0004$), making it a substantially stronger population-level accuracy proxy at roughly one-third the inference cost of the strongest sample-based baseline (whose own cross-model rates never exceed $r{=}{+}0.80$ on any dataset). At the per-question level, OC contributes information beyond self-consistency, semantic entropy, $P(\text{True})$, and constructed decomposition-aware baselines on every dataset (cluster-robust $p \leq 10^{-19}$). We translate the regression coefficients into deployment-ready selective-prediction lifts ($\Delta_{\text{AUARC}}$ up to $+0.088$, $\Delta_{\text{AUROC}}$ up to $+0.167$ over a tuned $\text{SC}_{K=3}$ baseline). On five frontier thinking models, where the decomposition is extracted from the model's own chain of thought, the same equal-cost comparison gives positive selective-prediction lift on all $16$ (dataset, budget, metric) cells tested.

Recent multimodal large language models (MLLMs) show great potential in natural image understanding. Yet, they perform well, mainly on reasoning in-view contents within the image frame. This paper presents the first study on out-of-view (OOV) understanding, i.e., the ability to reason about objects, activities, and scenes beyond the visible frame of a perspective view. Our technical contributions are threefold. First, we design OpenView, a four-stage pipeline that leverages panoramic imagery for large-scale and diverse multi-choice VQA synthesis, where full-scene coverage enables flexible view framing while preserving global scene awareness. Second, we curate OpenView-Dataset, a high-quality synthetic dataset from diverse real-world panoramas to empower MLLMs upon supervised fine-tuning. Third, we build OpenView-Bench, a benchmark that jointly measures choice and rationale accuracy for interpretable and diagnosable evaluation. Experimental results show that despite having a large gap from human performance in OOV VQA answer selection, multiple MLLMs, upon empowered by OpenView, can consistently boost their performance, uplifted from 48.3% to 62.3% on average. Code, benchmark, and data will be released.