Science & Mathematics

Reference-based 3D scene stylization transfers the appearance of a reference image onto a reconstructed 3D scene while preserving scene geometry. Recent methods perform this transfer by optimizing the appearance representation of neural scene models including Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS). These methods optimize an RGB appearance representation, in which a single per-point colour combines the effects of material reflectance, environment illumination, together with the shading it produces, and the viewing direction. When stylization modifies this combined colour, the illumination effects present during stylization are baked into the stored appearance, so the stylized scene cannot be rendered under different illumination. To address this limitation, we present RxGS, a relightable 3D scene stylization framework built on a Bidirectional Reflectance Distribution Function (BRDF)-aware 3D Gaussian representation. Rather than modifying a per-point RGB colour, RxGS optimizes only the diffuse albedo field recovered by inverse rendering, while leaving roughness, metallicity, surface normals, and the estimated environment lighting unchanged. As a result, the stylized scene keeps all the components a physically-based renderer needs other than the edited albedo, and can therefore still be relit. RxGS optimizes the albedo field using a nearest-neighbour feature matching loss together with a spatial smoothness prior over Gaussian centres. We evaluate the method using a Geometry Consistency Cost (GCC) based on optical-flow warping error. RxGS achieves the lowest GCC on all evaluated NeRF-Synthetic scenes and remains competitive on real-captured scenes, while additionally supporting relightable stylization, which existing reference-based 3D scene stylization methods do not provide.

Zero-shot diffusion image compression based on reverse channel coding achieves high perceptual quality at low bitrates without retraining, yet its bitstream remains vulnerable to bit flips. Prior work has shown that a joint combinatorial representation can amplify a single bit flip into substitutions of multiple atoms and that indexing each atom independently confines this error propagation. However, independent indexing reduces the number of atoms that fit within the bit budget and gives the decoder no indication of which atoms are corrupted. To address these issues, we propose BitFlat, a training-free error-localizing bitstream framework with two components. Blockwise Locality Alignment (BLA) decomposes the globally coupled atom selection at each diffusion step into independent selection units, each describing one signed atom from a dedicated codebook with a fixed-width field, so that any bit error affects only the unit in which it occurs. Reliability-Aware Diffusion Control (RADC) uses the bits saved by BLA to impose structured valid-codeword constraints within each field and converts the decoding state of each field into a binary reliability gate, which turns detected corrupted atoms into erasures at known positions and excludes them from noise synthesis. BitFlat requires neither target residuals nor additional network forward passes and leaves the diffusion model unchanged. Experiments on Kodak, DIV2K, and CLIC2020 show that BitFlat outperforms conventional, learned, and diffusion-based codecs at bit error rates from $10^{-4}$ to $10^{-2}$ and, even with $26.7\%$ fewer bits, surpasses Robust Turbo-DDCM.

Harmful fine-tuning can rapidly erode the safety alignment of large language models, creating substantial risks when aligned models are adapted to downstream tasks. Existing alignment-stage defenses use harmful examples and simulated harmful perturbations during safety alignment to prepare models for later harmful fine-tuning, yet the collective effects of harmful weight perturbations remain underexplored. We identify a collective failure pattern: disjoint parts of a harmful perturbation have limited effects in isolation, but their joint application increases the model's harmfulness beyond the sum of their individual effects. Guided by this observation, we propose Joint Weight Perturbations (JWP), an alignment-stage defense that masks different parts of a harmful perturbation estimated from the current model and trains the model to respond safely in the resulting states. This exposes the model to different combinations of harmful perturbations and reduces its sensitivity to their accumulation during subsequent harmful fine-tuning. Experiments across downstream tasks and model families show that JWP offers stronger resistance to harmful fine-tuning than prior defenses while largely preserving task performance.

In network analysis, nodes with different degrees are usually weighted differently in tasks such as link prediction. For example, two individuals connected to the same high-degree node (e.g., a celebrity) are often considered less likely to know each other than if they share a low-degree connection. In this paper, we theoretically justify the reasoning behind this heuristic using a graph generation model. In this model, nodes are embedded in a latent metric space, and edges are more likely to form between nearby nodes. We analytically derive degree-based weighting schemes for commonly used link prediction and similarity measures. In particular, we identify the optimal weighting scheme for the Common Neighbors index and extend our analysis to other popular heuristics, including the graph distance. Our theoretical findings align with previous empirical observations, and we further validate them through experiments on both synthetic and real-world networks.

Continuous vision-and-language navigation (VLN) is typically evaluated by goal-reaching success, assuming it reflects an agent's ability to execute language-grounded behavior. We show that this assumption can fail: permissive contact models in current simulators may mask physically infeasible actions, creating a physical executability gap between semantic competence and embodied realization. To bridge this gap, we introduce Visual-Semantic Traversability Alignment (VeSTA), a framework that aligns VLN policies with physically executable futures. We define visual-semantic traversability as the ability to predict instruction-conditioned trajectories that are semantically valid, geometrically feasible, and physically robust. VeSTA addresses this misalignment through two complementary mechanisms: improving executable support in trajectory generation and risk-aware selection among multimodal futures. Specifically, Trajectory Relative Group Policy Optimization uses strict-contact feedback to reshape the candidate distribution toward executable regions, while the Traversability Risk Critic learns a risk-adjusted scalar utility for candidate-wise future selection. Experiments on R2R and RxR reveal substantial dependence of existing VLN agents on simulator contact semantics. Under strict-contact evaluation, VeSTA substantially reduces collision-induced failures while preserving or improving goal-reaching performance, and controlled real-world robot experiments provide evidence that the learned preference transfers beyond simulation.

Direct wind sensing from a multi-rotor UAV enables localized atmospheric measurements, but the reliability of the anemometer depends on vehicle motion. This work presents a vehicle-state-aware confidence method that uses horizontal and vertical vehicle speed to quantify the acquisition of each wind observation. The resulting confidence score is used to adapt measurement influence to a two-dimensional vector estimator. Experiments were conducted during horizontal translation, vertical motion, and held-out autonomous waypoint flight. The proposed two-feature formulation achieved an ROC-AUC of 0.995, with 98.84% of translation observations identified as unreliable. The confidence-aware adaptation reduced median estimator gain during translation from 0.207 to 0.064 while retaining median gain of 0.151 during hover. The results show that vehicle state provides useful proprioceptive context for regulating the influence of direct airborne wind observations.

As deep neural networks are required to continuously adapt to evolving data, class-incremental learning (CIL) has become an active research area. However, most existing works primarily focus on improving accuracy while overlooking confidence calibration, which measures the trustworthiness of the model's predictions. Although temperature scaling (TS) is effective for static calibration, it has rarely been explored in continual settings where data distributions evolve over time. In this paper, we introduce Multi-Group Temperature Scaling for Continual Calibration (MT-CC), a novel post-hoc calibration framework that addresses the asymmetric calibration behavior in CIL, where the rate of confidence changes does not align with the rate of accuracy degradation across tasks, leading to different levels of miscalibration. Our key contribution is to design multiple temperature parameters through a lightweight grouping network that separates samples by confidence statistics and task-aware signals in CIL. Furthermore, a Groupwise Difference-between-Confidence-and-Accuracy (GDCA) regularization is incorporated to promote intra-group consistency. Experiments demonstrate that MT-CC significantly improves calibration performance while maintaining accuracy, offering a principled framework for reliable continual learning.

We study language-conditioned local navigation, where a robot must navigate to a text-specified target visible in its initial observation and stop at the described location. End-to-end navigation policies require diverse training data to generalize across scenes, yet real-robot data is costly to collect and simulated data leaves a sim-to-real gap. Internet egocentric videos offer a scalable source of real-world navigation experience. Our key insight is to learn goal-directed visual dynamics from Internet egocentric videos, and use these dynamics to support robot action generation. To this end, we propose ScaleNav, a unified world-action framework with three-stage training. We first perform instruction-conditioned video pretraining on these videos to learn how observations evolve as the camera approaches instructed targets across diverse real-world scenes. Action alignment then uses action-labeled simulated data to train the model to predict executable action chunks. Joint fine-tuning subsequently optimizes all trainable components together for navigation. ScaleNav achieves success rates of 86.5% on a single-view adaptation of Short-Horizon OVON and 94.9% on our HabitatGS-LN benchmark, outperforming the evaluated baselines on both benchmarks. We further deploy and evaluate ScaleNav on a real quadruped robot, with results supporting the model's ability to generalize to real-world navigation. Code and videos are available at https://zju3dv.github.io/ScaleNav/

Data-driven learning for general-purpose robotic manipulation presents a fundamental trade-off between data-collection cost and adaptability to real-world environments. Imitation learning (IL) requires costly human demonstrations, whereas simulation-based learning is vulnerable to sim-to-real gaps in dynamics and sensor observations. We propose ReDOA (Real-World Distillation via Observation Alignment), a framework that decouples control learning from perception learning to learn image-based policies in the real world. A state-based teacher policy trained in simulation is executed on a real robot, and its control knowledge is distilled into an image-based student policy using the collected trajectories. Observation Alignment resolves observation mismatches between data collection and deployment. This framework collects successful, efficiently executed trajectories paired with real-world visual observations, without requiring demonstrations from human experts. In real-robot experiments on three non-prehensile manipulation tasks, including bimanual manipulation, our method achieves better overall performance when jointly considering task success and execution efficiency than both IL based on human demonstrations and policy distillation in simulation. https://isri-aist.github.io/redoa-project-page/

Generative policies have achieved strong performance in imitation learning tasks, but their success is often constrained by the limited availability of expert demonstrations. Although preference-based fine-tuning can leverage suboptimal rollouts to improve performance, existing approaches heavily rely on expensive human annotations. In this work, we introduce Progress-Guided Preference Generation (PGPG), an unsupervised preference alignment framework that eliminates the need for human-in-the-loop supervision. Our method leverages ordered coverage alignment to measure sequential progress relative to a small set of expert demonstrations, enabling the automatic extraction of pseudo-preference labels from autonomously collected rollouts. Furthermore, we introduce key enhancements to improve the reliability of these progress-guided preferences, enabling effective fine-tuning of pretrained generative policies without requiring manually designed reward functions. Extensive experiments across a suite of challenging long-horizon manipulation tasks demonstrate that PGPG outperforms standard supervised fine-tuning baselines. Moreover, our conceptually simple framework is broadly applicable across multiple classes of generative policies, demonstrating strong versatility when integrated with both diffusion and flow-matching policies.

On-policy distillation (OPD) enhances the capabilities of small language agents by providing dense, token-level supervision from a teacher model on student-generated trajectories. However, standard OPD implicitly assumes that credit is uniformly distributed across long-horizon trajectories, assigning the same strength of teacher supervision to different actions. Through credit-distribution analysis and outcome-reversal counterfactual experiments, we find that trajectory credit is concentrated on a small number of critical actions. Moreover, even when the trajectory is held fixed, reversing only the final outcome substantially changes the credit assigned to multiple interaction turns. These findings show that the uniform-credit assumption fails to identify critical decisions and ultimately degrades the model's reasoning accuracy. To address these challenges, we propose ToolOPD, an outcome-aware on-policy distillation framework designed for long-horizon tool use. ToolOPD first summarizes each completed trajectory and its outcome into a reusable hindsight skill. It then measures how this skill changes the teacher's probability of each student action token and converts the change into outcome-aware credit. The resulting credit preserves supervision that supports successful decisions or discourages failed ones, while suppressing irrelevant and conflicting signals. ToolOPD uses these credits to reweight the action-level distillation loss and jointly optimizes it with reinforcement learning. Extensive experiments across multiple long-horizon tool interaction benchmarks demonstrate that ToolOPD significantly outperforms existing reinforcement learning and self-distillation baselines.

Infrared and visible images provide complementary cues for low-light image generation, and physics-based methods can better exploit this complementarity. However, existing physics-based methods still rely heavily on paired data. To address these limitations, we propose Multi-View Routing Mixture-of-Experts (MVR-MoE), a framework for physically consistent TeX generation that learns multi-view structural representations. MVR-MoE mines intrinsic structural cues from infrared and visible images to organize them into multi-view representations, while the proposed multi-view hyper connections expand the feature space and preserve both shared evidence and modality-specific details. A hierarchical routing mechanism further regulates feature utilization at both the structural-view level and the expert level, enabling adaptive exploitation of heterogeneous cues. The predicted temperature, emissivity, and texture maps are constrained by the Stefan-Boltzmann law to ensure physically consistent. Extensive experiments on challenging datasets demonstrate that MVR-MoE achieves superior robustness and discriminability while remaining lightweight.

Large language models have recently been used to discover heuristics through iterative program search, offering a promising approach to automated algorithm design. However, candidates selected by average performance may degrade across instance families, random seeds, and problem scales because mean-based objectives can obscure rare but severe failures. We propose CVaR-RMD, a risk-sensitive evolutionary framework that uses an LLM to mutate executable heuristic programs and ranks candidates by empirical conditional value-at-risk over heterogeneous evaluation conditions. Across three combinatorial optimization domains, CVaR-based selection improves upper-tail performance and reduces failure rates while maintaining competitive mean performance. These gains persist on larger unseen instances and independently curated within-domain benchmarks

Linear stochastic approximation (SA) is widely used to estimate fixed points from noisy observations, but most finite-time analyses assume that the underlying operator remains fixed. We study linear SA with time-varying affine operators, where the fixed points change over time and the observation noise may be correlated across rounds. For constant-stepsize updates, we establish a bound on the cumulative expected squared distance between the iterates and the corresponding fixed points. The bound accounts for the initial error, observation noise, and changes in the fixed points over time. We also establish lower bounds for scalar Gaussian models, showing that, under specified conditions, a suitably chosen constant stepsize achieves the smallest possible worst-case cumulative error up to constant factors, even compared with estimators that use future observations. These results further quantify the effect of temporal correlation on estimation accuracy. Since choosing an appropriate stepsize generally requires knowledge of the noise level and how much the fixed points change, we develop an adaptive method that uses an additional bounded, unbiased observation of the fixed point to achieve an almost optimal algorithm without needing the knowledge of the noise level and the fixed point variation over time. Finally, we apply our results to value-function estimation using synchronous tabular temporal-difference learning in changing environments, assuming generative-model access and a common invariant distribution.

Inference-time steering is crucial for deploying Large Language Models (LLMs) in real-world applications, where they must flexibly adjust behavior preferences. For instance, a model should dynamically adjust its caution level based on the risk of the scenario. This requirement introduces a challenge: modulating the intensity of LLM preferences. Prompt engineering can require task-specific prompt design, while fine-tuning requires additional optimization and data. These requirements can make repeated behavioral adjustments costly. To address this, we introduce NeuroSteer, a plug-and-play framework that enables flexible, quantitative adjustment of LLM preferences at inference time. NeuroSteer employs a Sparse Autoencoder (SAE) as its controller. Comparing neuron activation frequencies between positive and negative examples isolates "feature neurons" that correspond to target preferences such as politeness and positive sentiment. During inference, the activations of selected SAE features are amplified in the latent representations to steer the output distribution. Users can adjust the strength of this steering on a per-scenario or per-user basis, granting the LLM fine-grained, quantifiable trait modulation. Our experiments show that NeuroSteer can alter sentiment, toxicity, stance, and politeness at inference time. It achieves higher target-direction steering scores than the compared baselines in several settings, with task- and strength-dependent trade-offs in generation quality.