NeurIPS 2026 论文列表

Understanding the 3D physical world has emerged as an essential component in vision-language-action models (VLAs), enabling robots to perform several challenging tasks by unifying \textit{spatial} perception, language, and action. Existing approaches typically either rely on explicit 3D inputs, such as depth maps or point clouds, or implicitly inject spatial knowledge into visual representations. However, explicit 3D inputs are often brittle in practice due to sensor noise, hardware heterogeneity, and incomplete depth coverage, while implicit alignment can distort the original 2D visual representation and weaken generic scene understanding. In this paper, we present SPARK-VLA, a simple yet effective framework to allow VLAs to implicitly possess generic 2D and spatial 3D visual comprehension within a unified visual backbone. SPARK-VLA consists of two key components: (1) spatial representation distillation, which transfers 3D understanding from a frozen 3D expert into a LoRA-adapted visual branch by aligning attention responses while preserving the original visual pathway, and (2) a lightweight knowledge router, which selectively forwards action-relevant visual tokens to the language model using routing supervision derived from action-loss gradients. Experiments in both simulation and real-world environments show that SPARK-VLA, with only a 0.5B language backbone and 50M trainable parameters, achieves performance comparable to or better than substantially larger VLAs.

Why must vision-language navigation be bound to detailed and verbose language instructions? While such details ease decision-making, they fundamentally contradict the goal for navigation in the real-world. Ideally, agents should possess the autonomy to navigate in unknown environments guided solely by simple and high-level intents. Realizing this ambition introduces a formidable challenge: Beyond-the-View Navigation (BVN), where agents must locate unseen targets without step-by-step guidance. Existing large language model (LLM)-based methods, though adept at following dense instructions, often suffer from short-sighted behaviors due to their reliance on short-horizon supervision. Simply extending the supervision horizon, however, destabilizes LLM training. In this work, we identify that video generation models (VGM) inherently benefit from long-horizon supervision to align with language instructions, rendering them uniquely suitable for BVN tasks. Capitalizing on this insight, we propose applying VGM as the main backbone to generate actions for the first time in this field. Yet, the prohibitive latency for generating videos spanning tens of seconds makes real-world deployment impractical. To bridge this gap, we propose SparseVideoNav, achieving sub-second trajectory inference guided by a generated sparse future spanning a 20-second horizon. This yields a remarkable 27$\times$ speed-up compared to the unoptimized counterpart. Extensive real-world zero-shot experiments demonstrate that SparseVideoNav achieves 2.5$\times$ the success rate of state-of-the-art LLM baselines on BVN tasks and marks the first realization of such capability in challenging night scenes.

Domain generalised semantic segmentation (DGSS) has recently benefited from combining multiple foundation models, such as vision-language and vision foundation models, to leverage their diverse representations for generalisation. Existing multi-model approaches share a common design choice: cross-model fusion is restricted to depth-aligned layers, implicitly assuming that mutually beneficial information resides at the same depth across models. However, foundation models pretrained under different objectives may develop distinct representation hierarchies, making the optimality of depth-aligned fusion questionable. Moreover, we observe that the foundation models have different depth-wise domain sensitivity. Motivated by these observations, we propose a Sparse All-Layer Connector (SALC) that enables each layer to access information from preceding layers of another model. As the number of accessible layers grows with depth, dense aggregation may hurt generalisation. SALC therefore learns to adaptively select a sparse subset of informative layers. We further introduce a candidate dropout regularisation that strengthens sparsity and encourages SALC to explore diverse layers, leading to more robust selection. Across four foundation model combinations and three DGSS evaluation settings, the proposed design demonstrates stronger generalisation than conventional depth-aligned fusion. Code will be public upon acceptance.

Despite remarkable progress in text-to-image generation, current Diffusion Transformers (DiTs) inherently struggle with learning intricate visual content, such as accurate multilingual scene text rendering and novel visual concepts. This limitation stems from the inherent contradiction between vague semantic guidance provided by text and the objects of accurate reconstruction during diffusion training. To address this bottleneck, we propose SpaG-DiT, a novel spatial grounding training framework. SpaG-DiT directly injects explicit spatial priors into the DiTs before the attention module via Contextual Diagonal Position Encoding (CDPE), which dynamically modulates spatial coordinates without destroying the inherent sequential structure of the linguistic input. Furthermore, we introduce VLM Grounding Guidance (VLM-GG) as an automated pipeline to extract implicit spatial priors from pre-trained Vision-Language Models, eliminating human annotation costs. To evaluate the grounding ability, we establish two new datasets: SpaG-DiT-MultiLingual for non-Latin/Chinese text rendering and SpaG-DiT-Creature for new concept learning. Extensive experiments demonstrate that SpaG-DiT achieves superior training efficiency and significantly outperforms vanilla self-finetuning on the above two tasks and the widely adopted GenEval benchmark. These results validate the plug-and-play nature of SpaG-DiT and its potential to enhance DiT training processes. Code, model, and data will be released.

Vision Transformers are increasingly applied to data defined on the sphere, particularly in physics, meteorology, and related scientific domains. Position embeddings provide the spatial information that attention itself does not encode, making their design central to the expressive power of Transformers. However, most existing position embedding schemes are designed for Cartesian grids and therefore do not naturally handle longitude periodicity, polar singularities, or geodesic relations on spherical domains. This mismatch limits the ability of standard attention to model spherical geometry without specialized architectural modifications. We propose Spherical Reflection Position Embedding (SpRePE), a drop-in position embedding scheme for Vision Transformers on spherical data. SpRePE encodes each absolute spherical position by applying Householder reflections to query and key representations. Although each token is encoded only from its own absolute coordinates, the resulting attention inner products induce an explicit spherical relative-position term with a clear geometric interpretation. This formulation injects sphere-aware relative geometric information into standard attention without constructing pairwise attention-bias matrices, introducing task-specific modules, or modifying the backbone architecture. It also avoids the quadratic overhead of relative position biases and preserves the same asymptotic overhead as RoPE. We evaluate SpRePE on spherical image classification, panoramic depth estimation, and global weather forecasting. Across these tasks, SpRePE achieves competitive performance compared with strong position embedding baselines, with particularly clear gains in settings where spherical geometry is important.

Open-weight Vision-Language Models (VLMs) score 85--94\% on existing spatial benchmarks (BLINK Spatial, What's Up CLEVR) yet collapse to chance on simple geometric questions when object, scene, and texture priors are stripped away. We introduce and open-source Sophon, a procedurally generated diagnostic of paired colored point-cloud stimuli across three coordinate systems (1D linear, 2D Cartesian, 2D polar) and four question types (binary, categorical, comparative, quantitative). Across four open architectures spanning every major projector design (LLaVA-NeXT, LLaVA-OneVision, Qwen2.5-VL, Idefics3), accuracy clusters at 47--52\% while frontier models (Opus 4.6, Gemini 2.5 Pro, GPT-5.4) clear 77--83\% and humans reach 84--94\%. We trace the open-weight failure end-to-end. The vision tower works: encoder probes recover Sophon spatial direction at 80--97\%. The language tower's reasoning circuit works: a text-only control restores $+40$ pp, and probes preserve the spatial signal through every LLM decoder layer at 80--86\%. Per-layer attention attribution localizes visual integration to mid-stack layers. Yet at the readout, the model elevates the correct answer-set tokens but selects among them at chance. Fine-tuning experiments rule out the unembedding as the mechanism. By moving the unembedding layer through the residual stream, we observe the emergence of a small but discriminative contrast at late-stack residual layers in fine-tuned models, which in practice converts the random selection into a confident one more likely to be correct. However, ablation experiments locate the mechanism upstream: training only the upstream layers with the late stack frozen recovers the full $+29$ pp gain, while training only the late stack captures less than a third. The late-stack contrast is the downstream readout of upstream changes. No layer in either base or fine-tuned models shows geometric alignment between the spatial signal and the unembedding's vocabulary axes, suggesting that 7B-class VLMs have the weights to reason spatially but lack the residual-stream pathways to convert that reasoning into answers.

Large language models (LLMs) can solve every intermediate step of a multi-step math problem on its own and still fail the full problem, even when given a roadmap of the steps and all of their answers. We introduce OracleLadder, a diagnostic evaluation that locates where LLM math reasoning fails by giving the model increasing levels of oracle help. For each problem, a teacher model writes a fixed roadmap of intermediate sub-goals (milestones), and a deterministic symbolic verifier grades every answer. Testing the model with no help, with the roadmap, with the roadmap plus the milestone answers, and on each milestone alone sorts each failure into one of five reasoning gaps. On 354 NuminaMath problems and six models from 8B to 671B parameters (Qwen3, gpt-oss, Llama 3.3, DeepSeek-V3.1), the largest gap for every model is the composition gap, a stricter form of the compositionality gap. It covers 33-48% of problems, and 24-37% after removing problems that an LLM review flags as grading errors. Accuracy and milestone-help recovery rank the two strongest models differently, and two RLVR runs with similar accuracy gains move problems differently. The roadmap effect replicates on MATH500 and AIME 2024/25, per-problem recovery agrees for 83-87% of problems under an independent second teacher, and the help ladder carries over to code generation. We release the data, roadmaps, prompts, and code at https://github.com/slark-prime/OracleLadder.

Multilingual large language models often produce inconsistent reasoning and answers across semantically equivalent prompts in different languages. Prior work suggests that intermediate representations can be relatively language-agnostic, but generation becomes increasingly language-specific as models commit to discrete output tokens. This is problematic because language-specific lexical choices can cause semantically equivalent reasoning paths to diverge across languages. This motivates searching for a cross-lingual alignment signal that is less tied to any single vocabulary item or script. We propose **SOLAR**, an auxiliary objective for supervised fine-tuning that aligns _soft-token_ representations across languages, using English as a pivot. Soft tokens are probability-weighted mixtures over the vocabulary embeddings, yielding continuous representations that can aggregate information from semantically related tokens across languages. We then align each non-English soft-token summary to its English counterpart in the shared embedding space. Across four multilingual reasoning benchmarks, SOLAR improves accuracy by up to +17.7 points over the base model and +3.8 over standard supervised fine-tuning, with the largest gains on low-resource languages. SOLAR also strengthens final-layer cross-lingual similarity and substantially reduces language-cluster separability, suggesting that aligning soft-token representations helps preserve shared semantic structure during multilingual reasoning.

Video generation has advanced rapidly, producing photorealistic videos from text or image prompts. Meanwhile, film production and social robotics increasingly demand multi-person videos with rich social interactions, including conversations, gestures, and coordinated actions. However, existing models offer no explicit control over interactions, such as who performs which action, when it occurs, and toward whom it is directed. This often results in wrong person performing unintended actions (actor-action mismatch), disordered social dynamics, and wrong action targets. To address these challenges, we present SocialDirector, a training-free interaction controller that enhances the generation model by modulating cross-attention maps. SocialDirector contains two modules: Social Actor Masking and Directional Reweighting. Social Actor Masking constrains each person's visual tokens to attend only to their own textual descriptions via a spatiotemporal mask, avoiding actor-action mismatch and disordered social dynamics. Directional Reweighting amplifies attention to directional words (e.g., "leftward", "right"), leading each action towards its intended target. To evaluate generated social interactions, we annotate existing datasets with interaction descriptions and build a fully automated evaluation pipeline powered by open-source VLMs. Experiments on different video generation models show that SocialDirector significantly improves interaction fidelity and approaches the upper bound set by real videos.

Diffusion-based super-resolution (SR) is a key component in video generation and video restoration, but is slow and expensive, limiting scalability to higher resolutions and longer videos. Our key insight is that many regions in video are inherently low-detail and gain little from refinement, yet current methods process all pixels uniformly. To take advantage of this, we propose SkipSR, a simple framework for accelerating video SR by identifying low-detail regions directly from low-resolution input, then skipping computation on them entirely, only super-resolving the areas that require refinement. This simple yet effective strategy preserves perceptual quality in both standard and one-step diffusion SR models while significantly reducing computation. In standard SR benchmarks, our method achieves up to 60$\%$ faster end-to-end latency than prior models on 720p videos with no perceptible loss in quality.

Composed Image Retrieval (CIR) retrieves a target image given a reference image and a textual edit. Existing training-free methods prompt a frozen multimodal large language model (MLLM) to infer the user's edit intent and rewrite the composed query into a single retrieval artifact, such as a target caption, an edited image, or a modality-fused embedding. These methods can reason about the edit, but they have limited means to execute that reasoning during retrieval. A key remaining bottleneck for training-free CIR appears to be not only intent understanding, but intent execution. To address this challenge, we propose SkillCIR, a training-free framework that tackles the intent execution gap by routing the inferred constraints to specialized retrieval skills. Specifically, the same frozen MLLM that infers the intent also emits a structured plan of active constraints; each constraint activates a typed skill that scores gallery candidates along one evidence channel, and the per-skill scores are composed in score space with explicit signs and re-checked by a local verification step. Across three benchmarks and three CLIP backbones, SkillCIR improves the primary recall/mAP metrics by 2.8 to 7.9 points at the latency of typical training-free CIR pipelines and around 23x faster than the latest training-free state of the art. These gains hold with a fixed skill pool and no model-parameter updates, suggesting that improving the means to execute the inferred edit is a useful complement to stronger reasoning about it.

Face anti-spoofing (FAS) is critical for protecting face recognition systems against presentation attacks in open environments. Beyond high cross-domain generalization, practical FAS deployments must provide interpretable and uncertainty-aware decisions when multimodal evidence is incomplete, degraded, or conflicting. However, current multimodal domain-generalized FAS methods focus primarily on feature fusion or domain alignment, compressing RGB, depth, and infrared cues into opaque logits. This makes it difficult to discern which modality drives the decision, whether modalities reinforce or suppress each other, and whether a given sample lies within the classifier's reliable operating regime. We introduce SEM-FAS, a quantum-inspired yet fully classical single-pass evidence measurement framework that jointly enhances generalization, interpretability, and reliability. SEM-FAS encodes each modality as a normalized complex-valued state whose amplitude captures evidence strength and whose phase encodes cross-modal interference; multiple coherent heads are then mixed into a density-matrix representation. A constrained POVM-like measurement head converts this representation into calibrated class probabilities while analytically decomposing every prediction into per-modality main effects and pairwise synergy/suppression terms. Simultaneously, three built-in uncertainty indicators are obtained without sampling overhead: predictive ambiguity, representation mixedness, and measurement-coverage mismatch. On standard multimodal DG-FAS benchmarks, SEM-FAS reduces the best baseline HTER by 2.18\% and improves AUC by 1.38\% under both complete and missing-modality protocols. Visual analyses further suggest that the learned modality effects, pairwise interactions, and uncertainty scores align with expected FAS behavior, even for hard samples, indicating that single-pass evidence measurement is a promising way to jointly enhance generalization, interpretability, and reliability in multimodal FAS.

The de facto standard way of training neural networks requires storing activations at each layer to memory, in order to compute gradients in backpropagation. As progress has often been made by stacking larger and more layers, memory consumption has now become a major bottleneck for making frontier models available to everyone. In this work, we take advantage of the Residual Network (ResNet) layer formulation that is now used in many state-of-the-art neural network architectures to recompute the activations at each layer on-the-fly during backpropagation instead of storing them in memory. We propose a modification to Reversible Residual Network (RevNet) that overcomes its limitation of requiring to replace the ResNet layer network function with two new network functions, by setting either one to an identity function. This modification is simple yet has a significant advantage because now there is no need to change the original ResNet architecture. Our method can work as a drop-in replacement for layers with residual connections such as in ordinary ResNets or Transformers with practically no loss in performance while offering significant reductions in memory consumption. With this simple modification, we are able to maintain the same computational cost and ease of use as activation checkpointing, which requires no architectural changes, while leveraging a reversible procedure such that we only need to store the activations from the last layer. The method is called Simplified RevNet, and compared to previous work in reversible architectures, we here propose a simpler and more streamlined approach that comes in two variants based on which of $F$ or $G$ in RevNet becomes an identity function. Empirically, we demonstrate performance at practically the same level as the non-reversible counterparts on ImageNet image classification with ResNets and on OpenWebText language modeling with Transformers.

Semi-supervised learning (SSL) has emerged as a practical solution for addressing data scarcity challenges by leveraging unlabeled data. Recently, vision-language models (VLMs), pre-trained on massive image-text pairs, have demonstrated remarkable zero-/few-shot performance that often surpasses SSL approaches due to their exceptional generalization capabilities. This gap motivates us to question: how can we effectively harness the powerful generalization capabilities of VLMs into task-specific models? Knowledge distillation (KD) offers a natural framework for transferring VLM capabilities, but we identify that it suffers from gradient conflicts between supervised and distillation losses. To address this challenge, we propose Dual-Head Optimization (DHO), which introduces dual prediction heads for each distinct signal. We observe that DHO resolves gradient conflicts, enabling improved feature learning compared to single-head KD baselines, with practical benefits of minimal computational overhead and test-time hyperparameter tuning without retraining. Extensive experiments across 15 datasets show that DHO consistently outperforms KD baselines, often outperforming teacher models with smaller student models. DHO also achieves new state-of-the-art performance on both in-distribution ImageNet semi-supervised learning and out-of-distribution generalization across ImageNet variants. We will publicly release our code and model checkpoints to facilitate future research.

LLM/VLM-based digital agents have advanced rapidly thanks to scalable sandboxes for coding, web navigation, and computer use, which provide rich interactive training grounds. In contrast, embodied agents still lack abundant, diverse, and automatically generated 3D environments for interactive learning. Existing embodied simulators rely on manually crafted scenes or procedural templates, while recent LLM-based 3D generation systems mainly produce static scenes rather than deployable environments with verifiable tasks and standard learning interfaces. We introduce SimWorld Studio, an open-source platform built on Unreal Engine 5 for generating evolving embodied learning environments. At its core is SimCoder, a tool/skill-augmented coding agent that writes and executes engine-level code to construct physically grounded 3D worlds from language/image instructions. SimCoder self-evolves by using verifier feedback (e.g., compilation errors, physics checks, VLM critiques) to revise environments and autonomously add reusable tools and skills to its library. Generated worlds are exported as Gym-style environments for embodied agent learning. SimWorld Studio further enables co-evolution between environment generation and embodied learning: agent performance feedback guides SimCoder to generate adaptive curricula near the learner’s capability frontier, so that environments become increasingly challenging as the embodied agent improves. Three case studies on embodied navigation show that self-evolution improves generation reliability, generated environments substantially improve embodied agent performance that generalizes to unseen benchmarks, and co-evolution yields an 18-point success-rate gain over fixed-environment learning and a 40-point gain over an untrained agent.

Vision-Language-Action (VLA) models often bundle architectural changes with different pretraining data, backbone scales, and optimization recipes, obscuring what actually drives progress. We introduce SimVLA, a deliberately minimal VLA—a standard vision-language backbone with a lightweight continuous-action head—as a controlled testbed for this attribution problem. Across LIBERO, CALVIN, WidowX, and Google Robot, one-knob-at-a-time ablations show that training dynamics are the largest measured driver, producing performance swings of $\Delta$$\approx$54-89\%, while task configuration and architecture have smaller effects under matched settings. Despite using a small backbone and no additional large-scale robot-data VLA pretraining before benchmark-specific training, SimVLA reaches 98.6\% average success on LIBERO, remains competitive across the other benchmarks, and transfers to held-out real-robot scenes. These results suggest that simple, carefully controlled VLA baselines remain underexplored and provide a calibration point for future architectural claims within the VLM-encoder plus lightweight-head family.

Apple Silicon is a major consumer platform for local LLM inference, with at least ten serving engines competing for users. Speed-only leaderboards on this platform mislead: the fastest stack often claims memory the rest of the machine needs, or silently produces wrong output. We introduce \ourbench, a benchmark that evaluates ten Apple Silicon inference stacks through three lenses: throughput and latency at concurrency 1, 8, and 16; peak memory under contention with the operating system; and fidelity as weighted F1 on a classification task against an NVIDIA reference run on identical weights. The benchmark covers three model releases and two workload types (short chat and multi-turn agent prompts averaging ${\sim}$4K input tokens). A maintainer agent re-runs the full benchmark weekly; in its first five runs it caught a single upstream library bump that broke three of ten stacks through three distinct failure modes and diagnosed a chat-template error that corrupted one stack's fidelity scores. No audited stack performs well across all three lenses, and the stack rankings shift when memory and fidelity are read alongside speed. The candidate set, stacks without disqualifying gaps under any lens, reduces from ten to four. A separate multi-node comparison shows that tensor parallelism over RDMA scales decode $1.3\times$ on two nodes, while pipeline parallelism over TCP regresses $16$--$19\%$. In every case the gaps are in the engines, not the hardware: Apple Silicon provides sufficient memory, compute, and interconnect; the engines have not yet exploited it. We release the harness, per-run results, and weekly journals.

3D subsurface geological models underpin hydrocarbon exploration, geothermal energy, underground gas storage, and geological carbon storage. A decade of deep generative facies modeling has yet to yield a model that generalizes across geological settings, scales to realistic reservoir extents, and conditions naturally on well data, bottlenecked by the absence of a large-scale, multi-environment, property-rich dataset. We release SiliciclasticReservoirs, the largest and most diverse public 3D dataset of siliciclastic (sandstone-bearing) subsurface reservoirs to date: $10^6$ voxelized reservoirs at $64{\times}64{\times}32$ resolution spanning eight depositional settings (deep-water turbidite lobes, deltas, and six fluvial channel architectures from the Alluvsim family), each annotated with binary and six-class facies labels, per-voxel porosity and permeability, and a natural-language caption describing the reservoir's geological setting and petrophysical attributes. To our knowledge, it is the first siliciclastic dataset to pair 3D geological models with petrophysical properties at this scale, the first to span this breadth of settings in a unified release, and the first to ship text captions, supporting future multimodal and language-conditioned training for text-to-reservoir generation and retrieval. Building on SiliciclasticReservoirs, we train ResFlow, a flow-matching foundation model that, from a single set of weights, generates any environment under continuous parametric control, conditions on arbitrary well configurations via classifier-free guidance, and assembles reservoirs at arbitrary spatial extent through overlapping-tile MultiDiffusion, with no retraining. SiliciclasticReservoirs (CC-BY 4.0), ResFlow (MIT), and the data-generation engine ResMill (MIT) are publicly released.

Modern LLM workflows increasingly move coordinate-indexed objects across checkpoints: steering vectors, sparse autoencoders, top-$k$ neuron sets, attribution lists, and merge alignments. This is only well posed after fixing the model's residual-stream gauge. We show that the native discrete gauge is architecture-dependent: LayerNorm residual charts have permutation gauge $S_d$, up to a global sign flip, while RMSNorm residual charts with generic per-channel gain have signed-permutation gauge $B_d = S_d \ltimes \{\pm 1\}^d$. Thus permutation-only alignment is symmetry-incomplete for RMSNorm models. We introduce sign-marginalized Hungarian matching and prove a sharp population failure mode: with decorrelated source coordinates, raw signed-correlation matching has a structural permutation-accuracy ceiling equal to the fraction of positive signs in the true gauge up to $O(d \cdot 2^{-d})$, whereas sign-marginalized matching removes this obstruction. We then make coordinate-preserving transport, rather than function-level merging, the primary object: composing saved-checkpoint local $B_d$ gauges along same-base fine-tuning trajectories recovers 91.1% of cross-run coordinates at 1500 steps versus 60.3% for endpoint matching, and the gain is not explained by merely routing through the base. The recovered gauge transfers tools that permutation-only alignment breaks: TinyLlama SAE reconstruction has NMSE $0.004$ under $B_d$ recovery versus $1.08$ under $S_d$; Qwen sentiment steering preserves 95.8% of its effect versus 17.2%; refusal steering reverses sign under $S_d$. Coordinate-preserving merge tests show the same mechanism. The same covariance governs stateful training: signed transport of AdamW state preserves the resumed trajectory, while permutation-only state transport starts from a functionally identical checkpoint but follows a different trajectory. Finally, we give gauge-sweep audits for index-level interpretability claims: coordinate names are reproducible only relative to an explicit gauge.