Computer Science & Mathematics

As the applications of AI continue to expand across various multimodal and multilingual domains, ensuring safe and responsible deployment of these systems has become an important concern. A wide-range of interventions have been proposed to make AI responses safer and more reliable. They often incur substantial computational or annotation costs, or cannot be extended to multilingual settings. In this work, we propose Selective Detection And Targeted Editing (SEDATE), a training-free approach based on detect-edit-verify strategy. SEDATE first detects the responsible neurons, edits them as and when necessary and finally verifies the model responses; all without any additional model training. We evaluate SEDATE across 7 languages and 4 model families. We further evaluate its effectiveness across 4 out-of-distribution languages and 4 additional datasets. We show that SEDATE achieves substantially superior safety scores across all languages and models over the best-performing baseline while retaining the general capabilities and incurring negligible overhead. Our results are robust, with the trends remaining similar for out-of-distribution languages and datasets.

Multi-agent systems have emerged as a promising paradigm for multimodal medical reasoning, but existing methods rely on two assumptions that hinder practical deployment: (i) always-active dense collaboration among multiple agents, and (ii) dependence on large cloud-based models that raise privacy and efficiency concerns. In this paper, we challenge both assumptions and propose SMART-Med, a fully local multi-agent method that treats collaboration as a learnable decision rather than a default behavior. Our key insight is that not every medical query requires costly multi-agent deliberation: many can be reliably solved by a single well-trained agent, while only uncertain cases benefit from expert collaboration. Based on this insight, SMART-Med introduces an uncertainty-driven routing mechanism that adaptively decides when to answer directly and when to recruit a small subset of expert agents, effectively converting dense multi-agent reasoning into selective, query-adaptive coordination. To make this routing reliable, we curate a new multimodal medical dataset MedVR-3K with high-quality reasoning traces and a difficulty hierarchy, and then fine-tune local agents using GRPO with two complementary rewards: a certainty-aware accuracy reward and a vision-language reward model for correctness and reasoning quality respectively. Experiments on four medical visual question answering benchmarks show that SMART-Med achieves competitive performance compared to existing MAS approaches while substantially reducing inference cost (reducing inference time by more than 3.8x and token usage by 72%). Moreover, after fine-tuning on MedVR-3K, SMART-Med yields an average improvement of 6.7% across Slake, PathVQA, VQA-RAD, and PMC-VQA.

Recent vision-and-language navigation (VLN) policies act on their visual history, and an incorrect action can take them off the instructed route. Paired interventions on four VLN policies show that clearing the history doubles pooled success at the onset of a deviation but lowers it on route, while later in a deviation, removing only the off-route frames helps much less. After a deviation, the retained history keeps the policy from turning where a turn is needed, which we call behavioral inertia. We propose MOON, a framework that equips a frozen VLN policy with two lightweight modules, adding less than 1% of its parameters. A deviation judge decides when to forget, raising an alarm that clears the history when it predicts that the agent is off route. A recovery policy, trained with GRPO from alarm states, learns how to move on, turning more of the successes found only by sampling into a single greedy rollout. Experiments on R2R-CE and RxR-CE show that MOON reaches the highest R2R-CE success rate among the compared methods and raises the success rate of its frozen base by 5.2 and 6.0 points, and an ablation with a random trigger indicates that this gain depends on when the history is cleared. Without retraining, MOON also improves zero-shot object-goal navigation on HM3D-OVON. Together, these results suggest that deviation recovery depends on knowing when to forget, not only how to act. Code is available at https://github.com/AutoLab-SAI-SJTU/MOON.

Crack propagation provides richer information than static crack observations for evaluating structural conditions of concrete infrastructure and planning maintenance strategies. Unlike static crack detection, which can now be largely automated using computer vision, vision-based crack tracking is a coupled cross-view and cross-time problem: the same physical crack must retain its identity as camera viewpoints change, while newly developed growth must be distinguished from existing crack segments in a unified reference view. Existing segmentation and 3D reconstruction methods can recover crack masks or scene geometry, but they do not preserve crack identity across views or quantify temporal extension. To address this gap, this study formulates crack growth monitoring as a cross-view and cross-time crack-instance tracking task and proposes a geometry-assisted, morphology-aware framework. The framework first uses image-based geometric registration to align observations from different viewpoints into a shared structural reference, then applies morphology-aware instantiation to assign physically interpretable crack identities, and finally performs temporal mapping in the unified view to distinguish newly developed growth from existing crack segments. To enable evaluation, a temporally annotated dataset was established and used, covering three representative cracking scenarios: single dominant crack propagation, mixed-mode branching, and multiple-crack interaction, with pixel-level masks and cross-time instance labels. Experimental results show that the proposed framework can establish loose cross-view identity correspondence under changing viewpoints, while strict spatial overlap and growth localization remain limited by early micro-crack segmentability and projection drift, highlighting directions for future crack-monitoring research.

High-resolution remote sensing (RS) imagery provides fine spatial detail over large areas, making it essential for fine-grained scene understanding. Referring Remote Sensing Image Segmentation (RRSIS) aims to segment regions described by natural language, yet many existing methods rely on sequential alignment--fusion pipelines with limited joint modeling of cross-scale visual hierarchies and linguistic dependencies. Moreover, dense attention mechanisms struggle with the sparsity and wide spatial dispersion of RS targets, while current datasets offer limited structured evaluation of multi-target references, multi-hop relational reasoning, and multi-granularity descriptions, limiting both model development and evaluation. To address these challenges, we propose a Cross-Scale Sparse Attention Model (CSSAM), which aligns multiscale visual features with complex referring expressions through two key modules. The Cross-scale Modality-aligned Fusion Module (CMFM) performs unified cross-scale cross-modal fusion, while the Dynamic Multimodal Sparse Hybrid Attention (DMSHA) selectively attends to informative visual--textual tokens, reducing dense visual interactions. Furthermore, we introduce M$^{3}$RRSIS-D, a RRSIS dataset explicitly designed around three core reasoning dimensions: multi-target (M$^{1}$), reference-dependent multi-hop (M$^{2}$), and multi-granularity (M$^{3}$) expressions. Extensive experiments on M$^{3}$RRSIS-D and two public RRSIS benchmarks demonstrate that CSSAM achieves competitive performance across accuracy, fine-grained reasoning, and high-IoU localization metrics, validating its effectiveness for complex semantic reasoning in realistic RS scenarios. Code, dataset, and model weights will be released upon acceptance.

Automated prostate lesion detection from diffusion-weighted MR imaging (DWI) conventionally relies on reconstructed images and does not fully use all acquisition information. We present an end-to-end network that uses complex multicoil DWI data and coil sensitivity information for prostate lesion segmentation. We included 269 fastMRI prostate examinations with additional voxel-level lesion annotations. Of these, 64 examinations contained PIRADS 4 or 5 lesions. The proposed network, VarNet-Seg, consists of a k-space variational network (VarNet) with six variational cascades and an image-space 3D segmentation U-Net. Comparator models used the same segmentation architecture with fully coil-combined images (CC-Seg), fixed root-sum-of-squares reconstruction (RSS-Seg), or fixed SENSE reconstruction (SENSE-Seg). We performed ablations with end-to-end training and data consistency (DC) enabled or disabled. We evaluated lesion segmentation using Dice and patient-level detection using AUROC in five-fold cross-validation with full sampling, simulated undersampling, added noise, or both. VarNet-Seg achieved the highest Dice at every undersampling level. At eight-fold acceleration, Dice was 0.22 for VarNet-Seg, 0.14 for SENSE-Seg and 0.16 for VarNet-Seg (no DC). The paired difference from SENSE-Seg was 0.08 (95% CI, 0.04-0.11). With full sampling, VarNet-Seg and RSS-Seg performed similarly (0.40 vs 0.39). Patient-level AUROC decreased from 0.75 with full sampling to 0.63 at eight-fold acceleration; corresponding RSS-Seg AUROCs were 0.76 and 0.64. The proposed multicoil network improved prostate lesion segmentation under simulated undersampling compared to conventional models. Matched ablations supported contributions from both end-to-end training and data consistency. Prospective evaluation is needed to establish whether this approach can be used clinically in accelerated prostate DWI.

In Ni-Mn-based Heusler alloys, the magnetocaloric effect originates from a first-order magnetostructural transition (FOMST). However, the influence of particle and grain size on thermal hysteresis, transition width, and nucleation and growth mechanism remains poorly understood, as mechanical grinding can introduce residual stresses and processing-induced defects. In this study, we examined the size dependence of the FOMST in gas-atomized Ni-Mn-Sn powder by combining microstructural characterization with temperature-dependent XRD, magnetometry, calorimetry and single-particle measurements, enabling decoupling of size dependence from processing-related influences. We observed a pronounced size dependence of the transition ranges (Ms-Mf) and (Af-As), which decrease with decreasing particle size and approach zero for single-crystalline particles below ~20-30 micrometer. The thermal hysteresis decreases with increasing particle size, reaching a lower limit of 17 K for large polycrystalline particles above 100 micrometer. EBSD reveals a nearly constant grain size, implying a transition from polycrystalline to single-crystalline particles as particle size decreases. In polycrystalline particles, multiple grains transform gradually over an extended temperature interval because grain boundaries act as barriers to the progression of transformation. In contrast, single-grain particles lack preferred nucleation sites, resulting in larger hysteresis, while the absence of barriers to transformation propagation causes the entire particle to transform in a narrow range. These results demonstrate that the size dependence of the FOMST is governed by the interplay between particle size, grain boundaries, and the free particle surface.

The mass concentration of particles smaller than 2.5um (PM2.5) is cheap to measure and widely monitored, whereas the particle number size distribution (PSD) requires costly instruments and is measured at only a few research stations. Yet many aerosol effects, such as vapor uptake, cloud droplet formation and deposition in the human respiratory tract, depend on the PSD rather than on mass. We explore whether recent generative deep learning models can nowcast the PSD: from the PM2.5 record and eight meteorological parameters over a 24h window, a model generates an ensemble of plausible hourly size distributions for the same window, reflecting that one mass can arise from many size distributions. Rather than proposing a single best model, we use two contrasting Finnish stations, the boreal forest site Hyytiälä and the urban site Kumpula, as a controlled testbed. On it, we benchmark six generative models against deterministic and climatological baselines, locally and after transfer between sites, and test physical constraints for neural network training. Flow matching performs best overall; its guided variant has the lowest continuous ranked probability score (CRPS) at both sites. A constraint based on a reduced aerosol general dynamic equation (GDE) lowers the CRPS by 5.5-11.5% at the training site and reduces errors after transfer between sites. Transfer remains harder than local estimation, identifying training on large multi-station data sets as the next step towards PSD estimates at unmonitored locations. This benchmark establishes physics-informed generative models as a promising route from routine air-quality monitoring to size-resolved aerosol information.

Video spatial reasoning requires multimodal large language models (MLLMs) to understand how objects are arranged and related across viewpoints. While these models can recognize objects in individual frames, integrating these observations to solve complex spatial tasks remains challenging. We examine route planning as a representative task and observe that existing models become substantially less accurate as the number of required turns increases, with several models achieving accuracy close to random guessing on routes requiring three or more turns. This finding suggests that recognizing landmarks alone is insufficient: models must also connect observations across viewpoints and track how landmarks are arranged in a shared space. Motivated by this observation, we introduce GeoLift, a training-free framework that uses feed-forward 3D reconstruction to map objects detected in individual frames into a shared metric 3D scene. GeoLift combines a frozen MLLM for 2D grounding with pretrained geometry models for scene reconstruction. It then uses shared geometric regions, called superpoints, to associate observations of the same object across views without a learned matcher. The reconstructed 3D map enables explicit computation of spatial relations and multistep routes, and provides geometric evidence for estimating distances and sizes. With all model weights frozen, GeoLift improves Qwen3-VL-4B's average score on VSI-Bench from 55.4% to 64.1%, including an increase in route-planning accuracy from 30.9% to 53.6%, and achieves 59.8% accuracy on MindCube. These results show that connecting frame-level object grounding through a shared 3D map can improve complex spatial reasoning without task-specific training.

Ensemble decoding improves decoding reliability by combining component decoders with complementary failure events. Existing approaches typically obtain this diversity from code-specific structure, such as code automorphisms, multiple parity-check representations, or targeted modifications of the message-passing graph. In this work, we instead learn the diversity of an ensemble of weighted belief propagation decoders by jointly optimizing trainable Tanner-graph edge weights. A central design question is how the training objective should distribute gradient among the ensemble members to promote complementary decoding behavior. To this end, we compare four ensemble objectives under direct gradient-based optimization and identify best-member training as the most effective approach. For a (64,32) progressive edge-growth (PEG) low-density parity-check (LDPC) code, the resulting ensemble outperforms the considered trained and untrained belief-propagation-based ensemble baselines. With 10 members, it provides a gain of approximately 0.66 dB over a single belief propagation decoder at a target frame error rate (FER) of $10^{-4}$. Similar behavior is observed for a (128,64) CCSDS LDPC code, indicating that the learned complementarity is not specific to PEG. Although the ensemble performance improves substantially, all trained members show a higher standalone FER. The gain therefore results from learned specialization and complementary failure events rather than uniformly improved component decoders.

Hybrid dynamical systems flow continuously but jump at discrete events. Physics-informed neural networks (PINNs) fit data while enforcing known dynamics, but a single network is smooth in time and cannot model these jumps. We show that the more closely a PINN follows a jump, the larger its physics loss must be. We propose hybrid-time PINNs (HT-PINNs), which assign one network to each flow interval of the hybrid time domain and train the networks jointly with the jump times and the unknown parameters. The number of jumps and the sequence of modes are found by enumeration. HT-PINNs recover the number of jumps and the restitution coefficient of a bouncing ball. In a MuJoCo simulation of robotic disassembly, where a robot pulls a press-fitted pin out of a housing through a spring, HT-PINNs find the number of jumps and the sequence of modes, and estimate the static and kinetic friction coefficients with small errors of 0.1% and 1.7%.

In the era of large-scale pretrained models, zero-shot generalization has increasingly blurred the boundary between known and unknown concepts. However, current evaluations do not directly reveal which open-world concepts different vision-language models can recognize. To study this problem, we introduce OODConceptBench, a hierarchical benchmark for open-world concept recognition on image-concept pairs. The full benchmark set (L1) contains 30,752 images, and the harder subset (L2) contains 14,285 images. The benchmark covers 560 concept phrases across seven open-world categories. We evaluate 37 vision-language models across release year, model family, and variant. We obtain three main findings. First, target-concept recognition is lower on L2 than on L1 across all 37 evaluated models. Second, later-released models do not necessarily achieve stronger open-world concept recognition. Third, an increase in overall recognition does not imply consistent improvement across all benchmark categories. Together, these results show that OOD remains a challenge even in the era of large-scale vision-language pretraining.

Robot policy evaluation informs deployment decisions and policy improvement, but physical testing and simulator preparation require substantial effort. World models offer a complementary route through closed-loop interaction in generated environments. Recent benchmarks have demonstrated their potential as policy evaluators through outcome-based evaluation (i.e., assessing task success or failure). However, such evaluation does not establish whether world models faithfully simulate how policies fail. Misrepresenting failures can lead to incorrect diagnoses of policy weaknesses and make policies appear ready for deployment even when small scene changes cause them to fail. We therefore ask: do world models faithfully simulate how and at which stages policies fail, and under what conditions they transition from success to failure? To answer this question, we introduce WorldFailBench with three evaluation tasks. Failure reason tests whether models reproduce the observable errors policies make. Failure stage assesses whether failures concentrate at the same stages of full-task execution. Failure boundary examines whether models reproduce the conditions under which policies transition from success to failure under controlled scene changes in simulation. The benchmark covers manipulation tasks in the LIBERO and RoboTwin simulation environments and real-world manipulation tasks performed by the G1 humanoid robot. Experiments across seven world models reveal substantial gaps in their understanding of policy failures, providing guidance for developing more reliable world models for policy evaluation.

Video world models are expected to generate videos that accurately follow specified camera trajectories while maintaining consistent 3D structures. However, existing models are predominantly trained on densely sampled videos with small inter-frame camera motions. When applied to sparse, large step trajectories, they often produce frozen frames, inaccurate camera motion, and severe geometric drift. We attribute this degradation to a mismatch between the gradual viewpoint transitions seen during training and the large viewpoint changes required at inference time. To address this problem, we propose Dense Trajectory Teacher Distillation (DTTD), a post-training framework that transfers a model's generation capability from dense trajectories to large camera motions. Given a sparse trajectory, DTTD interpolates additional camera poses to construct a dense trajectory covering the same path. The frozen base model then generates reliable teacher videos along this familiar dense trajectory, and pose aligned teacher observations are used to evaluate candidate videos generated under the original sparse trajectory. In this way, the model serves as its own generation teacher, requiring neither additional video data nor a stronger external video generator. We further combine dense teacher supervision with geometry aware rewards that assess cross frame consistency and camera trajectory following, and use the resulting preferences for offline preference optimization. To stably expand the model's motion range, we introduce a cumulative curriculum that progressively incorporates larger motion steps while retaining simpler trajectories to reduce capability forgetting.

Successful tool-use trajectories are not necessarily useful supervision. Even successful interactions may underrepresent multi-step dependencies or contain actions that are unjustified by the information visible to the agent at decision time. We identify two coupled requirements for effective trajectory supervision: compositional coverage, so that important observation-action dependencies are represented, and decision grounding, so that each supervised action is supported by the learner-visible interaction history. We introduce Trajectory2Gym, which constructs executable interactions around under-supported compositional structures and derives supervision only from grounded decisions in their realized executions. Structural-support selection allocates the resulting experience, while parent-normalized supervised fine-tuning preserves the semantic interaction as the unit of training weight. This reframes trajectory generation as the deliberate construction of learnable interaction experience, rather than the collection of successful trajectories. Across five Qwen3 scales, Trajectory2Gym improves Tau1, Tau2, and BFCL at every tested scale, increasing the equal-weight three-benchmark mean by 16.9-29.8 percentage points. The gains are strongest in stateful and multi-turn interaction, where later decisions depend on earlier observations and feedback. Our results suggest that scaling tool-use agents depends not only on collecting more successful trajectories, but on constructing interactions whose dependencies are represented and whose decisions are learnable from the agent's visible history.