NeurIPS 2026 Papers

Streaming 3D reconstruction from long monocular video sequences requires maintaining a key-value(KV) cache that grows linearly with sequence length, creating a severe memory bottleneck. Existing approaches either truncate the cache to a fixed set of anchor frames, leading to reconstruction quality degradation, or rely on attention-score heuristics that are agnostic to 3D scene structure,failing to preserve geometrically valuable tokens. To address these problems, we present GHOST (Geometry-Hierarchical Online Streaming Token Eviction), a training-free KV cache management framework that exploits the model's own 3D geometry outputs to evict redundant tokens online. GHOST introduces three mutually reinforcing innovations: a hierarchical dual-level importance scoring scheme, a privilege mechanism that protects special tokens from eviction, and a cosine-similarity-guided layer-wise budget allocation. Experiments on various benchmarks show that GHOST preserves excellent reconstruction quality while cutting the KV cache by nearly half and delivering $1.75\times$ faster inference compared to state-of-the-art methods. Our code and model will be available to the public.

Feed-forward 4D Gaussian Splatting (4DGS) offers an efficient paradigm for reconstructing dynamic scenes from monocular videos, bypassing the need for lengthy per-scene optimization. However, existing methods suffer from three fundamental bottlenecks: redundant frame-wise grid-aligned Gaussian representations, suboptimal single-step 2D-to-4D regression architectures, and restrictive dependencies on 3D/4D signals during training or inference. In response to these limitations, we introduce a novel Gaussian-Grounded Query Refinement (GGQR) framework. To reduce representation redundancy, GGQR learns a compact set of Gaussian queries equipped with a newly designed plateau-shaped temporal kernel, enabling single primitives to persistently model motion-consistent regions across extended frames. To alleviate regression ambiguity, these queries are iteratively refined through a Gaussian-grounded attention mechanism operating in 3D space. Specifically, GGQR leverages foundation models to lift 2D observations into a shared 3D space, explicitly grounds the 3D attention in intermediate Gaussian states, and residually updates those Gaussian states. Trained entirely self-supervised on casual videos, GGQR achieves a new state-of-the-art on dynamic reconstruction benchmarks among feed-forward 4DGS methods, delivering superior performance in novel view-time rendering alongside downstream motion modeling tasks.

Whether a video is physically plausible is a question current generative models answer poorly and current evaluators answer expensively, through physics-targeted training, billion-parameter video world models, or multimodal-language model judges. We ask whether frozen images vision encoders, with no video or physics supervision, already contain a useful signal for the same question. Across four frozen backbones, self-supervised vision transformers (DINOv2, DINOv3) and biologically-inspired models of the ventral stream (CORnet-S, VOneNet), a small set of geometric signals (curvature, speed variation, acceleration, prediction residual) on per-frame feature trajectories separates plausible from physics violated videos with no retraining. We propose GEOPHYS, a training-free framework that 11 reaches 97.7% on LikePhys and 93.3% on IntPhys2 accuracy for physics-violation detection, surpassing V-JEPA 2, GPT-4o, Gemini, and twelve modern video diffusion models. The same signals track human EEG responses to object-permanence violations and scale with object number. Deployed unchanged as a best-of-N reward during video generation, GEOPHYS lifts MAGI-1 4.5B from 53.8\% to 64.7\% PhysicsIQ score at 5.8× lower wall-clock and 7.8× lower memory than V-JEPA 2 world-model rewards. A useful proxy for physical plausibility is already encoded as a geometric regularity of natural-video statistics in vision features.

Accurate characterization of deep subsurface structures is a fundamental problem in geoscience, but geophysical inversion is inherently non-unique and therefore requires joint interpretation of gravity, magnetic, electromagnetic, and seismic observations. Despite recent advances in deep learning for geophysical inversion and PDE-based scientific modeling, existing large-scale open datasets are largely limited to a single modality and rarely provide physically consistent multi-source responses generated from the same 3D geological target. This limitation hinders the development and evaluation of multimodal joint inversion methods for deep subsurface characterization. We introduce GEMS-3D, a large-scale synthetic benchmark for 3D geological multiphysics learning. GEMS-3D couples four forward modeling engines—$\textbf{g}$ravity, $\textbf{e}$lectromagnetic, $\textbf{m}$agnetic, and 3D acoustic $\textbf{s}$eismic—to generate sample-aligned observations from shared geological realizations. Guided by facies-aware rock-physics relationships, the dataset provides spatially aligned volumes of P-wave velocity ($v_p$), density ($\rho$), resistivity ($\phi$), and magnetic susceptibility ($\chi$) across nine representative classes of complex geological anomalies, including dike swarms, salt domes, and gas reservoirs. Each sample includes gravity responses, magnetic responses, electromagnetic responses, and 3D acoustic seismic data, together with acquisition metadata and anomaly labels. By packaging co-registered property volumes, anomaly support, acquisition metadata, and aligned multiphysics responses as sample-level bundles, GEMS-3D provides a unified benchmark for operator learning, multimodal inverse interpretation, and missing-modality robustness in deep subsurface settings. The anonymized forward-modeling framework and corresponding dataset link is available at https://anonymous.4open.science/r/GEMS-3D-E4CA.

Large language models encode rich representations in their hidden states, yet these representations remain largely opaque and disconnected from the model's vocabulary. We introduce Geometric Embedding Mixture (GEM), a pre-training regularization method that aligns transformer hidden states with the geometry of the vocabulary embedding space. GEM utilizes a self-supervised objective to encourage the final hidden state to approximate a probability-weighted mixture of token embeddings. This effectively minimizes the free energy of the models and pushes the representation into the convex hull of the vocabulary. We pre-train models at three scales (135M, 360M, and 760M) and demonstrate that GEM yields superior predictive uncertainty compared to standard training, reducing perplexity by up to 53\% on benchmarks. Furthermore, we show that this geometric structural constraint mitigates representation anisotropy and unlocks \emph{native interpretability}. Unlike baseline models, GEM enables direct layer-wise decoding of intermediate hidden states and better causal faithfulness using the untuned language head, revealing coherent semantic trajectories without the need for auxiliary probes. Our work demonstrates that imposing geometric structure during pre-training improves both performance and transparency, offering a path toward language models that are interpretable by construction.

Real-world networks typically exhibit small-world properties: dense local clustering yet efficient global propagation. However, current methods struggle to reconcile this hierarchy: Graph Neural Networks (GNNs) suffer from over-smoothing in long-range modeling, while sequence/state-space models often lack the inductive bias to preserve complex graph topology. To bridge this gap, we propose Graph Encoding Mamba (GEM), a dual-scale architecture for graph-level hierarchical representation learning. First, GEM employs a dual-branch encoder that synergizes the local sensitivity of GNNs with the selective scanning of Mamba, effectively capturing both local topology and global shortcuts. Second, a modularity-aware hierarchical pooling mechanism is designed to retain salient community structures during graph coarsening. Crucially, to mitigate structural redundancy from multi-scale fusion, we incorporate an Information Bottleneck objective to distill task-relevant structural patterns from redundant features, encouraging compact, task-relevant representations and supporting intrinsic interpretability. Extensive experiments on graph-level benchmarks demonstrate that GEM improves performance across biological, social, and brain network datasets. We further provide a lightweight node-level adaptation on long-range benchmarks as a boundary analysis of GEM's transferability. Code is available at \url{https://anonymous.4open.science/r/GEM-C34D}.

Selective state space models (SSMs) such as Mamba-3 are efficient and increasingly competitive with Transformers. However, their fixed recurrent state remains a bottleneck for using context as temporary memory. To address this issue, we propose $ \textit{GEAR} $ (GEnerator-Adaptive state space models for associative Recall), a slow-hypernetwork that produces coefficients for LoRA updates to the SSM token-to-recurrence parameter generators. GEAR makes the parameter generator context-adaptive while preserving the original linear-time Mamba-3 scan. To evaluate whether this adaptation improves in-context recall, we pretrain 180M-parameter models for 2B FineWeb-Edu tokens and compare against Mamba-3 baseline and ablations at the same scale. In the Multi-Query Associative Recall (MQAR) task, models see $N$ in-context key-value bindings and must predict the value paired with a queried key. In compact MQAR evaluation where the adaptive path is active at the query, GEAR improves candidate-restricted NLL over Mamba-3 by 0.70, 0.69, and 0.48 at $ N=32,64,128 $, respectively. Beyond compact MQAR, GEAR reduces the penalty on older queried bindings and gives the clearest gain within the Mamba-3 family under moderate random-filler interference. These results show that context-adaptive SSM generator improves associative recall while preserving Mamba-3 language modeling capabilities.

We present a novel GPU-accelerated global solver for constrained nonlinear programming (NLP) inspired by the efficient linear bound propagation framework in neural network (NN) verification. Existing work on NN verification, a problem that can be cast as feasibility detection for a nonlinear objective, demonstrated orders-of-magnitude speedups with the linear bound propagation-based framework when compared to traditional solvers. Extending this paradigm to global nonlinear programming requires addressing many challenges: a global constrained NLP solver must effectively exploit all, possibly unstructured, nonlinear constraints to improve the objective, whereas NN verification typically handles simple input box constraints over a well-structured NN and does not directly optimize an objective function. We address these challenges by fully exploiting the linear bounds produced by bound propagation: these linear bounds can be viewed as relaxations of nonlinear constraints and used to provide dual objectives, check infeasibility, and guide the search for a primal solution. In addition, we proposed tighter linear relaxations (essential for linear bound propagation) for bilinear functions commonly used in NLP problems, and a projected augmented Lagrangian method tightly coupled to our dual solving procedure to produce primal solutions. Our method solves 399 of 505 bounded and continuous NLP problems in the GAMS and MINLPLib benchmarks under a 180-second limit, outperforming many strong global solvers (including commercial ones) such as SCIP, BARON, and LindoGlobal. Compared to SCIP (the strongest solver on these problems), we solve 104 problems exclusively, including large-scale, highly nonlinear ones that are often out of reach for traditional approaches. Our method demonstrates a distinct, highly complementary regime for GPU-accelerated global optimization of nonlinear programming problems.

Multimodal Large Language Models (MLLMs) are typically aligned with caption-pair data, but not all answer-token supervision is equally visual. In LLaVA-style alignment training, many tokens are predictable from language context alone, yet their losses can still update vision-related parameters. Such language-dominant updates may add noisy supervision for tokens that require visual evidence. To diagnose this vulnerability, we probe the model against visual counterfactuals to reveal the intrinsic grounding requirements of each token, distilling them into two soft indicators: visual necessity and evidence specificity. Guided by this diagnostic insight, we present $\textbf{GEAR-Align}$, a controlled post-training framework that makes supervision destination explicit. During training, it converts token-level visual reliance into parameter-level gradient allocation: all answer tokens update language-side parameters, high-necessity tokens update route-flow parameters, and the high-specificity subset additionally updates visual-evidence parameters. Across two controlled Qwen3-backbone settings, GEAR-Align improves hallucination-sensitive and local-evidence-heavy benchmarks while remaining competitive on general multimodal benchmarks. These results provide evidence for token-specific visual gradients as a useful principle for controlled MLLM post-training.

Reinforcement learning (RL) can be used to improve the policy (denoiser) of diffusion large language models (dLLMs), while being hindered by the intractability of the policy likelihood. A dominant and efficient family of methods replaces the likelihood in standard RL with its evidence lower bound (ELBO), estimated from randomly masked sequences. Despite being well-aligned with pre-training, these approaches introduce bias from training-inference mismatch by applying the likelihood surrogate ELBO, which can potentially cause training collapse. In this work, we propose **Guided Denoiser Self-Distillation (GDSD)** to directly distill the denoiser of dLLMs from an advantage-guided self-teacher, derived from the closed-form optimum of reverse-KL regularized RL. GDSD matches the dLLM's denoiser logits to the teacher's via a normalization-free objective, which reduces RL to likelihood-free self-distillation and thus bypasses the TIM biases. Recent ELBO-based methods emerge as instances of applying different distillation divergences, but with diagnosable pathologies that GDSD avoids. On planning, math, and coding benchmarks with LLaDA-8B and Dream-7B, GDSD consistently outperforms prior state-of-the-art ELBO-based methods with a more stable training reward dynamics, achieving test-accuracy improvements of up to $+19.6\\%$. These results suggest that direct denoiser self-distillation, without relying on ELBO likelihood surrogate, can provide a more stable and effective RL procedure for dLLMs.

Diffusion distillation, exemplified by Distribution Matching Distillation (DMD), has shown great promise in few-step generation but often sacrifices quality for sampling speed. While integrating Reinforcement Learning (RL) into distillation offers potential, a naive fusion of these two objectives relies on suboptimal raw sample evaluation. This sample-based scoring creates inherent conflicts with the distillation trajectory and produces unreliable rewards due to the noisy nature of early-stage generation. To overcome these limitations, we propose GDMD, a novel framework that redefines the reward mechanism by prioritizing distillation gradients over raw pixel outputs as the primary signal for optimization. By reinterpreting the DMD gradients as implicit target tensors, our framework enables existing reward models to directly evaluate the quality of distillation updates. This gradient-level guidance functions as an adaptive weighting that synchronizes the RL policy with the distillation objective, effectively neutralizing optimization divergence. Empirical results show that GDMD sets a new SOTA for few-step generation. Specifically, our 4-NFE (Number of Function Evaluations) models outperform the quality of their multi-step teacher and substantially exceed previous DMDR results in GenEval and human-preference metrics, exhibiting strong scalability potential.

Climate downscaling of global General Circulation Models (GCMs) is a challenging ill-posed inverse problem. In practice, the lack of paired data makes end-to-end supervised training infeasible, while downstream assessments require processing large, heterogeneous GCM ensembles under strict computational limits. Consequently, developing an efficient, zero-shot downscaling approach is crucial. While vanilla Diffusion Posterior Sampling has recently shown immense potential by bypassing paired training, it remains bottlenecked by geographic misalignments, spectral biases, and prohibitive inference costs. To address these limitations, we propose the GCM-consistent Diffusion for Zero-shot Downscaling (GCD) framework. First, GCD conditions the diffusion prior on static geographic boundaries to ensure physical fidelity. Second, we introduce a filter measurement operator, effectively mitigating spectral discrepancies between GCMs and the prior to provide stable inference-time guidance. Third, to overcome the inference bottleneck, we design a joint acceleration scheme: by combining progressive distillation with Conjugate Gradient guidance to correct large-step deviations, GCD compresses the generation process to just 6 steps. Extensive experiments demonstrate that GCD achieves highly competitive 99th percentile accuracy across five heterogeneous GCMs and successfully recovers the high-frequency details of tropical cyclones, entirely without model-specific retraining.

Off-policy reinforcement learning (RL) for large language models bottlenecks at the rollout stage: every parameter update invalidates the key-value (KV) cache of in-flight agent trajectories, forcing either expensive recomputation or staleness that destabilizes training. Recent advances such as M2PO and asynchronous RLHF tolerate token-level staleness by reweighting the policy gradient with second-moment importance corrections, but they implicitly assume that a stale rollout is sampled from a well-defined behavior policy $\pi_{\theta_{\text{old}}}$. We show that this assumption is silently violated by every modern partial-rollout system: once a stale KV cache is reused under updated parameters $\theta_{\text{new}}$, the resulting token distribution is a hybrid $\pi_{\text{hyb}}$ whose attention projections consume stale keys and values while its MLP, output, and gating projections use the new weights. The token-level importance weight $w_t = \pi_{\theta_{\text{new}}}(y_t \mid x_{<t}) / \pi_{\theta_{\text{old}}}(y_t \mid x_{<t})$ corrects the sampling distribution but not this hidden-state-level mismatch, leaving an uncorrected source of policy-gradient bias that grows with both staleness and weight-update magnitude. We introduce Gradient-Coupled KV Drift (GCD), an exception-safe correction comprising (i) a high-probability concentration bound on the drift, with explicit per-layer terms for full-attention and Gated DeltaNet layers; (ii) a tail-event detector that flags inputs outside the bound’s regime of validity; (iii) a composite safety guarantee that treats the bias correction as trustworthy only when both the analytical estimate and the detector agree; and (iv) a variance-preservation theorem showing that GCD composes with the M2PO trust region without re-tuning. On Qwen3.5-9B trained with GRPO on MATH-500 and $\tau$-bench, GCD enables stable training at staleness 512 (a regime where M2PO collapses), reduces rollout wall-clock by 38\% versus full recomputation, and lifts $\tau$-bench retail pass$^4$ by 5.7 points over the M2PO baseline.

Few-Shot Industrial Visual Anomaly Detection (FS-IVAD) is a critical task in modern manufacturing, where automated product inspection systems must identify rare defects using only a handful of normal, defect-free training samples. This paper introduces GATE-AD, a reconstruction-based framework that casts few-shot normality modeling as a masked, representation-aligned graph reconstruction problem on a $k$-nearest-neighbor graph, built from frozen self-supervised ViT patch tokens. Since defects typically disrupt the contextual consistency between a patch and its spatial neighbors, GATE-AD emphasizes such neighborhood relations by attending anisotropically over each patch's local neighborhood using a Graph Attention Network (GAT) encoder. To prevent GAT over-smoothing in the low-shot setting, the encoder output is aligned with the input ViT tokens through a learnable latent space, where reconstruction inconsistency is scored with a Scaled Cosine Error (SCE) objective. On the MVTec AD, VisA, and MPDD benchmarks, GATE-AD attains state-of-the-art image-level AUROC in the wide majority of $1$- to $8$-shot settings, with low per-image inference cost that does not scale with the size of the support set. Code is included in the supplement material and will be publicly released upon acceptance.

Fine-tuning diffusion models via online reinforcement learning (RL) has shown great potential for enhancing text-to-image alignment. However, since precisely specifying a ground-truth objective for visual tasks remains challenging, the models are often optimized using a proxy reward that only partially captures the true goal. This mismatch often leads to reward hacking, where proxy scores increase while real image quality deteriorates and generation diversity collapses. While common solutions add regularization against the reference policy to prevent reward hacking, they compromise sample efficiency and impede the exploration of novel, high-reward regions, as the reference policy is usually sub-optimal. To address the competing demands of sample efficiency, effective exploration, and mitigation of reward hacking, we propose \textbf{G}ated and \textbf{A}daptive \textbf{R}egularization with \textbf{D}iversity-aware \textbf{O}ptimization (\textbf{GARDO}), a versatile framework compatible with various RL algorithms. Our key insight is that regularization need not be applied universally; instead, it is highly effective to selectively penalize a subset of samples that exhibit high uncertainty. To address the exploration challenge, GARDO introduces an adaptive regularization mechanism wherein the reference model is periodically updated to match the capabilities of the online policy, ensuring a relevant regularization target. To address the mode collapse issue in RL, GARDO amplifies the rewards for high-quality samples that also exhibit high diversity, encouraging mode coverage without destabilizing the optimization process. Extensive experiments across diverse proxy rewards and hold-out unseen metrics consistently show that GARDO mitigates reward hacking and enhances generation diversity without sacrificing sample efficiency or exploration, highlighting its effectiveness and robustness.

Reconstruction error is the standard signal for unsupervised time-series anomaly detection and has been widely adopted by time-series foundation models (TFMs). It is, however, vulnerable to input noise: the resulting score tracks the noise level itself rather than whether a sample is truly anomalous, and this failure is especially pronounced in domains where the noise envelope itself carries information about the underlying process state. We propose the *adaptation gap* $\text{Diff} = L_{FT} - L_{ZS}$, the difference between fine-tuned and zero-shot reconstruction losses, as a noise-robust complement. We prove that its conditional expectation cancels the input-noise variance and yields a calibration-anchored, population-level mean-separation guarantee. Building on these results we introduce **GAPS** (Gradient-Aware Adaptation-Gap Scoring), which selectively augments $L_{FT}$ with $\text{Diff}$ under a calibration-only routing layer with no label-tuned hyperparameters. Across a synthetic suite, an HRV benchmark whose normal class is intrinsically noisier than its anomaly class, and TSB-AD-U with a pre-registered noise-stress condition, GAPS preserves performance under standard conditions and yields substantial gains under noise stress.

We present GAMMA, a scalable feed-forward model that reconstructs 4D Gaussian Splatting from monocular videos, enabling dynamic novel view synthesis in real-time. Most existing approaches for dynamic scene modeling require multi-view videos as input or costly per-scene optimization. In contrast, GAMMA learns a unified 4D scene representation from monocular videos and generatively predicts 4D Gaussian primitives within seconds. The key insight of GAMMA lies in its unified Gaussian reconstruction model, which jointly estimates temporally consistent geometry, appearance, and motion from the input video. We further explore its scalability through large-scale training on a comprehensive dataset encompassing both synthetic and real-world videos. The reconstructed GAMMA representation supports interactive scene exploration via real-time rendering across timesteps and local viewpoints. Extensive experiments demonstrate that GAMMA outperforms existing methods in both reconstruction fidelity and efficiency.

Graph-level anomaly detection plays a critical role in a wide range of applications, such as fraud detection in financial networks and malicious program detection, yet it remains challenging in real-world scenarios where labeled anomalous graphs are extremely scarce and graph structures are highly diverse. Existing few-shot methods often suffer from severe performance degradation due to the effects of class imbalance and structural imbalance, and they struggle to generalize to structurally deviant or rare anomalies. In this paper, we propose an adaptive few-shot graph-level anomaly detection model with multi-view structure-aware prompting (GADMVP). To mitigate structural imbalance, we construct a graph-of-graphs representation that facilitates cross-graph message passing, capturing both intra-graph semantics and inter-graph dependencies. Building upon these priors, we introduce a structure-aware graph prompting mechanism that extracts and adaptively refines anomaly-relevant prompts from a small set of labeled graphs. Extensive experiments on multiple benchmark datasets demonstrate that the proposed method consistently outperforms state-of-the-art few-shot graph anomaly detection baselines, while exhibiting strong robustness under extremely limited labeled data. The code and implementation details are available at https://anonymous.4open.science/r/GLAD-3456543/.

Real-world multi-modal image fusion faces two fundamental challenges: spatial misalignment and complex degradations induced by heterogeneous imaging sensors. These factors are intrinsically coupled, but most existing fusion methods typically address only one of them, inevitably leading to failure under the other challenge. In this paper, we propose G2Fusion, the first fusion framework that simultaneously addresses image registration and information restoration. By developing a novel geometric-to-generative paradigm, it can directly produce high-quality fused images from unregistered and degraded inputs captured by heterogeneous imaging sensors. This framework consists of two key modules: a flow-based geometric deformation reduction module (Flow-GDR) and a DiT-based generative fusion module (DiT-GF). The former reduces large non-rigid discrepancies through dense flow estimation. The latter performs generative fusion, progressively refining residual misalignments while restoring degraded content through iterative denoising. To enable effective interaction among registration, restoration, and fusion, we design two complementary mechanisms. On the one hand, a target distribution mining strategy is introduced to construct a joint objective distribution from registration, restoration, and fusion, effectively guiding the optimization of DiT-GF. On the other hand, we develop a mutual promotion mechanism that establishes a closed-loop interaction between Flow-GDR and DiT-GF by re-estimating the residual deformation between the fused output and the infrared reference. Extensive experiments demonstrate that G2Fusion consistently outperforms state-of-the-art methods in terms of registration, restoration, and fusion.