While recent advancements in generative image editing models have achieved stunning visual fidelity, it remains an open question whether these systems possess an intrinsic knowledge of real-world lighting. Existing benchmarks typically evaluate high-level plausibility of perceptual light transport on curated internet imagery, using VLMs or human judgement, or they rely on synthetically generated datasets. In this work, we introduce the 3D-anchored Light Probe (3DLP) benchmark, for which we have captured a new high-fidelity HDR dataset of real-world lighting changes. The dataset consists of 1K image pairs of diverse indoor scenery in which light probes are physically turned on and off. To allow for a granular performance analysis, we annotated specific image regions such as cast shadows or metallic surfaces. With this data, we evaluate a range of state-of-the-art image editing models by measuring how well their light probe edits align with reality. The evaluation uses two new scores to compensate for AI-generated photographic effects, such as adjusted white balance. Our results show that the overall performance of models differs considerably, with differences slightly less pronounced for specular highlights. The best image editing models are remarkably consistent with real-world physics, however, they still leave room for improvement. We observe that image regions that receive less light from the light probe are more prone to errors for all models. Furthermore, building on their success in evaluating macroscopic lighting plausibility, we test VLMs on our task but find that they are unsuitable for pixel-level light transport analysis. We will make the benchmark, together with the real-world dataset, publicly available to encourage future research on this topic.
展开完整摘要收起摘要↓
While recent advancements in generative image editing models have achieved stunning visual fidelity, it remains an open question whether these systems possess an intrinsic knowledge of real-world lighting. Existing benchmarks typically evaluate high-level plausibility of perceptual light transport on curated internet imagery, using VLMs or human judgement, or they rely on synthetically generated datasets. In this work, we introduce the 3D-anchored Light Probe (3DLP) benchmark, for which we have captured a new high-fidelity HDR dataset of real-world lighting changes. The dataset consists of 1K image pairs of diverse indoor scenery in which light probes are physically turned on and off. To allow for a granular performance analysis, we annotated specific image regions such as cast shadows or metallic surfaces. With this data, we evaluate a range of state-of-the-art image editing models by measuring how well their light probe edits align with reality. The evaluation uses two new scores to compensate for AI-generated photographic effects, such as adjusted white balance. Our results show that the overall performance of models differs considerably, with differences slightly less pronounced for specular highlights. The best image editing models are remarkably consistent with real-world physics, however, they still leave room for improvement. We observe that image regions that receive less light from the light probe are more prone to errors for all models. Furthermore, building on their success in evaluating macroscopic lighting plausibility, we test VLMs on our task but find that they are unsuitable for pixel-level light transport analysis. We will make the benchmark, together with the real-world dataset, publicly available to encourage future research on this topic.
Modern image generation model rapidly grows their sizes to meet high-fidelity image synthesis. However, they gradually become unaffordable for their enormous parameter consumption and computation budget that lead to massive resources requirement and gpu memory footprint. In this paper, we propose TMP, the first Tree-structured Mixed-policy Pruning framework that generalizes prevalent image tasks (T2I and TI2I) and architectures (Mixture-of-Experts (MoE) and Diffusion transformer (DiT)). It could be applied to the step-distilled models and contribute as the last stage. We perform experiments upon current open-sourced SOTA HunyuanImage-3.0 instruct and a popular efficient model Z-Image turbo. The proposed pruning framework manages to compress HunyuanImage 3.0 from 80B to 20B parameters at 75% reduction ratio, sacrificing limited generation quality. We also optimize to enable the inference of the pruned 20B version of HunyuanImage 3.0 on a single 24GB 4090 GPU by engineering skills. The inference script and model weight have been integrated into the existing HunyuanImage3.0 open-source github and huggingface repository. Besides, we prove the efficacy of TMP by compressing Z-Image turbo from 6B to 4B (33% reduction) with negligible degradation.
展开完整摘要收起摘要↓
Modern image generation model rapidly grows their sizes to meet high-fidelity image synthesis. However, they gradually become unaffordable for their enormous parameter consumption and computation budget that lead to massive resources requirement and gpu memory footprint. In this paper, we propose TMP, the first Tree-structured Mixed-policy Pruning framework that generalizes prevalent image tasks (T2I and TI2I) and architectures (Mixture-of-Experts (MoE) and Diffusion transformer (DiT)). It could be applied to the step-distilled models and contribute as the last stage. We perform experiments upon current open-sourced SOTA HunyuanImage-3.0 instruct and a popular efficient model Z-Image turbo. The proposed pruning framework manages to compress HunyuanImage 3.0 from 80B to 20B parameters at 75% reduction ratio, sacrificing limited generation quality. We also optimize to enable the inference of the pruned 20B version of HunyuanImage 3.0 on a single 24GB 4090 GPU by engineering skills. The inference script and model weight have been integrated into the existing HunyuanImage3.0 open-source github and huggingface repository. Besides, we prove the efficacy of TMP by compressing Z-Image turbo from 6B to 4B (33% reduction) with negligible degradation.
Unlike diffusion-based models that operate in continuous latent spaces, autoregressive unified multimodal models produce images by sequentially predicting discretized visual tokens. These tokens are derived from a codebook that maps embeddings to quantized visual patterns. The language-like architecture enables unified multimodal models to effectively capture text conditional information for generation, making them promising for text-to-image tasks. This also raises an interesting question: how safe are the images generated in such an autoregressive way? In this work, we propose iterative self-improving codebooks for safe autoregressive generation. We leverage the understanding and judgment capabilities of the unified multimodal model itself to identify unsafe generated images without human annotation. Subsequently, the inherent representations in the codebook are fixed to eliminate harmful mappings. Our method comprises two steps: first, we use the unified model to identify unsafe generations and construct corresponding harmful and safe image-text pairs. These pairs are used to construct the Harmful Space and guide updates to the codebook, thereby eliminating harmful outputs. Second, we perform adaptive fine-tuning on the codebook within the harmless space using safe image-text pairs to ensure the quality of generated images. These two steps are repeated until no further improvement is observed, producing a safety-enhanced model codebook. Without additional external feedback, the safety of models is improved iteratively.
展开完整摘要收起摘要↓
Unlike diffusion-based models that operate in continuous latent spaces, autoregressive unified multimodal models produce images by sequentially predicting discretized visual tokens. These tokens are derived from a codebook that maps embeddings to quantized visual patterns. The language-like architecture enables unified multimodal models to effectively capture text conditional information for generation, making them promising for text-to-image tasks. This also raises an interesting question: how safe are the images generated in such an autoregressive way? In this work, we propose iterative self-improving codebooks for safe autoregressive generation. We leverage the understanding and judgment capabilities of the unified multimodal model itself to identify unsafe generated images without human annotation. Subsequently, the inherent representations in the codebook are fixed to eliminate harmful mappings. Our method comprises two steps: first, we use the unified model to identify unsafe generations and construct corresponding harmful and safe image-text pairs. These pairs are used to construct the Harmful Space and guide updates to the codebook, thereby eliminating harmful outputs. Second, we perform adaptive fine-tuning on the codebook within the harmless space using safe image-text pairs to ensure the quality of generated images. These two steps are repeated until no further improvement is observed, producing a safety-enhanced model codebook. Without additional external feedback, the safety of models is improved iteratively.
Modern image generation demands a single model that unifies diverse capabilities, including text-to-image (T2I), local editing, and global editing. However, these capabilities are rarely naturally aligned and often conflict. For instance, editing tends to degrade T2I performance, while global and local editing interfere with each other. Consequently, effectively composing these capabilities has become a central challenge for image generation model training. To tackle this, we introduce DanceOPD, an on-policy generative field distillation framework for flow-matching models that routes each sample to one capability field, queries one low-noise student-induced state, and trains with a simple velocity MSE objective. With each capability source defined as a velocity field over the shared flow state space, the student learns from fields queried on its own rollout states to compose expert capabilities. This formulation also absorbs operator-defined fields such as classifier-free guidance. Comprehensive experiments on T2I, editing, realism-field absorption, and CFG absorption show that our approach improves multi-capability composition, strengthening target capabilities while preserving anchor generation quality. We believe this work establishes a practical route for generative field distillation in flow-matching models.
展开完整摘要收起摘要↓
Modern image generation demands a single model that unifies diverse capabilities, including text-to-image (T2I), local editing, and global editing. However, these capabilities are rarely naturally aligned and often conflict. For instance, editing tends to degrade T2I performance, while global and local editing interfere with each other. Consequently, effectively composing these capabilities has become a central challenge for image generation model training. To tackle this, we introduce DanceOPD, an on-policy generative field distillation framework for flow-matching models that routes each sample to one capability field, queries one low-noise student-induced state, and trains with a simple velocity MSE objective. With each capability source defined as a velocity field over the shared flow state space, the student learns from fields queried on its own rollout states to compose expert capabilities. This formulation also absorbs operator-defined fields such as classifier-free guidance. Comprehensive experiments on T2I, editing, realism-field absorption, and CFG absorption show that our approach improves multi-capability composition, strengthening target capabilities while preserving anchor generation quality. We believe this work establishes a practical route for generative field distillation in flow-matching models.
作者Minghao Yin, Jiahao Lu, Wenbo Hu, Wang Zhao, Shan Ying, Kai Han
Modern video diffusion transformers position their tokens through RoPE on the (u,v,t) axes — a description of the camera's sampling grid that says nothing about the 3D structure of the scene. We observe that the geometric relation between two camera rays is captured by the Plucker reciprocal product, which is bilinear in the two rays — the same algebraic form as the dot product in Transformer attention. Building on this analogy, we propose RayPE, a positional-encoding extension that injects per-token 6D Plucker coordinates additively into the queries and keys of self-attention, with a query/key flip arrangement under which the symmetric identity configuration coincides exactly with the reciprocal product. The injection is additive, the resulting attention score decomposes into a content term, a geometry term, and two content and geometry cross-terms — all of which our experiments find individually necessary. To make the encoding stable across video data with heterogeneous camera-translation scales (SfM, deep SLAM, metric), we further decouple ray direction from moment magnitude, gate the encoding by a learned function of the log-magnitude, and apply RMSNorm to align it with the QKNorm-normalized content branch. The full module adds less than 0.1% parameters to a pretrained video DiT, is zero-initialized to start from the pretrained weights, and improves camera controllability, cross-frame 3D consistency, and overall video quality on a four-dataset training mixture.
展开完整摘要收起摘要↓
Modern video diffusion transformers position their tokens through RoPE on the (u,v,t) axes — a description of the camera's sampling grid that says nothing about the 3D structure of the scene. We observe that the geometric relation between two camera rays is captured by the Plucker reciprocal product, which is bilinear in the two rays — the same algebraic form as the dot product in Transformer attention. Building on this analogy, we propose RayPE, a positional-encoding extension that injects per-token 6D Plucker coordinates additively into the queries and keys of self-attention, with a query/key flip arrangement under which the symmetric identity configuration coincides exactly with the reciprocal product. The injection is additive, the resulting attention score decomposes into a content term, a geometry term, and two content and geometry cross-terms — all of which our experiments find individually necessary. To make the encoding stable across video data with heterogeneous camera-translation scales (SfM, deep SLAM, metric), we further decouple ray direction from moment magnitude, gate the encoding by a learned function of the log-magnitude, and apply RMSNorm to align it with the QKNorm-normalized content branch. The full module adds less than 0.1% parameters to a pretrained video DiT, is zero-initialized to start from the pretrained weights, and improves camera controllability, cross-frame 3D consistency, and overall video quality on a four-dataset training mixture.
Recent game world models can synthesize visually plausible, action-conditioned rollouts. However, their interaction behaviors often remain limited to exploratory or wandering trajectories, and physical dynamics are typically learned as implicit correlations from data rather than as controllable variables. This limitation hinders their applicability to authored game environments, where physical rules are deliberately designed and require explicit manipulation. We introduce PhysEditWorld, a multimodal dataset with physical parameters, with a primary focus on gravity in this initial version. At its core, PhysEditWorld is built upon a replay paradigm implemented with a UE5 replay-and-rendering pipeline. Each scenario records a normalized action trace and replays the same initial state, character controller, action sequence, and camera policy under multiple gravity configurations, enabling controlled and attributable physical variation. PhysEditWorld contains 12 cinematic UE5 scenes, over 100 hours of gameplay interactions, and more than 60 million rendered rollout frames. Each sample provides synchronized multimodal signals, including RGB, depth, normals, audio, action traces, camera trajectory, engine states, semantic annotations, and explicit gravity labels. We further conduct initial utility studies on both generative video models and world understanding models, demonstrating that PhysEditWorld enables improved gravity-faithful dynamics modeling, enhances consistency under physical edits, and provides a scalable foundation for controllable world modeling research.
展开完整摘要收起摘要↓
Recent game world models can synthesize visually plausible, action-conditioned rollouts. However, their interaction behaviors often remain limited to exploratory or wandering trajectories, and physical dynamics are typically learned as implicit correlations from data rather than as controllable variables. This limitation hinders their applicability to authored game environments, where physical rules are deliberately designed and require explicit manipulation. We introduce PhysEditWorld, a multimodal dataset with physical parameters, with a primary focus on gravity in this initial version. At its core, PhysEditWorld is built upon a replay paradigm implemented with a UE5 replay-and-rendering pipeline. Each scenario records a normalized action trace and replays the same initial state, character controller, action sequence, and camera policy under multiple gravity configurations, enabling controlled and attributable physical variation. PhysEditWorld contains 12 cinematic UE5 scenes, over 100 hours of gameplay interactions, and more than 60 million rendered rollout frames. Each sample provides synchronized multimodal signals, including RGB, depth, normals, audio, action traces, camera trajectory, engine states, semantic annotations, and explicit gravity labels. We further conduct initial utility studies on both generative video models and world understanding models, demonstrating that PhysEditWorld enables improved gravity-faithful dynamics modeling, enhances consistency under physical edits, and provides a scalable foundation for controllable world modeling research.
While personalized image generation has achieved remarkable progress, multi-reference image generation (MRIG) remains a challenging task. Most existing benchmarks fail to adequately evaluate complex MRIG scenarios, hindering further progress in this area. To better assess model performance on complex MRIG tasks, we introduce OmniRef-Bench, a benchmark that covers complex combinations of reference image types and a large number of reference images. Evaluations on OmniRef-Bench show that mainstream open-source models struggle in complex MRIG scenarios, and their performance deteriorates significantly as the number of mixed-type reference images increases. To address this issue, we propose DyRef, a two-stage training framework. In the first stage, supervised fine-tuning equips the model with the basic capability to handle complex MRIG tasks. In the second stage, we introduce Difficulty-aware Advantage Reweighting (DAR) and Discriminative Reward Scaling (DRS). DAR dynamically adjusts the optimization objective to improve performance when handling a large number of mixed-type reference images. DRS enlarges intra-group reward differences for more effective policy optimization. Experiments demonstrate that DyRef significantly improves the performance of open-source models on OmniRef-Bench and single-image editing benchmarks, demonstrating the effectiveness and generalization capability of our approach.
展开完整摘要收起摘要↓
While personalized image generation has achieved remarkable progress, multi-reference image generation (MRIG) remains a challenging task. Most existing benchmarks fail to adequately evaluate complex MRIG scenarios, hindering further progress in this area. To better assess model performance on complex MRIG tasks, we introduce OmniRef-Bench, a benchmark that covers complex combinations of reference image types and a large number of reference images. Evaluations on OmniRef-Bench show that mainstream open-source models struggle in complex MRIG scenarios, and their performance deteriorates significantly as the number of mixed-type reference images increases. To address this issue, we propose DyRef, a two-stage training framework. In the first stage, supervised fine-tuning equips the model with the basic capability to handle complex MRIG tasks. In the second stage, we introduce Difficulty-aware Advantage Reweighting (DAR) and Discriminative Reward Scaling (DRS). DAR dynamically adjusts the optimization objective to improve performance when handling a large number of mixed-type reference images. DRS enlarges intra-group reward differences for more effective policy optimization. Experiments demonstrate that DyRef significantly improves the performance of open-source models on OmniRef-Bench and single-image editing benchmarks, demonstrating the effectiveness and generalization capability of our approach.
While instruction-based image editing, enabled by multi-modal generative models, has advanced significantly, existing benchmarks lack a comprehensive evaluation of physics-based reasoning, a critical capability for handling real-world scenarios. To address this, we introduce PhyEditBench, a benchmark designed to assess the physical understanding of editing models. Guided by a hierarchical taxonomy, we establish 4 primary classes and 12 subclasses. It comprises 238 high-quality, high-resolution, real-world instances meticulously extracted from videos to capture authentic physical dynamics, alongside 35 synthetic Anti-Physics instances. Our empirical analysis of current SOTA editing methods exposes substantial limitations in their physics-based reasoning. We further propose a training-free baseline named PhyWorld that uses test-time scaling and a latent reduction strategy. PhyWorld outperforms comparable models and suggests that the video generation process can effectively serve as a reasoning mechanism for image editing. The project page is available at https://github.com/Previsior/PhyEditBench.
展开完整摘要收起摘要↓
While instruction-based image editing, enabled by multi-modal generative models, has advanced significantly, existing benchmarks lack a comprehensive evaluation of physics-based reasoning, a critical capability for handling real-world scenarios. To address this, we introduce PhyEditBench, a benchmark designed to assess the physical understanding of editing models. Guided by a hierarchical taxonomy, we establish 4 primary classes and 12 subclasses. It comprises 238 high-quality, high-resolution, real-world instances meticulously extracted from videos to capture authentic physical dynamics, alongside 35 synthetic Anti-Physics instances. Our empirical analysis of current SOTA editing methods exposes substantial limitations in their physics-based reasoning. We further propose a training-free baseline named PhyWorld that uses test-time scaling and a latent reduction strategy. PhyWorld outperforms comparable models and suggests that the video generation process can effectively serve as a reasoning mechanism for image editing. The project page is available at https://github.com/Previsior/PhyEditBench.
Video customization based on Text-to-Video (T2V) models aims to learn specific features from reference data to generate controllable videos. While significant strides have been made in image stylization and video motion customization, simultaneously controlling multiple concepts, such as content, style, and motion, remains a major challenge. In this work, we systematically define the task of multi-concept video customization, which requires the joint control of content, style, and motion. To facilitate research in this area, we construct a comprehensive benchmark and propose Disco-LoRA, a unified framework designed to tackle this problem by disentangling and flexibly recombining different concepts in two stages: (1) We decompose the objective into two sub-tasks: Content-Style and Content-Motion. Each sub-task is addressed using our Iterative Dual-LoRA Disentanglement Framework, which effectively disentangles distinct concepts within the data. (2) We identify layer-wise weight trends as crucial for LoRA identity, while weight magnitudes dictate composability. To harmonize these scales, we propose a Z-score-based statistical regularization that aligns weight distributions, preserving layer-wise trends while minimizing interference between different LoRAs. Extensive experiments show that Disco-LoRA excels in multi-concept video customization, effectively preserving appearance, style, and motion for controllable text-to-video generation.
展开完整摘要收起摘要↓
Video customization based on Text-to-Video (T2V) models aims to learn specific features from reference data to generate controllable videos. While significant strides have been made in image stylization and video motion customization, simultaneously controlling multiple concepts, such as content, style, and motion, remains a major challenge. In this work, we systematically define the task of multi-concept video customization, which requires the joint control of content, style, and motion. To facilitate research in this area, we construct a comprehensive benchmark and propose Disco-LoRA, a unified framework designed to tackle this problem by disentangling and flexibly recombining different concepts in two stages: (1) We decompose the objective into two sub-tasks: Content-Style and Content-Motion. Each sub-task is addressed using our Iterative Dual-LoRA Disentanglement Framework, which effectively disentangles distinct concepts within the data. (2) We identify layer-wise weight trends as crucial for LoRA identity, while weight magnitudes dictate composability. To harmonize these scales, we propose a Z-score-based statistical regularization that aligns weight distributions, preserving layer-wise trends while minimizing interference between different LoRAs. Extensive experiments show that Disco-LoRA excels in multi-concept video customization, effectively preserving appearance, style, and motion for controllable text-to-video generation.
LLM agents are transitioning from experimental tools to permanent infrastructure — a computational layer as enduring as the electrical grid. Like any infrastructure, they carry a cost chain from physical capital through enterprise investment to user consumption, ending at the user's most irreplaceable resource: lifetime. When unoptimized, this chain leaks, consuming user lifetime without adequate compensation. This paper proposes Pingquanqi (Equalizer), a cross-domain sociotechnical framework for Human-Agent Interaction Governance (HAIGF). Its product form is an Agent framework-level embedded design specification, analogous to WCAG for web accessibility, whose goal is not to be purchased but adopted as a standard. Pingquanqi consists of four integrated components deployable as native middleware: (1) a user-state discrimination model enabling proactive knowledge leveling, (2) a Bayesian progressive stop-loss rule capping per-session interaction cost, (3) controlled friction mechanisms breaking self-reinforcing dependency loops, and (4) Lsteal, a transparency metric rendering token-to-lifetime cost conversion visible. A fifth mechanism, reflective summarization (F5), enables guided cognitive recollection. The framework is grounded in cross-cultural philosophy: Mao's epistemology of practice (On Practice, 1937) provides the basis for cross-session knowledge accumulation; Wang Yangming's unity of knowledge and action (zhi xing he yi, c. 1509) illuminates Lsteal's root — knowing without acting is incomplete; and Hegel's unity of theory and practice demonstrates cross-traditional convergence. This paper argues Pingquanqi's primary economic beneficiary is the enterprise deploying Agent services — through reduced wasted computation, improved user satisfaction, and sustained subscription revenue — with individual user benefit as the natural downstream consequence.
展开完整摘要收起摘要↓
LLM agents are transitioning from experimental tools to permanent infrastructure — a computational layer as enduring as the electrical grid. Like any infrastructure, they carry a cost chain from physical capital through enterprise investment to user consumption, ending at the user's most irreplaceable resource: lifetime. When unoptimized, this chain leaks, consuming user lifetime without adequate compensation. This paper proposes Pingquanqi (Equalizer), a cross-domain sociotechnical framework for Human-Agent Interaction Governance (HAIGF). Its product form is an Agent framework-level embedded design specification, analogous to WCAG for web accessibility, whose goal is not to be purchased but adopted as a standard. Pingquanqi consists of four integrated components deployable as native middleware: (1) a user-state discrimination model enabling proactive knowledge leveling, (2) a Bayesian progressive stop-loss rule capping per-session interaction cost, (3) controlled friction mechanisms breaking self-reinforcing dependency loops, and (4) Lsteal, a transparency metric rendering token-to-lifetime cost conversion visible. A fifth mechanism, reflective summarization (F5), enables guided cognitive recollection. The framework is grounded in cross-cultural philosophy: Mao's epistemology of practice (On Practice, 1937) provides the basis for cross-session knowledge accumulation; Wang Yangming's unity of knowledge and action (zhi xing he yi, c. 1509) illuminates Lsteal's root — knowing without acting is incomplete; and Hegel's unity of theory and practice demonstrates cross-traditional convergence. This paper argues Pingquanqi's primary economic beneficiary is the enterprise deploying Agent services — through reduced wasted computation, improved user satisfaction, and sustained subscription revenue — with individual user benefit as the natural downstream consequence.
Video Diffusion Models (VDMs) is constrained by immense computational costs. While offline calibration-based acceleration suffers from calibration data dependency, prohibitive calibration duration, and susceptibility to distribution shifts, offline calibration-free methods eliminate these hurdles. However, since they rely on instantaneous zero-order approximations where the mapping between input and output differences varies in real-time, they are susceptible to observational noise and ignore the intrinsic momentum within the diffusion trajectory. In this paper, we propose NaviCache, a plug-and-play test-time self-calibration method re-conceptualizing feature evolution as an Inertial Navigation System (INS) problem. NaviCache bridges the fundamental domain gap and the non-stationary nature of diffusion by modeling the relative coupling between input and output variations. We introduce a dual-state estimation architecture that adaptively tracks the feature change ratio and its latent drift, initialized via a specialized Initial Alignment phase. By integrating a time-dependent noise schedule with an uncertainty-aware Measurement Update mechanism, NaviCache provides a theoretically grounded mechanism for error-bounded computation skipping. Extensive experiments on the HunyuanVideo, Wan, and Open-Sora series demonstrate that NaviCache exhibits more accurate error judgment for computation skipping and achieves outstanding comprehensive performance.
展开完整摘要收起摘要↓
Video Diffusion Models (VDMs) is constrained by immense computational costs. While offline calibration-based acceleration suffers from calibration data dependency, prohibitive calibration duration, and susceptibility to distribution shifts, offline calibration-free methods eliminate these hurdles. However, since they rely on instantaneous zero-order approximations where the mapping between input and output differences varies in real-time, they are susceptible to observational noise and ignore the intrinsic momentum within the diffusion trajectory. In this paper, we propose NaviCache, a plug-and-play test-time self-calibration method re-conceptualizing feature evolution as an Inertial Navigation System (INS) problem. NaviCache bridges the fundamental domain gap and the non-stationary nature of diffusion by modeling the relative coupling between input and output variations. We introduce a dual-state estimation architecture that adaptively tracks the feature change ratio and its latent drift, initialized via a specialized Initial Alignment phase. By integrating a time-dependent noise schedule with an uncertainty-aware Measurement Update mechanism, NaviCache provides a theoretically grounded mechanism for error-bounded computation skipping. Extensive experiments on the HunyuanVideo, Wan, and Open-Sora series demonstrate that NaviCache exhibits more accurate error judgment for computation skipping and achieves outstanding comprehensive performance.
Learning governing equations from observed solution data is a fundamental challenge in scientific machine learning, yet the theoretical conditions under which a ground-truth ODE can be uniquely and stably identified from multiple solution observations remain largely undeveloped, and no quantitative analysis of the sample complexity of such learning tasks exists in the literature. To address this gap, we introduce the Hausdorff distance on solution sets as the natural metric for comparing differential equations, since it captures the worst-case separation between two equations over all admissible initial conditions and thus encodes the minimax structure of the identification problem. We establish identifiability bounds for governing ODEs across a wide class of structure equations--ranging from linear ODEs to nonlinear classes with Lipschitz (Hölder)-continuous vector fields--characterizing precisely when two distinct equations can be distinguished from solution data. Using this metric, we derive metric entropy estimates for the relevant ODE classes and analyze sample complexity bounds, quantifying how many solution observations are needed to reliably recover the governing equation.
展开完整摘要收起摘要↓
Learning governing equations from observed solution data is a fundamental challenge in scientific machine learning, yet the theoretical conditions under which a ground-truth ODE can be uniquely and stably identified from multiple solution observations remain largely undeveloped, and no quantitative analysis of the sample complexity of such learning tasks exists in the literature. To address this gap, we introduce the Hausdorff distance on solution sets as the natural metric for comparing differential equations, since it captures the worst-case separation between two equations over all admissible initial conditions and thus encodes the minimax structure of the identification problem. We establish identifiability bounds for governing ODEs across a wide class of structure equations--ranging from linear ODEs to nonlinear classes with Lipschitz (Hölder)-continuous vector fields--characterizing precisely when two distinct equations can be distinguished from solution data. Using this metric, we derive metric entropy estimates for the relevant ODE classes and analyze sample complexity bounds, quantifying how many solution observations are needed to reliably recover the governing equation.
Analog hardware platforms such as coupled oscillators and Analog Ising Machines naturally solve differential equations at a fraction of the energy cost of digital computation, making them attractive for low-power generative modeling, yet a fundamental mismatch exists: modern generative models assume flexible, software-defined dynamics, whereas analog hardware imposes fixed, physics-determined differential equations with limited approximation capacity. This paper introduces Analog Interaction Systems (AIS), a unified framework for hardware-implementable dynamical systems, and empirically characterizes their expressivity gap relative to neural network baselines. Two hardware-compatible mechanisms are proposed to narrow this gap - time-varying piecewise parameters and hidden physical states - and a Wasserstein GAN training procedure is developed to enable training of these models without requiring them to follow a specific trajectory. We characterize how area and power scale with connection density and precision, showing that sparse connectivity and low-bit-width quantized parameters are necessary for practical implementation, and estimate an energy cost of 23uJ per generated image for the chosen architecture, representing a 2-orders-of-magnitude improvement over digital baselines. On MNIST and Fashion-MNIST, our oscillator-based AIS achieves FID scores of 27.6 and 80.8, outperforming the best prior hardware-implementable analog generative models by 3-4x with a 4-bit sparse architecture.
展开完整摘要收起摘要↓
Analog hardware platforms such as coupled oscillators and Analog Ising Machines naturally solve differential equations at a fraction of the energy cost of digital computation, making them attractive for low-power generative modeling, yet a fundamental mismatch exists: modern generative models assume flexible, software-defined dynamics, whereas analog hardware imposes fixed, physics-determined differential equations with limited approximation capacity. This paper introduces Analog Interaction Systems (AIS), a unified framework for hardware-implementable dynamical systems, and empirically characterizes their expressivity gap relative to neural network baselines. Two hardware-compatible mechanisms are proposed to narrow this gap - time-varying piecewise parameters and hidden physical states - and a Wasserstein GAN training procedure is developed to enable training of these models without requiring them to follow a specific trajectory. We characterize how area and power scale with connection density and precision, showing that sparse connectivity and low-bit-width quantized parameters are necessary for practical implementation, and estimate an energy cost of 23uJ per generated image for the chosen architecture, representing a 2-orders-of-magnitude improvement over digital baselines. On MNIST and Fashion-MNIST, our oscillator-based AIS achieves FID scores of 27.6 and 80.8, outperforming the best prior hardware-implementable analog generative models by 3-4x with a 4-bit sparse architecture.
作者Gia-Cat Bui-Le, Tuong-Vy Truong-Thuy, Hai-Dang Nguyen, Trung-Nghia Le
In recent years, advancements in deep learning and generative models have revolutionized music-driven dance generation. This paper introduces a novel platform, namely DanceDuo, leveraging diffusion models to generate AI-choreographed dance sequences synchronized with a variety of music genres, to encourage dancing practice. The system allows users to interact with AI by selecting music tracks, humanoid models, and importing personal dance videos for comparison, fostering a rich and engaging user experience. DanceDuo not only offers dance generation but also integrates human pose estimation models to provide users with insightful comparisons of their own performances with AI-generated sequences. We conducted a comprehensive user study, revealing that users found the interface intuitive, with particular praise for the dance comparison feature. Our DanceDuo contributes significantly to the integration of AI in dance choreography, offering novel avenues for both recreational and professional applications.
展开完整摘要收起摘要↓
In recent years, advancements in deep learning and generative models have revolutionized music-driven dance generation. This paper introduces a novel platform, namely DanceDuo, leveraging diffusion models to generate AI-choreographed dance sequences synchronized with a variety of music genres, to encourage dancing practice. The system allows users to interact with AI by selecting music tracks, humanoid models, and importing personal dance videos for comparison, fostering a rich and engaging user experience. DanceDuo not only offers dance generation but also integrates human pose estimation models to provide users with insightful comparisons of their own performances with AI-generated sequences. We conducted a comprehensive user study, revealing that users found the interface intuitive, with particular praise for the dance comparison feature. Our DanceDuo contributes significantly to the integration of AI in dance choreography, offering novel avenues for both recreational and professional applications.
作者Johnny Peng, Thanh Tung Khuat, Ellen Otte, Katarzyna Musial, Bogdan Gabrys
Mammalian cell-culture processes underpin the manufacture of many biopharmaceuticals, yet keeping a run on track is hard: critical process parameters drift over days, and an off-specification trend is often confirmed too late to intervene. Early-stage, multi-day forecasts could enable timely adjustment of feeding, sampling, and control, but bioprocess forecasting is challenging because measurements are sparse and irregularly sampled, operating conditions are heterogeneous across cell lines and media, and runs with near-identical early behaviour can diverge into different futures. We propose an adaptive framework combining a Gated Bottleneck Latent Ordinary Differential Equation (GB-Latent ODE) with Multi-Path Just-In-Time Fine Tuning (MP-JIT-FT). The GB-Latent ODE augments the stan dard Latent ODE with learnable variable-wise gating and a mask-aware bottleneck that compress high-dimensional sparse inputs, improving learning under limited data. Given a partially observed run, MP-JIT-FT retrieves similar historical trajectories, clusters the local neighbourhood into candidate regimes, and fine-tunes a separate model per regime to produce multiple plausible paths, each with a reconstruction-based confidence score, not a single averaged forecast. We further fuse Raman spectroscopy data: a machine-learning soft sensor turns dense Raman spectra into pseudo-observations that enrich the sparse offline measurements for more robust training. On 38 fed-batch 5L bioreactor runs spanning 14 conditions, MP-JIT-FT with Raman fusion achieves the best average rank and outperforms a global Latent ODE baseline on 8 of 9 target variables. Using local-divergence metrics, we show the multi-path gains are largest when locally similar prefixes diverge, whereas Raman fusion helps most when early dynamics are representative of later behaviour.
展开完整摘要收起摘要↓
Mammalian cell-culture processes underpin the manufacture of many biopharmaceuticals, yet keeping a run on track is hard: critical process parameters drift over days, and an off-specification trend is often confirmed too late to intervene. Early-stage, multi-day forecasts could enable timely adjustment of feeding, sampling, and control, but bioprocess forecasting is challenging because measurements are sparse and irregularly sampled, operating conditions are heterogeneous across cell lines and media, and runs with near-identical early behaviour can diverge into different futures. We propose an adaptive framework combining a Gated Bottleneck Latent Ordinary Differential Equation (GB-Latent ODE) with Multi-Path Just-In-Time Fine Tuning (MP-JIT-FT). The GB-Latent ODE augments the stan dard Latent ODE with learnable variable-wise gating and a mask-aware bottleneck that compress high-dimensional sparse inputs, improving learning under limited data. Given a partially observed run, MP-JIT-FT retrieves similar historical trajectories, clusters the local neighbourhood into candidate regimes, and fine-tunes a separate model per regime to produce multiple plausible paths, each with a reconstruction-based confidence score, not a single averaged forecast. We further fuse Raman spectroscopy data: a machine-learning soft sensor turns dense Raman spectra into pseudo-observations that enrich the sparse offline measurements for more robust training. On 38 fed-batch 5L bioreactor runs spanning 14 conditions, MP-JIT-FT with Raman fusion achieves the best average rank and outperforms a global Latent ODE baseline on 8 of 9 target variables. Using local-divergence metrics, we show the multi-path gains are largest when locally similar prefixes diverge, whereas Raman fusion helps most when early dynamics are representative of later behaviour.
作者Xuyue Huang, Zhe Chen, Wang Shen, Xiao-Ping Zhang
Diffusion Transformers (DiTs) have driven substantial progress in image and video generation but suffer from prohibitive computational costs. Feature caching accelerates inference by reusing intermediate representations. Existing methods rely on historical features for implementation simplicity, yet suffer from severe error accumulation at high acceleration ratios. To address this limitation, we investigate the nature of the requisite feature correction. We demonstrate that the optimal calibration update is characterized by a shared low-rank subspace across diverse prompts. Guided by this structural insight, we propose LearniBridge, a learnable calibration mechanism for feature caching that bridges multiple timesteps through lightweight LoRA updates. This mechanism enables effective calibration requiring only 3-5 training samples. Extensive experiments on image and video generation show that LearniBridge achieves up to $5.87\times$, $5.75\times$, and $4.10\times$ acceleration on FLUX, HunyuanVideo, and WAN2.1, respectively. On WAN2.1, it improves VBench by 1.28% over the previous SOTA at $4.10\times$ acceleration. Our code is available at https://github.com/Iiiiiiirene/LearniBridge.
展开完整摘要收起摘要↓
Diffusion Transformers (DiTs) have driven substantial progress in image and video generation but suffer from prohibitive computational costs. Feature caching accelerates inference by reusing intermediate representations. Existing methods rely on historical features for implementation simplicity, yet suffer from severe error accumulation at high acceleration ratios. To address this limitation, we investigate the nature of the requisite feature correction. We demonstrate that the optimal calibration update is characterized by a shared low-rank subspace across diverse prompts. Guided by this structural insight, we propose LearniBridge, a learnable calibration mechanism for feature caching that bridges multiple timesteps through lightweight LoRA updates. This mechanism enables effective calibration requiring only 3-5 training samples. Extensive experiments on image and video generation show that LearniBridge achieves up to $5.87\times$, $5.75\times$, and $4.10\times$ acceleration on FLUX, HunyuanVideo, and WAN2.1, respectively. On WAN2.1, it improves VBench by 1.28% over the previous SOTA at $4.10\times$ acceleration. Our code is available at https://github.com/Iiiiiiirene/LearniBridge.
Mixture-of-Experts (MoE) architectures have emerged as a powerful paradigm for scaling diffusion models in visual generation. Recent advancements have focused on adaptively allocating computational resources across diverse tokens to improve efficiency and performance. However, we identify a routing assignment problem in existing diffusion MoE frameworks: the router fails to accurately allocate more computational resources to salient tokens. Our analysis attributes this failure to the router's reliance on noise-corrupted latent features throughout the denoising process. Such stochastic noise obscures the critical structural and textural information, thereby preventing the router from effectively distinguishing salient tokens. To address this, we propose SharpMoE, a post-training framework with a saliency-harnessing accurate routing mechanism, which utilizes clean latent features as a noise-free guidance signal for routing. By bypassing the noise-distorted inputs, SharpMoE provides the router with clear saliency guidance, enabling the identification of salient tokens even in high-noise stages. Furthermore, we introduce a trajectory routing loss to constrain the compute allocation throughout the multi-step denoising trajectory, ensuring precise resource allocation along the generation rollout. Extensive experiments demonstrate that SharpMoE serves as a versatile, plug-and-play solution that further enhances the pretrained, converged MoE models, achieving state-of-the-art performance in visual generation.
展开完整摘要收起摘要↓
Mixture-of-Experts (MoE) architectures have emerged as a powerful paradigm for scaling diffusion models in visual generation. Recent advancements have focused on adaptively allocating computational resources across diverse tokens to improve efficiency and performance. However, we identify a routing assignment problem in existing diffusion MoE frameworks: the router fails to accurately allocate more computational resources to salient tokens. Our analysis attributes this failure to the router's reliance on noise-corrupted latent features throughout the denoising process. Such stochastic noise obscures the critical structural and textural information, thereby preventing the router from effectively distinguishing salient tokens. To address this, we propose SharpMoE, a post-training framework with a saliency-harnessing accurate routing mechanism, which utilizes clean latent features as a noise-free guidance signal for routing. By bypassing the noise-distorted inputs, SharpMoE provides the router with clear saliency guidance, enabling the identification of salient tokens even in high-noise stages. Furthermore, we introduce a trajectory routing loss to constrain the compute allocation throughout the multi-step denoising trajectory, ensuring precise resource allocation along the generation rollout. Extensive experiments demonstrate that SharpMoE serves as a versatile, plug-and-play solution that further enhances the pretrained, converged MoE models, achieving state-of-the-art performance in visual generation.
The prevalent dual-branch paradigm, i.e., training a side network to encode visual conditions and fusing its intermediate-layer features to a frozen pretrained main network, has shown remarkable success in visual-condition controllable generation. Despite its widespread adoption, the role of the side branch and its training efficiency remain underexplored. In this paper, we first revisit this mainstream paradigm through the lens of score-based generative modeling: 1) The main network preserves visual perceptual quality by providing a prior unconditional score. 2) The side network steers conditional control by implicitly contributing a likelihood score. Guided by this perspective, we propose LIkelihood Score Alignment (LISA), an effective regularization method that explicitly aligns the intermediate feature of the side network with an approximated likelihood score. Specifically, we first hook features from a designated layer of the side network and project them into the score latent space by a lightweight decoder. Then, we construct an approximated likelihood score target and calculate the distance between the decoder's output and this target as an additional regularization loss. Finally, we jointly optimize the side network and decoder with both standard diffusion loss and our regularization loss. Experiments across various image/video tasks, architectures, and diffusion/flow models demonstrated that LISA can not only consistently accelerate the training convergence and improve final synthetic results, but also encourage the side network's features to be more disentangled for conditional modeling with negligible additional training cost and zero extra inference cost.
展开完整摘要收起摘要↓
The prevalent dual-branch paradigm, i.e., training a side network to encode visual conditions and fusing its intermediate-layer features to a frozen pretrained main network, has shown remarkable success in visual-condition controllable generation. Despite its widespread adoption, the role of the side branch and its training efficiency remain underexplored. In this paper, we first revisit this mainstream paradigm through the lens of score-based generative modeling: 1) The main network preserves visual perceptual quality by providing a prior unconditional score. 2) The side network steers conditional control by implicitly contributing a likelihood score. Guided by this perspective, we propose LIkelihood Score Alignment (LISA), an effective regularization method that explicitly aligns the intermediate feature of the side network with an approximated likelihood score. Specifically, we first hook features from a designated layer of the side network and project them into the score latent space by a lightweight decoder. Then, we construct an approximated likelihood score target and calculate the distance between the decoder's output and this target as an additional regularization loss. Finally, we jointly optimize the side network and decoder with both standard diffusion loss and our regularization loss. Experiments across various image/video tasks, architectures, and diffusion/flow models demonstrated that LISA can not only consistently accelerate the training convergence and improve final synthetic results, but also encourage the side network's features to be more disentangled for conditional modeling with negligible additional training cost and zero extra inference cost.
Graph Neural Differential Equations (GNDEs) model continuous-time graph dynamics by parameterizing Neural ODE velocity fields with Graph Neural Networks. Their local, size-independent filters suggest a zero-shot size-transfer principle: train on a small graph and deploy on larger, similar graphs without retraining. We develop a quantitative theory for this principle on sparse random graphs sampled from graphons. We consider Graphon Neural Differential Equations (Graphon-NDEs) and adjoint Graphon-NDEs as the infinite-node limits of the forward and adjoint GNDE systems, and establish well-posedness. For an $n$-node random graph with sparsity parameter $α_n$, we prove trajectory-wise convergence of GNDE solutions to Graphon-NDE solutions at rate $O((α_n n)^{-1/2})$, up to logarithmic factors, with high probability. We also establish uniform-in-time convergence bounds for adjoint systems governing hidden-state and parameter gradients. We further study discretize-then-optimize (DTO) and optimize-then-discretize (OTD) training. Under explicit Euler discretization with $M$ steps, we show that DTO and OTD are asymptotically consistent, with hidden-state and local parameter-gradient discrepancies of orders $O(1/M)$ and $O(1/M^2)$, respectively, up to sparsity and logarithmic factors. Experiments on HSBM and tent graphons support the theoretical rates, while zero-shot transfer experiments across four graphon classes demonstrate accurate deployment of learned GNDEs on larger independently sampled graphs.
展开完整摘要收起摘要↓
Graph Neural Differential Equations (GNDEs) model continuous-time graph dynamics by parameterizing Neural ODE velocity fields with Graph Neural Networks. Their local, size-independent filters suggest a zero-shot size-transfer principle: train on a small graph and deploy on larger, similar graphs without retraining. We develop a quantitative theory for this principle on sparse random graphs sampled from graphons. We consider Graphon Neural Differential Equations (Graphon-NDEs) and adjoint Graphon-NDEs as the infinite-node limits of the forward and adjoint GNDE systems, and establish well-posedness. For an $n$-node random graph with sparsity parameter $α_n$, we prove trajectory-wise convergence of GNDE solutions to Graphon-NDE solutions at rate $O((α_n n)^{-1/2})$, up to logarithmic factors, with high probability. We also establish uniform-in-time convergence bounds for adjoint systems governing hidden-state and parameter gradients. We further study discretize-then-optimize (DTO) and optimize-then-discretize (OTD) training. Under explicit Euler discretization with $M$ steps, we show that DTO and OTD are asymptotically consistent, with hidden-state and local parameter-gradient discrepancies of orders $O(1/M)$ and $O(1/M^2)$, respectively, up to sparsity and logarithmic factors. Experiments on HSBM and tent graphons support the theoretical rates, while zero-shot transfer experiments across four graphon classes demonstrate accurate deployment of learned GNDEs on larger independently sampled graphs.
We propose latent-space diffusion posterior sampling (L-DPS), an approximate Bayesian framework for high-dimensional inverse problems governed by partial differential equations (PDEs). The method addresses three challenges in PDE-constrained inversion: implicit sample-based priors without tractable densities, high-dimensional spatially distributed parameters, and the high cost of repeated forward-model evaluations during posterior sampling. L-DPS combines a variational autoencoder, an unconditional latent diffusion model, diffusion posterior sampling, and a differentiable neural surrogate. The VAE maps the parameter field to a lower-dimensional latent space, the diffusion model learns an implicit prior score in this latent space, and DPS combines this learned prior with likelihood-based guidance. The likelihood gradient is evaluated through the decoder-surrogate composition, avoiding repeated calls to the full numerical PDE solver. We evaluate the method on an inverse Darcy flow problem with an unknown spatially distributed permeability field inferred from sparse and noisy pressure observations. L-DPS produces accurate and robust inverse solutions, reduces inference cost relative to full-space DPS, and outperforms amortized inverse baselines such as conditional latent diffusion and inverse FNO in sparse and noisy regimes. We further compare L-DPS with a KLE-MAP baseline and study mixed-prior generalization and the sensitivity of inversion accuracy to surrogate forward-model error.
展开完整摘要收起摘要↓
We propose latent-space diffusion posterior sampling (L-DPS), an approximate Bayesian framework for high-dimensional inverse problems governed by partial differential equations (PDEs). The method addresses three challenges in PDE-constrained inversion: implicit sample-based priors without tractable densities, high-dimensional spatially distributed parameters, and the high cost of repeated forward-model evaluations during posterior sampling. L-DPS combines a variational autoencoder, an unconditional latent diffusion model, diffusion posterior sampling, and a differentiable neural surrogate. The VAE maps the parameter field to a lower-dimensional latent space, the diffusion model learns an implicit prior score in this latent space, and DPS combines this learned prior with likelihood-based guidance. The likelihood gradient is evaluated through the decoder-surrogate composition, avoiding repeated calls to the full numerical PDE solver. We evaluate the method on an inverse Darcy flow problem with an unknown spatially distributed permeability field inferred from sparse and noisy pressure observations. L-DPS produces accurate and robust inverse solutions, reduces inference cost relative to full-space DPS, and outperforms amortized inverse baselines such as conditional latent diffusion and inverse FNO in sparse and noisy regimes. We further compare L-DPS with a KLE-MAP baseline and study mixed-prior generalization and the sensitivity of inversion accuracy to surrogate forward-model error.
Agents often repeatedly solve similar task instances from scratch, leading to unnecessary reasoning cost and long execution traces. Prior work has explored workflow reuse and executable skill induction, but it remains unclear which task scenarios admit procedural skills and how the shared procedural structure should be represented across successful traces. We study this problem in FSM-defined scenarios, where successful traces can be viewed as paths in an unknown transition graph, and formulate procedural skills as reusable parameterized control-flow subgraphs. Based on this view, we introduce SkillDisCo, a distillation-and-compilation framework that distills reusable PFSM subgraphs from successful traces and compiles them into callable, executable, and verifiable procedural skills. Experiments on ALFWorld and WebArena show that SkillDisCo improves success rates and reduces agent turns across benchmarks and model scales, demonstrating the benefits of representing shared experience as reusable execution structures.
展开完整摘要收起摘要↓
Agents often repeatedly solve similar task instances from scratch, leading to unnecessary reasoning cost and long execution traces. Prior work has explored workflow reuse and executable skill induction, but it remains unclear which task scenarios admit procedural skills and how the shared procedural structure should be represented across successful traces. We study this problem in FSM-defined scenarios, where successful traces can be viewed as paths in an unknown transition graph, and formulate procedural skills as reusable parameterized control-flow subgraphs. Based on this view, we introduce SkillDisCo, a distillation-and-compilation framework that distills reusable PFSM subgraphs from successful traces and compiles them into callable, executable, and verifiable procedural skills. Experiments on ALFWorld and WebArena show that SkillDisCo improves success rates and reduces agent turns across benchmarks and model scales, demonstrating the benefits of representing shared experience as reusable execution structures.
LLM-based code agents navigate repositories through keyword search but miss the structural relationships, such as call graphs, inheritance hierarchies, and configuration dependencies, that define how software actually works. This makes agent navigation stochastic and difficult to reproduce across runs. We investigate whether lightweight static analysis can provide deterministic anchors for these agents: stable structural facts injected as plain-text comments that constrain probabilistic exploration and make navigation more predictable. Starting from a strong baseline, Codex from OpenAI, we systematically inject varying granularities of structural annotations and measure their effects on localization, trajectory behavior, and run-to-run stability. Our study identifies what we call the deterministic anchoring effect: static structure helps less by making agents "smarter" and more by making their navigation disciplined and reproducible. Three observations support this finding: (1) Anchoring works: lightweight call/inheritance topology improves function-level localization (+2.2pp Func@5) and shortens trajectories (-1.6 interaction rounds); (2) Anchoring is scale-sensitive: the optimal granularity and directionality depend on repository characteristics, where denser semantics show diminishing returns and hub-heavy projects benefit from inverse-only links that expose "who-calls-me" without forward edges; (3) Anchoring stabilizes: tags raise link-following rate from 0.15-0.18 to 0.21-0.24, roughly halve run-to-run variance, and improve single-run reliability (Pass@1 +3.4 pp) on medium-scale repositories, at the cost of roughly 10% more input tokens. These observations suggest practical guidelines: default to lightweight topology on medium projects, prune forward edges in large repositories, and reserve dense tags for implicit-dependency cases.
展开完整摘要收起摘要↓
LLM-based code agents navigate repositories through keyword search but miss the structural relationships, such as call graphs, inheritance hierarchies, and configuration dependencies, that define how software actually works. This makes agent navigation stochastic and difficult to reproduce across runs. We investigate whether lightweight static analysis can provide deterministic anchors for these agents: stable structural facts injected as plain-text comments that constrain probabilistic exploration and make navigation more predictable. Starting from a strong baseline, Codex from OpenAI, we systematically inject varying granularities of structural annotations and measure their effects on localization, trajectory behavior, and run-to-run stability. Our study identifies what we call the deterministic anchoring effect: static structure helps less by making agents "smarter" and more by making their navigation disciplined and reproducible. Three observations support this finding: (1) Anchoring works: lightweight call/inheritance topology improves function-level localization (+2.2pp Func@5) and shortens trajectories (-1.6 interaction rounds); (2) Anchoring is scale-sensitive: the optimal granularity and directionality depend on repository characteristics, where denser semantics show diminishing returns and hub-heavy projects benefit from inverse-only links that expose "who-calls-me" without forward edges; (3) Anchoring stabilizes: tags raise link-following rate from 0.15-0.18 to 0.21-0.24, roughly halve run-to-run variance, and improve single-run reliability (Pass@1 +3.4 pp) on medium-scale repositories, at the cost of roughly 10% more input tokens. These observations suggest practical guidelines: default to lightweight topology on medium projects, prune forward edges in large repositories, and reserve dense tags for implicit-dependency cases.
作者Zhengyuan Liu, Stella Xin Yin, Min-Yen Kan, Nancy F. Chen
We present a conceptual framework for analyzing dialogue in collaborative problem-solving contexts, with an emphasis on the emerging dynamics of human-AI and multi-agent collaboration. As intelligent systems become active agents capable of autonomous reasoning and strategic cooperation, understanding the dialogic interaction during collaborative problem solving is increasingly important for optimizing and evaluating such partnerships. Our framework addresses key limitations in current analytical approaches through a hierarchical two-layer coding scheme that integrates cognitive and non-cognitive problem solving with metacognitive regulatory mechanisms. We demonstrate its effectiveness and generalizability across nine datasets spanning multiple domains, and provide insights into how humans and agents coordinate their knowledge, skills, and efforts to solve complex problems, showing in particular that metacognitive regulation can be an essential discriminator of deeper collaboration.
展开完整摘要收起摘要↓
We present a conceptual framework for analyzing dialogue in collaborative problem-solving contexts, with an emphasis on the emerging dynamics of human-AI and multi-agent collaboration. As intelligent systems become active agents capable of autonomous reasoning and strategic cooperation, understanding the dialogic interaction during collaborative problem solving is increasingly important for optimizing and evaluating such partnerships. Our framework addresses key limitations in current analytical approaches through a hierarchical two-layer coding scheme that integrates cognitive and non-cognitive problem solving with metacognitive regulatory mechanisms. We demonstrate its effectiveness and generalizability across nine datasets spanning multiple domains, and provide insights into how humans and agents coordinate their knowledge, skills, and efforts to solve complex problems, showing in particular that metacognitive regulation can be an essential discriminator of deeper collaboration.
Personalized language-model assistants are often evaluated through a memory lens: can a model recall preferences users have explicitly stated in dialogue? More comprehensive personalization demands a harder capability — inferring what users care about from the multimodal traces they naturally leave behind. We introduce SocialPersona, a benchmark for evaluating whether multimodal large language models (MLLMs) can recover revealed preferences from longitudinal social-media timelines and use them in dialogue. Built from longitudinal timelines of 171 everyday, non-promotional social-media users, SocialPersona contains text, images, timestamps, and 2,597 human-verified preference tags across seven interest domains, separating stable interests from recent interests. It supports two tasks: constructing structured user profiles from multimodal context and generating responses aligned with inferred profiles. Experiments with proprietary and open-weight MLLMs show that models can identify broad interest domains, yet their performance drops on fine-grained and recent interests and degrades further when inferred profiles must be used to personalize dialogue. Together with evidence that text and images provide complementary preference signals, these results indicate that robust cross-modal, long-horizon user modeling remains a key challenge, and that SocialPersona can help measure and advance progress toward assistants that infer and act on revealed preferences.
展开完整摘要收起摘要↓
Personalized language-model assistants are often evaluated through a memory lens: can a model recall preferences users have explicitly stated in dialogue? More comprehensive personalization demands a harder capability — inferring what users care about from the multimodal traces they naturally leave behind. We introduce SocialPersona, a benchmark for evaluating whether multimodal large language models (MLLMs) can recover revealed preferences from longitudinal social-media timelines and use them in dialogue. Built from longitudinal timelines of 171 everyday, non-promotional social-media users, SocialPersona contains text, images, timestamps, and 2,597 human-verified preference tags across seven interest domains, separating stable interests from recent interests. It supports two tasks: constructing structured user profiles from multimodal context and generating responses aligned with inferred profiles. Experiments with proprietary and open-weight MLLMs show that models can identify broad interest domains, yet their performance drops on fine-grained and recent interests and degrades further when inferred profiles must be used to personalize dialogue. Together with evidence that text and images provide complementary preference signals, these results indicate that robust cross-modal, long-horizon user modeling remains a key challenge, and that SocialPersona can help measure and advance progress toward assistants that infer and act on revealed preferences.
Large language models (LLMs) can make scientific software easier to use. However, a general model does not automatically know which measurements a particular sensor can support, which algorithms are implemented in the current software, or which conclusions are justified by a computed result. These distinctions are especially important for low-channel electroencephalography (EEG), where sparse spatial coverage and variable signal quality make plausible but unsupported interpretations easy to produce. We present NeuraDock Agent, an open-source architecture that separates a deterministic local EEG engine from a hardware-aware language layer. The numerical engine parses recordings, performs quality control, executes reviewed spectral workflows, and writes machine-readable artifacts. The LLM receives only a compact, allowlisted summary and a versioned context pack. The context describes the seven-channel hardware, reviewed workflows, result fields, implementation boundaries, scientific limits, and reference cases. Raw EEG and dense per-sample arrays remain local We evaluate the system at three levels. First, 12 recordings produced identical structured results over ten numerical repetitions, and a complete Rest/Task run produced identical result, report, and figure hashes over three repetitions. Second, request-capture and failure-injection experiments confirmed the tested data boundary and preservation of local artifacts under HTTP, malformed-output, and connection failures. Third, a boundary-awareness benchmark tested 36 ordinary and adversarial questions under four context ablations and two LLMs, yielding 288 outputs.These results support hardware- and implementation-aware grounding as a practical mechanism for calibrating what an EEG agent accepts, qualifies, or refuses; they do not establish clinical validity or a validated absolute cognitive-load index.
展开完整摘要收起摘要↓
Large language models (LLMs) can make scientific software easier to use. However, a general model does not automatically know which measurements a particular sensor can support, which algorithms are implemented in the current software, or which conclusions are justified by a computed result. These distinctions are especially important for low-channel electroencephalography (EEG), where sparse spatial coverage and variable signal quality make plausible but unsupported interpretations easy to produce. We present NeuraDock Agent, an open-source architecture that separates a deterministic local EEG engine from a hardware-aware language layer. The numerical engine parses recordings, performs quality control, executes reviewed spectral workflows, and writes machine-readable artifacts. The LLM receives only a compact, allowlisted summary and a versioned context pack. The context describes the seven-channel hardware, reviewed workflows, result fields, implementation boundaries, scientific limits, and reference cases. Raw EEG and dense per-sample arrays remain local We evaluate the system at three levels. First, 12 recordings produced identical structured results over ten numerical repetitions, and a complete Rest/Task run produced identical result, report, and figure hashes over three repetitions. Second, request-capture and failure-injection experiments confirmed the tested data boundary and preservation of local artifacts under HTTP, malformed-output, and connection failures. Third, a boundary-awareness benchmark tested 36 ordinary and adversarial questions under four context ablations and two LLMs, yielding 288 outputs.These results support hardware- and implementation-aware grounding as a practical mechanism for calibrating what an EEG agent accepts, qualifies, or refuses; they do not establish clinical validity or a validated absolute cognitive-load index.
作者Yulin Mao, Shiyu Li, Shuping Song, Yuling Zhang, Yajun Song
We present MedSWFlow, an open-source, model-agnostic LLM workflow for drafting medical social work case plans. The framework translates professional case-planning tasks into six stages: assessment, problem analysis, goal setting, intervention planning, risk anticipation, and planned effect evaluation. Drawing on established social work and behavioral frameworks, MedSWFlow standardizes case inputs, builds structured case profiles, and generates reviewable assessment forms and service plans through staged prompting. The system is released as an open-source research framework for reproducible case-plan generation across LLM providers. Outputs are intended as practitioner-reviewed drafts rather than final service decisions. Source code: https://github.com/santhiyacw-droid/MedSWFlow/tree/main.
展开完整摘要收起摘要↓
We present MedSWFlow, an open-source, model-agnostic LLM workflow for drafting medical social work case plans. The framework translates professional case-planning tasks into six stages: assessment, problem analysis, goal setting, intervention planning, risk anticipation, and planned effect evaluation. Drawing on established social work and behavioral frameworks, MedSWFlow standardizes case inputs, builds structured case profiles, and generates reviewable assessment forms and service plans through staged prompting. The system is released as an open-source research framework for reproducible case-plan generation across LLM providers. Outputs are intended as practitioner-reviewed drafts rather than final service decisions. Source code: https://github.com/santhiyacw-droid/MedSWFlow/tree/main.
作者Hashmat Shadab Malik, Anees Ur Rehman Hashmi, Numan Saeed, Muzammal Naseer, Salman Khan, Christoph Lippert
Reasoning in multimodal large language models (MLLMs) has shown strong promise in medical imaging. However, this reasoning is usually free-form text judged only by its final answer, making it hard to interpret and verify, especially in 3D radiology, where a diagnosis should be traceable to evidence in the scan. Existing chest CT question-answering datasets compound this by reducing expert radiology reports to answer-only pairs, dropping the reasoning that links findings to conclusions and omitting the patient history clinicians rely on. As a result, reasoning-capable 3D chest CT MLLMs remain out of reach, as neither the structured supervision needed to train them nor the protocol needed to verify their reasoning yet exists. We introduce CORTEX (Clinically Organized Reasoning and sTructured EXplanation), a structured reasoning benchmark for 3D chest CT. For each question, CORTEX restores the missing reasoning as a four-stage diagnostic trace mirroring a radiologist's workflow: task understanding, visual observation, diagnostic reasoning, and answer synthesis. We generate these traces using frontier large language models with broad medical and general-domain knowledge, then filter and verify them with a stage-level evaluation protocol combining automated rubric scoring with expert radiologist review. Crucially, both the reasoning structure and evaluation rubrics are designed in close collaboration with clinicians. Built on CT-RATE, a large, publicly available chest CT dataset without reasoning annotations, CORTEX comprises 76,177 validated reasoning traces across open-ended VQA, closed-ended VQA, and report generation, providing both the structured supervision and the stage-level evaluation protocol needed to build and evaluate trustworthy reasoning models for 3D chest CT. Our dataset and evaluation code will be made publicly available upon acceptance.
展开完整摘要收起摘要↓
Reasoning in multimodal large language models (MLLMs) has shown strong promise in medical imaging. However, this reasoning is usually free-form text judged only by its final answer, making it hard to interpret and verify, especially in 3D radiology, where a diagnosis should be traceable to evidence in the scan. Existing chest CT question-answering datasets compound this by reducing expert radiology reports to answer-only pairs, dropping the reasoning that links findings to conclusions and omitting the patient history clinicians rely on. As a result, reasoning-capable 3D chest CT MLLMs remain out of reach, as neither the structured supervision needed to train them nor the protocol needed to verify their reasoning yet exists. We introduce CORTEX (Clinically Organized Reasoning and sTructured EXplanation), a structured reasoning benchmark for 3D chest CT. For each question, CORTEX restores the missing reasoning as a four-stage diagnostic trace mirroring a radiologist's workflow: task understanding, visual observation, diagnostic reasoning, and answer synthesis. We generate these traces using frontier large language models with broad medical and general-domain knowledge, then filter and verify them with a stage-level evaluation protocol combining automated rubric scoring with expert radiologist review. Crucially, both the reasoning structure and evaluation rubrics are designed in close collaboration with clinicians. Built on CT-RATE, a large, publicly available chest CT dataset without reasoning annotations, CORTEX comprises 76,177 validated reasoning traces across open-ended VQA, closed-ended VQA, and report generation, providing both the structured supervision and the stage-level evaluation protocol needed to build and evaluate trustworthy reasoning models for 3D chest CT. Our dataset and evaluation code will be made publicly available upon acceptance.
Retrieval-augmented generation (RAG) gives agents access to accumulated knowledge, but has no model of time. When a fact changes (e.g., a function is renamed or API restructured), RAG retrieves both the stale and current value with near-identical embedding similarity. The agent then either abstains or serves the superseded fact. We show this is a structural problem: on a calibrated dataset, cosine similarity distinguishes a contradicted fact from a duplicated one with AUROC 0.59 (near chance), as contradictions are often more embedding-similar to the original than rephrased duplicates. We present MemStrata, a retrieval memory maintaining temporal validity. It stores facts like RAG, preserving static recall, but when a fact's value is contradicted, a deterministic (subject, relation, object) supersession rule retires the stale value in a bi-temporal ledger - with no similarity threshold and no LLM call. Across six benchmarks run locally with a 7B model, MemStrata ties RAG on static knowledge and reaches 0.95-1.00 accuracy on evolving knowledge (where RAG reaches 0.20-0.47). The central result is the stale-fact-error rate: when required to answer, RAG serves superseded values 15-40% of the time; MemStrata drives this to ~0%, a failure class RAG cannot avoid. MemStrata achieves this at retrieval latency (~2.1s) versus ~16-18s for LLM-reranking baselines. We release the harness, datasets, and a marker-free evaluation protocol for memory under knowledge evolution.
展开完整摘要收起摘要↓
Retrieval-augmented generation (RAG) gives agents access to accumulated knowledge, but has no model of time. When a fact changes (e.g., a function is renamed or API restructured), RAG retrieves both the stale and current value with near-identical embedding similarity. The agent then either abstains or serves the superseded fact. We show this is a structural problem: on a calibrated dataset, cosine similarity distinguishes a contradicted fact from a duplicated one with AUROC 0.59 (near chance), as contradictions are often more embedding-similar to the original than rephrased duplicates. We present MemStrata, a retrieval memory maintaining temporal validity. It stores facts like RAG, preserving static recall, but when a fact's value is contradicted, a deterministic (subject, relation, object) supersession rule retires the stale value in a bi-temporal ledger - with no similarity threshold and no LLM call. Across six benchmarks run locally with a 7B model, MemStrata ties RAG on static knowledge and reaches 0.95-1.00 accuracy on evolving knowledge (where RAG reaches 0.20-0.47). The central result is the stale-fact-error rate: when required to answer, RAG serves superseded values 15-40% of the time; MemStrata drives this to ~0%, a failure class RAG cannot avoid. MemStrata achieves this at retrieval latency (~2.1s) versus ~16-18s for LLM-reranking baselines. We release the harness, datasets, and a marker-free evaluation protocol for memory under knowledge evolution.
Industrial advertising recommender models are continuously improved through architecture evolution. Upgrades such as RankMixer, TokenMixer-Large, and MixFormer show that better structures remain a key source of quality and business gains. Yet developing such upgrades in production is expert-intensive and difficult to scale. Existing automation is insufficient: AutoML mainly tunes hyper-parameters, while effective gains often require cross-module changes under strict constraints; generic LLM coding agents optimize for runnable code, but runnable code does not imply a valid recommender architecture. Candidates may pass local tests while causing silent failures that degrade performance. We present NOVA, a level-aware agent harness for verification-aware architecture evolution. NOVA uses an architecture gradient, an SGD-inspired, non-differentiable update signal that aggregates prior modifications, verification diagnostics, metric feedback, and trajectory memory to guide the next modification. A verification cascade checks structure semantics, local executability, offline effectiveness, and online impact; invalid candidates are blocked early, with failure patterns recorded as forbidden directions. L1--L4 task-level control matches automation to task complexity and risk, routing high-risk tasks to Copilot for human oversight. Deployed in an industrial advertising system, NOVA achieves the highest effective pass rate on L2 ScaleUp and L3 Literature-to-Production tasks (54.5% and 60.0%), reduces silent failures compared with coding-agent baselines, and shortens one literature-to-production cycle by over 13x in human-attended time. In online A/B testing, the selected L3 candidate improves GMV on three pCVR objectives by +1.25%, +1.70%, and +2.02%, while reducing pCVR bias by 58.8%, 66.7%, and 37.3%.
展开完整摘要收起摘要↓
Industrial advertising recommender models are continuously improved through architecture evolution. Upgrades such as RankMixer, TokenMixer-Large, and MixFormer show that better structures remain a key source of quality and business gains. Yet developing such upgrades in production is expert-intensive and difficult to scale. Existing automation is insufficient: AutoML mainly tunes hyper-parameters, while effective gains often require cross-module changes under strict constraints; generic LLM coding agents optimize for runnable code, but runnable code does not imply a valid recommender architecture. Candidates may pass local tests while causing silent failures that degrade performance. We present NOVA, a level-aware agent harness for verification-aware architecture evolution. NOVA uses an architecture gradient, an SGD-inspired, non-differentiable update signal that aggregates prior modifications, verification diagnostics, metric feedback, and trajectory memory to guide the next modification. A verification cascade checks structure semantics, local executability, offline effectiveness, and online impact; invalid candidates are blocked early, with failure patterns recorded as forbidden directions. L1--L4 task-level control matches automation to task complexity and risk, routing high-risk tasks to Copilot for human oversight. Deployed in an industrial advertising system, NOVA achieves the highest effective pass rate on L2 ScaleUp and L3 Literature-to-Production tasks (54.5% and 60.0%), reduces silent failures compared with coding-agent baselines, and shortens one literature-to-production cycle by over 13x in human-attended time. In online A/B testing, the selected L3 candidate improves GMV on three pCVR objectives by +1.25%, +1.70%, and +2.02%, while reducing pCVR bias by 58.8%, 66.7%, and 37.3%.
Recent work (2024 to 2026) has converged on a strategy for defending tool-using LLM agents against indirect prompt injection: rather than training the model to refuse malicious instructions, enforce security outside the model with a deterministic policy that mediates the agent's actions. Systems such as CaMeL, FIDES, Progent, RTBAS, and FORGE realize this with capabilities, information-flow labels, and reference monitors, and several report near-elimination of attacks on the AgentDojo benchmark. We make two contributions. First, we organize these out-of-band defenses as instances of classical integrity protection (Biba), reference monitoring, and least privilege, yielding a structured comparison of what they do and do not cover. Second, we warn that every one of them is validated only on static benchmarks (a fixed set of injection attempts), the same methodology that made in-band defenses look strong until adaptive, defense-aware attacks broke twelve of them at over 90% success; we specify the threat model and protocol an adaptive evaluation requires. We then run that protocol as an independent reproduction and extension of Progent's own adaptive-attack analysis, on AgentDojo, with an open-weight agent (Qwen2.5-7B) self-hosted on a single H200, a setting its authors did not test. Averaged over three runs, the defense held: Progent cut mean attack success roughly sixfold (25.8% to 4.2%), and a hand-crafted adaptive attack did not raise it (2.6%). This is one small-scale data point on a weak model with a single black-box attack template; a stronger optimized (white-box GCG) attack remains open. The result is consistent with, but does not establish, the hypothesis that deterministic out-of-band enforcement is a harder target for an adaptive attacker than in-band detection.
展开完整摘要收起摘要↓
Recent work (2024 to 2026) has converged on a strategy for defending tool-using LLM agents against indirect prompt injection: rather than training the model to refuse malicious instructions, enforce security outside the model with a deterministic policy that mediates the agent's actions. Systems such as CaMeL, FIDES, Progent, RTBAS, and FORGE realize this with capabilities, information-flow labels, and reference monitors, and several report near-elimination of attacks on the AgentDojo benchmark. We make two contributions. First, we organize these out-of-band defenses as instances of classical integrity protection (Biba), reference monitoring, and least privilege, yielding a structured comparison of what they do and do not cover. Second, we warn that every one of them is validated only on static benchmarks (a fixed set of injection attempts), the same methodology that made in-band defenses look strong until adaptive, defense-aware attacks broke twelve of them at over 90% success; we specify the threat model and protocol an adaptive evaluation requires. We then run that protocol as an independent reproduction and extension of Progent's own adaptive-attack analysis, on AgentDojo, with an open-weight agent (Qwen2.5-7B) self-hosted on a single H200, a setting its authors did not test. Averaged over three runs, the defense held: Progent cut mean attack success roughly sixfold (25.8% to 4.2%), and a hand-crafted adaptive attack did not raise it (2.6%). This is one small-scale data point on a weak model with a single black-box attack template; a stronger optimized (white-box GCG) attack remains open. The result is consistent with, but does not establish, the hypothesis that deterministic out-of-band enforcement is a harder target for an adaptive attacker than in-band detection.