World-action models jointly predict future observations and actions, yet reconstruction-oriented visual representations do not explicitly emphasize the spatial relationships and object-state changes needed for manipulation. We introduce Phoenix, which learns representations of scene structure for world-action modeling. Specifically, Phoenix Encoder enriches DINOv3's general visual semantics through dual-teacher distillation of geometric and spatiotemporal semantic knowledge. Next, CS-VAE causally aggregates and compresses Phoenix Encoder features into generative latents while preserving scene information and spatial layout. Finally, Phoenix jointly predicts these latents, complementary appearance latents, and robot actions, linking scene evolution with action generation. To study pretraining data efficiency, we curate approximately 39,000 hours of egocentric and robot manipulation data for representation learning and use the robot demonstration subset for WAM pretraining. Following benchmark-specific post-training, Phoenix achieves state-of-the-art performance across simulation and real-robot benchmarks spanning single-arm, bimanual, and dexterous manipulation. It demonstrates strong robustness to perturbations, achieving success rates of 94.8% on LIBERO-Plus and 72.4% on RoboTwin 2.0 C2R, alongside strong cross-embodiment transfer and real-world manipulation performance on Bench2Dex and RoboDojo-RealWorld, respectively. Pretraining-scale analysis shows larger performance gains from data scaling than a video-VAE-only baseline, indicating stronger scaling benefits and more effective data utilization. Moreover, Phoenix maintains high success rates when adapted with limited downstream demonstrations, demonstrating strong few-shot adaptation. Our code, pretraining data, and models are available on the project page to help the community reproduce our results.
Controllable real-world video generation models can condition on a rendered graybox proxy video, which fixes layout, motion and camera. Coding agents can build such videos by generating a 4D scene in Blender, yet no benchmark checks both these rendered proxy videos and the complex, multi-shot scenes behind them. We introduce VibePrevisBench, a benchmark for 4D scene generation: an agent turns a brief into an executable 4D scene and renders its proxy video. Its 242 tasks span four input kinds, including scripts, photos, continuations, and reference videos. Each delivery is scored along two evaluation dimensions. Rule-based Scene Check re-runs the code and compares the exported 4D scene with hidden ground truth facts in nine dimensions. VLM-based Screen Check asks a VLM judge whether facts that pass the rules are recognizable in the proxy video. We analyze the evaluation results and trajectories of 11 frontier coding agents and identify three key observations. First, multimodal understanding separates the agents, especially in understanding space and camera motion. Second, all agents seem to already master basic camera work and staging, and fail at challenging requests. Third, agents inspect their own renders but fail to correct their errors. Agents thus need better visual understanding of space and camera, explicit checks for challenging requests, and item-by-item verification of renders against the brief.
Jailbreak Detection on Large Language Models via Contrastive Prototype Structures
Yan Qiu, Yuanyuan Li
AI 速读 基于摘要
JDPI利用对比学习表示和原型距离进行越狱检测,无需额外下游分类器,在多个基准上表现优于基线。
cs.CL
查看原文摘要
Jailbreak attempts and harmful prompts remain a challenge for the safe deployment of large language models. We study whether contrastively learned representations can support safety detection without training an additional downstream classifier. We propose Jailbreak Detection with Prototype Inference (JDPI), which uses binary safety labels to construct harmful and benign prototypes and makes predictions from relative prototype distances. JDPI uses only binary safety labels for encoder adaptation, prototype construction, and validation. In our setup, it uses fewer labeled examples than the classifier baseline (Method A) by avoiding the separate split used to train the downstream classifier. Across four benchmarks covering different input formats, contrastive fine-tuning improves JDPI's detection accuracy by 9.1-18.0 percentage points. Using the same encoder, JDPI achieves slightly higher ROC-AUC on one benchmark and stays within 0.017 of the baseline across all four. With dataset-specific adaptation, JDPI exceeds the released Llama Guard 3-8B baseline by 0.173-0.310 in ROC-AUC across the four benchmarks. In multi-turn replay on TomGibbs, JDPI uses accumulated user history as the conversation unfolds, achieving higher conversation-level accuracy than when it uses the current turn alone. Across 12 directed transfers, JDPI outperforms the classifier by an average of 1.59 percentage points in accuracy, demonstrating stronger cross-dataset robustness.
Grana: Compositional Reward Modeling for Long-Horizon Robot Manipulation
Zi-ang Cao, Kaiyuan Zheng, Yinzhen Xu, Haonan Duan, Jiasheng Gu, Fengyuan Hu, Nan Zhu, Kun Lei 等 17 位作者
AI 速读 基于摘要
Grana 是一种视觉-语言机器人奖励模型,通过开放词汇子任务描述识别未完成子任务并估计其连续进度,再组合子任务进度与已完成子任务信用来估计长时程任务的整体进度。
cs.RO
查看原文摘要
Robot reward modeling faces two challenges: generalizing across tasks and accurately tracking progress in long-horizon manipulation. General-purpose reward models demonstrate broad task transfer, but transfer from short-task pretraining to long-horizon tasks remains underexplored. Long-horizon progress models often rely on task-specific subtask classifiers or fixed action vocabularies, limiting the reuse of learned progress knowledge when adapting to new tasks. To address these challenges, we introduce Grana, a vision-language robot reward model that learns to identify unfinished subtasks and estimate their continuous progress using open-vocabulary subtask descriptions. We train a language-conditioned model to estimate progress in short-horizon tasks and individual subtasks of long-horizon tasks. For long-horizon tasks, we combine the selected subtask's predicted progress with credit for preceding completed subtasks to estimate whole-task progress. General progress prediction requires training the model across diverse tasks with progress supervision that accounts for imperfect execution. Elapsed-time proxies can misrepresent progress when relative subtask durations vary, or failures reverse local progress. We therefore introduce an annotation protocol that converts sparse subtask boundaries and action-quality annotations into dense progress targets, without requiring annotators to assign numerical progress scores. Pretraining with this supervision on 200 diverse, mostly short-horizon manipulation tasks enables zero-shot reward prediction on unseen tasks and improves progress prediction after adaptation to multi-minute, long-horizon tasks. Experiments demonstrate (i) generalizable reward prediction, including adaptation to five long-horizon tasks, (ii) benefits from pretraining and composing subtask progress estimates, and (iii) improved downstream policy performance.
We construct a four-buyer online combinatorial auction with independent subadditive valuations and a fixed, known arrival order in which every online rule earns strictly less than $1/2$ of the expected offline optimum. A common partition, revealed at the third arrival, gives the first three buyers one unit of welfare per item. If the rule uses three distinct partitions, either an item allocated early contributes zero or the fourth buyer's expected value for the remaining bundle is one less than its size. We compute the optimal online and offline welfare exactly and show that a $1/2$ guarantee always holds with at most three buyers.
In this paper, we propose a Unified Law for Visual Tasks, which could be further divided into Four Laws of Computer Vision, with the goal of enabling a single framework or method to uniformly train on and perform nearly all visual tasks. Our starting point is an inspiration and hypothesis: if the world information is encoded into image, which we call World Snapshots, then image processing is world processing, and image generation is World Snapshot generation. Furthermore, if all visual and visualization tasks can be reformulated as generation tasks, then a unified method, World Snapshot Models (WSMs), can perform the vast majority of visual tasks, much as large language models have largely unified natural language processing. This paper validates these laws across more than 1,000 trained and evaluated visual tasks (defined by distinct input-output mappings) and presents a unified training framework for future visual-task research and applications.
Agentic image restoration composes operations to handle mixed degradations, yet its toolbox often requires storing multiple restoration networks. We propose PrismRestore, which constructs a compact bank of callable restoration actions from a single frozen all-in-one model. Our approach builds on task-dependent yet overlapping channel responses in the shared model. An offline task-prior search uses activation and gradient statistics to group channels with similar response profiles, then searches task-specific modulation coefficients over non-representative channels while leaving representative channels unmodulated. The resulting prior bank enables explicit action selection and composition without additional restoration backbones or task-specific adapter modules. A state-conditioned policy learns to compose these actions using Group Relative Policy Optimization. Its objective separately normalizes fidelity and perceptual feedback within groups of trajectories sampled for the same input, and combines terminal evaluation with return-to-go derived from stepwise quality changes. Experiments on synthetic composite-degradation and real-world multi-task benchmarks show improvements over the evaluated all-in-one baselines and restoration quality competitive with multi-model agents. PrismRestore achieves these results with a single restoration backbone and faster inference than the evaluated agentic baselines.
Computer-Aided Design (CAD) underpins modern engineering, yet converting existing shapes into editable models still demands substantial expert effort. Most AI systems emit the entire CAD program in a single pass, never inspecting the intermediate geometry. In contrast, human engineers build a part feature by feature, checking after each operation what remains to be modeled. We introduce CADENA (Spanish for "chain"), a model that reconstructs a 3D mesh as a parametric CAD program, growing its sequence of operations one at a time and comparing the target with the currently predicted geometry at every step. We also address the lack of benchmarks for evaluating reverse-engineering methods on mechanical parts, introducing CADENA-Bench, a benchmark that measures performance across categories of mechanical parts. CADENA outperforms prior methods on CADENA-Bench and on the DeepCAD, Fusion 360, and MCB datasets. Code is available at https://github.com/zhemdi/cadena, model weights at https://huggingface.co/kulibinai/cadena, and CADENA-Bench at https://huggingface.co/datasets/kulibinai/cadena-bench.
The final processing step in compression systems is generally an entropy coding stage, which converts symbols into a bit stream based on their probability distribution. A distinct class of entropy coding methods operates not by mapping input symbols to discrete codewords but by operating on intervals or ranges. This approach enables a more accurate approximation of the source entropy, particularly for sources with highly skewed or varying symbol distributions. Representative techniques in this category include traditional arithmetic coding, range coding, and methods based on asymmetric numeral systems (ANS). Due to their near-optimal compression efficiency, these interval-based entropy coders are integral to modern data compression systems. The complexity of these methods depends mainly on three processing steps: the core routines of encoding and decoding, the interval-based determination of the correct symbol at decoder, and the efforts of keeping updated with respect to the varying symbol distribution. The interval-based symbol determination at decoder typically demands for a searching procedure. Previous literature showed that the search can be replaced by a table-based approach with only O(1)-complexity but having the side-effect that the adaptation of the multi-symbol statistic becomes infeasible because of the high time-consumption of adapting the table. We compare different approaches to fast adaptive coding of multi-symbol alphabets and propose a new adaptation process using a special ring-buffer technique enabling the adaptive table-based decoding procedure in combination with the replacement of a division by a bit-shift operation at encoder and decoder core routines. The proposed technique is faster than alternative approaches for alphabets from about 12 to 64 different symbols when comparing the overall encoder+decoder time in a practical implementation.
Equivariant Modeling with Polynomial Features via Minimal Homogeneous Generators
Michael Liaofan Liu, Joe Kileel
AI 速读 基于摘要
提出一个基于极小齐次生成元构造等变多项式特征的通用框架,并配套自动计算与拟合的Python包。
cs.LG · stat.ML
查看原文摘要
We present a general framework for modeling and learning equivariant functions between finite dimensional semisimple real group representations, unifying constructions developed across many application areas. Our model is a linear combination of equivariant polynomial features whose coefficients are arbitrary learnable functions of invariant polynomial features. We take as features minimal homogeneous generators for the invariant algebra and the covariant module over the invariant algebra, up to prescribed degrees. We provide a Python package that automatically computes these features and fits the resulting models to training data. The algorithm for feature computation requires explicit isotypic decompositions of the input and output representations, along with explicit Clebsch-Gordan decompositions of the tensor product of irreducible representations, and achieves speedups over alternative methods in computational algebra. Numerical experiments in scientific modeling demonstrate the full learning pipeline and yield accurate predictors with few trainable parameters, using the model both for end-to-end prediction and within layers of larger architectures.
MAVE: Reconstructing Editable 3D Blockouts from Film Shots
Yifan Chang, Jian Yang, Jiangnan Shao, Xiaofeng Wang, Zhaoxin Fan, Renjing Xu, Ming Lu, Xingang Wang 等 9 位作者
AI 速读 基于摘要
作者提出MAVE框架与MAVEBench基准,从电影镜头自动重建可编辑的Blender 3D blockout场景与帧对齐视频。
cs.CV · cs.AI
查看原文摘要
Animated 3D blockouts support previsualization and provide structural conditions for video generation, but manually reconstructing them from reference videos is costly. This limits both their use by creators and the production of paired training data. We introduce MAVEBench, a benchmark of 50 film shots and 416 keyframes for evaluating camera motion, spatial layout, and character blocking. We propose MAVE, a video-to-blockout framework combining Metric observation, Agentic modeling, and evidence-based VErification. It first recovers cameras, human motion, and geometric evidence in a shared metric coordinate system. An agent then constructs an editable environment under constraints from these observations, while programmatic checks report inconsistencies to guide refinement. The outputs are editable Blender scenes and frame-aligned blockout videos. On MAVEBench, MAVE improves overall scores by 27.5-41.1 points across three agents over their corresponding vanilla baselines. With DeepSeek-1017, it achieves a score of 73.2 at an average API cost of only USD 0.054 per shot, outperforming the strongest vanilla baseline while reducing total reconstruction time by 53%. We also construct MAVEData-3K, a dataset of 3,000 blockout-video pairs for model training, and will release it publicly. Creators can change viewpoints and character paths, then render the edited scenes to condition video generation. Code is available at https://github.com/yfChang-cv/MAVE.
CAD practitioners often work with proprietary parts that they are unwilling to share with external services. For individuals and small companies, bringing this work in-house also requires computational resources they can afford. We introduce CADMAGEDDON, an open-source agent that reconstructs editable CAD from meshes and can run locally on single-GPU systems and current-generation laptops. The method decomposes this complex task into a chain of simple visual decisions about what to construct and what to correct. Geometric tools support precise construction, allowing the agent to focus on visual reasoning. Using the mid-sized Qwen3.8-27B model, CADMAGEDDON outperforms prior approaches, including systems based on frontier models, in reconstruction quality on all the main benchmarks. CADMAGEDDON integrates with FreeCAD, bringing its reconstructions into practitioners' existing CAD workflows. Code: https://github.com/kulibinai/cadmageddon
User simulators offer a scalable way to evaluate interactive agents, but how to assess their quality remains underexplored. In particular, it is unclear whether existing user simulator evaluation frameworks reliably detect meaningful failures in simulated user behavior. This uncertainty makes it difficult to interpret simulator evaluation. To understand what a high benchmark score tells us about simulator quality, we (meta-)evaluate the metrics themselves. To instantiate this meta-evaluation, we deliberately perturb simulated replays to induce failures such as altered goals, omitted constraints, premature stopping, and acceptance of broken work, then test whether the metrics penalize these failures. Evaluation methods from three prior studies detect only 5-6 of the 10 perturbations; their combined scores even favor populations instructed to accept broken work over unmodified controls. Motivated by these gaps, we propose a new evaluation suite that combines utility of user simulators for evaluating agents and two measures of fidelity to real users: population-level fidelity assesses collective interaction patterns, while trace-level fidelity assesses the preservation of individual users' goals, constraints, and persona. Under the same detection criterion, our evaluation suite detects all 10 perturbations. Evaluating six user simulators on 351 real multi-turn sessions in which 348 distinct users across 23 languages and agents work together to create and refine artifacts such as full-stack web apps, games, design assets, and automated workflows, we find that population-level measures can miss when a simulated user pursues the wrong goal, ignores a constraint, or behaves out of character, highlighting the need to assess fidelity at comprehensive granularities.
General-purpose object detectors lose accuracy on UAV footage, where targets span only a handful of pixels and onboard compute is limited. Prior work composes independently-validated architectural techniques into one detector, assuming gains reported in isolation transfer once combined. We stress-test that assumption directly. LAF-YOLOv10 integrates four techniques into YOLOv10n: a Partial Convolution C2f (PC-C2f) backbone block, an Attention-Guided Feature Pyramid Network (AG-FPN), a P2 detection head replacing the large-object P5 head, and Wise-IoU v3 regression, asking whether their combined effect matches what each contributes alone. We train LAF-YOLOv10 three times (seeds 42, 123, 256) on VisDrone-DET2019, benchmark against unmodified YOLOv10n, and use TIDE error decomposition, per-category breakdown, per-component ablation, attention/loss comparisons, zero-shot transfer to UAVDT, and held-out/test-dev evaluation to localize where the combination succeeds or fails. Composability does not hold here. LAF-YOLOv10 reaches 24.0+/-0.4% mAP@0.5 at 2.14M parameters, 7.8 points below YOLOv10n (31.8%), a deficit that transfers to UAVDT (-10.0 points) and is confirmed by held-out and test-dev evaluation (23.5%, 22.5%). Background false positives, localization error, and duplicate detections move in the direction AG-FPN and Wise-IoU were designed to push. Ablation traces the deficit to a specific source: the P2/-P5 head swap costs 2.5 points independently plus a 2.5-point interaction penalty when layered onto a backbone weakened by PC-C2f, whose own 2.0-point loss is consistent with a partial pretrained-weight transplant (73/150 backbone tensors transfer). The failure is attributable to a specific interaction, not the four components individually. Composability must be verified directly, not assumed. Code/checkpoints: https://github.com/Mudassiruddin7/Small-Object-Detection-in-UAV-Imagery.
Turning a CAD model into a machining program requires both process knowledge and geometric precision. Language models can generate such workflows by writing geometry and toolpath code for each part, but this makes numerical execution part of the model's task. We introduce MillBill, an agent that separates machining decisions from their geometric execution and verification. Given a boundary-representation part, stock, and material, a locally hosted, mid-sized open model proposes operations and cutting parameters for indexed three-axis machining. A geometric harness generates toolpaths, simulates material removal, and returns feedback for plan refinement. Setup planning accounts for what the available cutters can physically reach, rather than visibility alone. We evaluate the resulting programs by the fraction of the target surface area within the specified tolerance after simulated machining. With Qwen3.8-27B, MillBill outperforms frontier-model approaches that construct their own geometry and toolpath workflows on the evaluated machining tasks. Code: https://github.com/kulibinai/millbill
Large language models have been widely used to generate plausible research ideas, yet it remains unclear what makes automated ideation effective and how generated novel proposals arise from prior scientific work. We introduce the Creative AI Researcher (CAR), an agentic framework combining literature grounding, iterative development and explicit conceptual operators. CAR constructs a semantic knowledge graph from retrieved literature and equips an ideation agent with graph-retrieval tools and four operators: bisociation, analogy, exploration and elaboration. Each operator call is checked against structural acceptance conditions and recorded, producing a trace of intermediate transformations and declared literature sources. We evaluate CAR with two language-model backbones across 40 LiveIdeaBench research topics and two execution benchmarks. CAR achieves significantly higher methodological novelty (MN) than single-shot prompting and The AI Scientist. Its methodological novelty does not differ significantly from the Research Ideation Agent, while its feasibility scores are higher on both backbones. With a shared experiment executor, CAR achieves the highest pairwise win rate among the evaluated AI methods on FML-bench-Lite and BioXArena. In a ten-condition ablation study, MN point estimates favor CAR over one-step proposal on both backbones, but no MN contrast survives correction for multiple comparisons. The results do not establish a consistent quality benefit from explicit operators beyond free-form iteration. CAR makes the sequence of operations and declared literature sources inspectable, separating judged idea quality from procedural traceability.
We present a large-scale experimental study of quantum-computing-based molecular simulation carried out on IQM's Sirius 24-qubit superconducting processor, utilizing up to 16 operational qubits. The work employs Sample-based Quantum Diagonalization (SQD) together with the Local Unitary Cluster Jastrow (LUCJ) ansatz to estimate ground-state energies for a set of benchmark molecules, including H$_2$, LiH, BeH$_2$, H$_2$O, and NH$_3$. In addition, we introduce a Linear-CNOT variant of the Unitary Coupled-Cluster Singles and Doubles (LCNot-UCCSD) ansatz within the SQD workflow, trading higher circuit depth for reduced classical preprocessing. A comparison between these ansätze is provided, clarifying their respective strengths, limitations, and suitability for near-term quantum hardware. We further explore potential energy landscapes through 1D scans for H$_2$ and HeH$^+$ using both STO-3G and 6-31G basis sets, and for LiH and BeH$_2$ in STO-3G. Extending beyond this, we demonstrate the experimental construction of a full 2D potential energy surface for the water molecule on quantum hardware, mapped over a 32 $\times$ 32 grid in bond length and bond angle. To move beyond small benchmark systems, we combine SQD(LUCJ) with Density Matrix Embedding Theory (DMET) to compute active-space energies for a set of ligand-like molecules, as well as the pharmacologically relevant amantadine system. Across all studies, the majority of quantum-computed energies agree with reference FCI results, as well as with DMET-CASCI energies for embedded systems, to within the chemical accuracy threshold for the chosen basis sets. These results demonstrate the reliability of sample-based diagonalization approaches and underscore the potential of hybrid embedding strategies for extending quantum simulations to increasingly complex molecular systems, while also highlighting their practicality on quantum hardware.
Teleoperated Haircutting with Delay Compensation and Passivity-Based Control for Unilateral Internet Teleoperation under Real Communication Uncertainties
Viridiana Luna-Zempoalteca, Zhendai Huang, Rosalba Galvan-Guerra, Antonio Rosales, Xiaoli Liu, Shuai Li
Robotic haircutting requires accurate motion execution and reliable communication to safely reproduce the operator's commands. Internet-based teleoperation is a key enabling technology to remotely perform precise manipulation tasks such as haircutting. However, communication uncertainties, including variable transmission delays and packet jitter, degrade synchronization between the master and slave systems and affect tracking performance. This paper presents a unilateral Internet teleoperation framework for robotic haircutting that takes into account experimentally measured communication uncertainties. Three delay-compensation techniques, namely Zero-Order Hold (ZOH), First-Order Hold (FOH), and Low-Pass Filtering (LPF), are evaluated together with a classical proportional-derivative (PD) controller and a passivity-based synchronization controller. Unlike conventional studies based on synthetic delay models, the proposed framework is validated using communication data acquired from a real cloud-based teleoperation platform, where both packet delay and effective delay are characterized. The results demonstrate that combining delay-compensation techniques with passivity-based synchronization improves the robustness of unilateral Internet teleoperation under realistic communication uncertainties, providing a reliable control framework for robotic haircutting applications.
Fixed shipment costs encourage large orders, while inventory and customer backlog costs favor frequent replenishment. In a serial system, the stages must also coordinate their lot sizes and material availability. We study the classical echelon stock $(R,nQ)$ policy: the downstream stage requests a fixed lot, the upstream stage orders an integer multiple of that lot, and transfers contain complete lots. For unit Poisson demand, full backlogging, and at both stages, deterministic lead times, and setup and holding costs, we prove that the optimized cost in this class is at most $5/3\approx1.667$ of the optimal admissible system cost, uniformly over all nonnegative cost rates, demand rates, and lead times. The guarantee holds even when the selected policy pays a setup charge for every lot while the optimal benchmark pays only once per aggregate shipment. To obtain the guarantee, we use a common inventory calculation that allows independently chosen replenishment frequencies while respecting upstream material availability. We compare its optimized value, an upper bound on the best integer-ratio policy cost, with a novel lower bound on the optimal system cost that combines allocated single-stage costs with the minimum inventory cost when setup costs are zero. We also show that no uniform factor below $(1+\sqrt{3})/2\approx1.366$ is possible for the classical class. Existing numerical experiments place best-found classical policies within $22.7\%$ of an evaluated lower bound across three broad grids. The results quantify both the reliability and the limits of a simple operating rule and explain how shipment consolidation, replenishment timing, and flexibility affect its value.
WorldRuler 用可执行 3D 规则图把规则执行、视频生成与感知反馈闭环连接,以支持用户自定义规则驱动的交互式世界生成。
cs.AI
查看原文摘要
Rules govern how player actions change interactive worlds. Different rules create distinct gameplay experiences within similar scenes and action spaces. For customizable gameplay, generative world models must follow user-defined rules beyond real-world behaviors. However, realizing user-defined rules presents two challenges: Rule-State Persistence and Rule-to-Visual Realization. The former requires preserving rule definitions, dependencies, and world states across saving and loading, so that restored worlds follow the established rules. The latter requires accurately realizing rule-induced changes on designated objects and attributes, particularly when affected objects are unseen. In this paper, we propose WorldRuler, which organizes rule definitions, dependencies, and object states in an executable 3D rule graph and connects rule execution, video generation, and perceptual feedback in a closed loop. The graph supports joint saving and restoration of rules and world states, while dynamic object binding grounds rule-specified entities in visual instances and uses stored states to constrain generation from their first appearance. We further introduce object attribute state replacement and counterfactual reconstruction training, enabling accurate attribute control and compatible appearance and motion changes while limiting interference with unrelated content. We develop an automated data collection and annotation pipeline using GTA-V and Habitat to construct RuleVerse, a dataset of rule-driven interactions with detailed state and event annotations, including cases where rules take effect before affected objects become visible. Extensive experiments demonstrate the effectiveness of WorldRuler in preserving rule-state consistency across saving and loading and accurately realizing rule outcomes on designated objects and attributes.
Visual agents that read dense documents must ration a finite observation budget across candidate regions, yet answer accuracy conflates where an agent looks, how it adapts to the reads that remain, and how it an?swers once observation stops. We introduce budget-paired counterfactual audits, an evaluation protocol that holds public evidence and region bindings fixed while perturbing only the remaining budget or de?clared task parameters, and credits every acquired history with exact continuation values. The resulting accounting decomposes lost expected accuracy into stepwise acquisition loss and terminal decision loss, separating planning from fixed-role heuristics. Across forty paired routing anchors, five open-weight
作者提出 NereidVLA,先以约两百万海洋样本进行海洋空间预训练(Sea2Know),再用仿真水下动力学下的专家轨迹适配,使视觉-语言-动作模型能根据语言指令预测离散运动 token 并解码为局部 3D 轨迹,用于闭环水下导航。
cs.CV · cs.RO
查看原文摘要
Underwater navigation requires translating visual spatial cues into motion, yet marine vision-language datasets rarely include paired robot trajectories. We present NereidVLA, a vision-language-action model that transfers marine spatial knowledge to language-conditioned 3D navigation. We first train Sea2Know on approximately two million marine examples to learn target grounding, spatial relations, and scene structure without paired robot-motion supervision. We then adapt it using expert trajectories executed under simulated underwater vehicle dynamics. Given RGB histories and a language instruction, NereidVLA predicts discrete motion tokens that decode into local 3D trajectories for closed-loop tracking. Evaluation combines an independent six-dimensional underwater navigation understanding benchmark with offline prediction and closed-loop navigation. Under matched navigation training and execution settings, marine spatial pretraining increases simulation success from 72.4% to 84.4%. Without fine-tuning on the real test environment, NereidVLA achieves 78.75% success in laboratory-pool robot trials, compared with 47.50% for InternVLA-N1. These results show that marine spatial supervision can support downstream motion learning and physical underwater navigation.
Human communication involves highly coordinated multimodal signals such as speech and co-speech gestures. Existing approaches for generating both modalities commonly follow a cascaded pipeline, where speech is generated first and subsequently used to drive gesture generation; speech generation is therefore independent of gestures. We introduce Gaspard, a unified speech--gesture generation model based on continuous autoregressive generation with next-step flow-matching predictions, and trained on Seamless Interaction, a large-scale multimodal dataset. The heterogeneous temporal properties of speech and gesture motivate modality-specific prediction heads and decoupled classifier-free guidance over the two generation streams. This guidance rebalances the influence of speech and gesture at inference time, improving expressivity while mitigating dynamics collapse in continuous autoregressive gesture generation. Objective and perceptual evaluations show that Gaspard is a strong gesture generation model against unimodal baselines and that it produces highly synchronous speech and gestures compared to a matched cascaded system.
Human vision and video stereo networks both reuse disparity evidence across time, yet a refresh of the input can invalidate a match formed milliseconds earlier. What tells a system whether an old match still counts? We answer this from limited human data in three steps. First, nine temporal mechanisms share a frozen binocular-energy front end and an equal-capacity readout, so held-out behavior isolates the temporal selection rule: on 240 conditions from 10 observers, down-weighting aged evidence that has lost binocular support lowers subject RMSE from 12.6% to 10.3% and reproduces a period-dependent reversed-depth illusion, across all 59 readings of the unpublished protocol and 8 front-end perturbations. Second, four propositions establish what such data can identify: gating is identified at the level of the refresh schedule, and candidate binding is invisible to standard stimuli but exposed by a single-eye geometry flip that separates the readouts by up to 7.1%. Third, the identified rule becomes a two-field stereo memory that separates sample validity from hypothesis support. It removes a 50 ms recovery lag after geometry flips, and its regime analysis explains why plain integration stays preferable on MPI-Sintel, where 71.1% of pixels change appearance but not disparity.
World models are increasingly used in autonomous driving to predict how a scene evolves under the ego vehicle's actions, supporting counterfactual reasoning and model-based planning. However, existing evaluations focus on prediction fidelity and planning performance, and rarely verify that the predictions depend on the conditioning action. In this paper, we study action dependence in action-conditioned semantic forecasting, which predicts future semantic segmentation from past segmentations and the ego action. We compare a forecaster with feature-wise and input-level action conditioning against an identical model without action input, and measure action dependence with test-time action interventions scored by standard segmentation metrics. The evaluation is validated in a synthetic environment with known action-dependent dynamics and applied to Cityscapes and BDD100K, both within each dataset and across the two vehicle platforms. Experiments show that the executed action improves mid-term mIoU by at most 2.5 points on Cityscapes and 0.6 points on BDD100K, whereas a wrong action reduces it by up to 23.2 and 7.0 points, below the model without action input. On real data, feature-wise conditioning depends 1.6 times more on the action than input-level conditioning at similar accuracy, whereas both depend on it equally in the synthetic environment, where the history does not determine the action. These results show that an accuracy gain indicates that a driving world model uses its action, but not how strongly its forecasts depend on it. Code is available at https://github.com/binayksingh/does-action-matter.
Deep brain stimulation (DBS) effectively treats motor symptoms of Parkinson's disease (PD), but it is unclear whether it changes fine hand movements in the same way for every patient. We analyzed 35 sessions of 19 patients with PD from a multi-center cohort performing the CopyDraw task, a drawing task designed to overcome the subjectivity and limited granularity of clinical assessments. Besides standard kinematics (mean velocity, acceleration, jerk, and drawing accuracy), we derived higher distributional moments and non-linear complexity measures, for 38 features in total. DBS effects were quantified with linear mixed models at the group level and with residualized regression and a random-effects meta-analysis per session. The DBS state was decoded per session with regularized linear discriminant analysis to compare feature sets as behavioral labels for deriving physiomarkers for adaptive DBS (aDBS). DBS modulated all 38 features at the group level, most strongly velocity-related features. However, for every feature, the between-subject variability of the DBS effect exceeded the group-level effect. Only one feature, the detrended fluctuation analysis of the distance to the template, showed a significant mean effect across sessions, and it is not a standard kinematic feature. In six sessions without any significant modulation of standard features, complexity measures were modulated instead. For decoding, standard features alone were the best choice in ten of 35 sessions, but combining all 38 features yielded the highest average area under the ROC curve (0.700 vs. 0.677), with the largest gains in sessions that were hard to decode from standard features. DBS effects on hand-motor behavior are individual. Deriving neural markers for aDBS from behavior therefore requires session-specific analyses with individualized features beyond standard kinematics and group-level effects.
Hybrid attention models combine full and linear attention to achieve a better quality--efficiency trade-off. Converting pretrained Transformers into such models requires identifying which layers should retain full attention under a prescribed budget. We identify two off-policy mismatches in FlashMorph's layer selection: (i) continuous branch interpolation and teacher-forced hidden states differ from discrete hybrid computation; and (ii) frozen model weights fail to account for parameter adaptation during subsequent recovery. We introduce FlashMorph2, an on-policy layer-selection method that (I) evaluates sampled discrete hybrid architectures end to end using their own hidden states, and (II) jointly optimizes layer-selection scores and model weights under the recovery objective to reflect post-adaptation performance. Experiments across Qwen3 model scales and linear-attention variants show that FlashMorph2 consistently outperforms existing layer-selection methods, improving average scores over FlashMorph by $+2.42$ points on RULER, $+1.08$ on LongBench, and $+0.76$ on recall-intensive tasks, while completing layer selection in only $0.72$--$3.57$ GPU hours across 0.6B--8B models, demonstrating high efficiency and scalability.
Global planners and path-tracking controllers are often evaluated independently, although their interaction determines how a vehicle follows a planned path. Whether simulation accurately predicts the performance of specific planner-controller combinations on Ackermann-steered hardware remains uncertain. This study evaluated three geometric controllers (Regulated Pure Pursuit, Stanley, and Vector Pursuit) paired with two global planners (Hybrid A* and NavFn) on a 1/10-scale Quanser QCar and in a matched Gazebo simulation. Both environments used shared custom planning and control algorithms, matched route geometry and goals, ground-truth pose references, and a common evaluation pipeline. Five runs per configuration and environment yielded 60 benchmark runs. Controller rankings for tracking accuracy and command smoothness transferred consistently between environments, with Spearman correlations of 0.77-0.94. However, planner-dependent effects showed substantially weaker agreement. In simulation, planner selection changed the root-mean-square (RMS) angular acceleration of the commanded velocities by no more than 0.030 rad/s^2, whereas Hybrid A* increased it by 0.141-0.843 rad/s^2 on hardware, depending on the controller. These differences coincided with planner-dependent achieved speeds that the simulation's idealised longitudinal response did not reproduce. Mean controller computation times remained below 0.5 ms on the Jetson TX2. The findings demonstrate that agreement in controller rankings does not guarantee accurate prediction of planner-dependent performance for specific controller pairings on hardware, highlighting the importance of physical validation.
Quantum Machine Learning (QML) has attracted increasing interest as a possible route to construct feature representations that are difficult to reproduce with classical models. However, near-term QML architectures based on trainable parametrized quantum circuits remain affected by limited hardware resources, noise, and severe optimization difficulties such as barren plateaus. In this work we propose Soft-MoQE, a Soft Mixture of Quantum Experts architecture that replaces variational quantum training with the input-dependent selection and combination of fixed quantum feature maps. The model is composed of a heterogeneous pool of non-trainable quantum circuits, each defining a distinct embedding, and a classical gating network that learns how to weight these embeddings for each data point. We benchmark Soft-MoQE on three PCA-reduced tasks: MNIST parity, MNIST 10-class, and FashionMNIST. Soft-MoQE outperforms the single best expert, the uniform average of all experts and a linear baseline, achieving task-dependent performance comparable to that of a trainable quantum neural network, without optimizing any quantum parameter and with a quantum training cost that is lower by orders of magnitude. The analysis of the class- conditioned expert usage shows that the learned routing combines a few experts for each input, with a selection that depends on the class. These results support the learned routing among fixed quantum circuits as a trainability-aware design principle for hybrid QML.
Neural networks with randomly initialized, fixed weights and trainable biases are universal approximators, yet the dynamics of bias-only training remain largely unexplored. We study these dynamics in one-hidden-layer ReLU networks through activation codes, the binary patterns of active neurons that partition the input space into convex polyhedral regions. We show that the gradients with respect to the biases depend on each input only through its activation code, and hence the training dynamics can be characterized with these codes. In particular, as long as the activation codes remain fixed, training reduces to a linear least-squares problem with the codes as features. Training can then be understood as a sequence of least-squares problems, with transitions occurring when neurons change their activation states. We show that the fraction of switching neurons vanishes as the network widens, and the training dynamics approach those of linear least squares on the initial activation codes in the infinite-width limit. Our experiments support these theoretical findings, showing that while activation switches can noticeably affect training in narrower networks, their relative frequency and the associated changes in the dynamics diminish as the width increases.