Learned particle simulators based on graph neural networks achieve strong one-step accuracy, but errors compound over long horizons. An underexplored variable is the interaction graph: existing methods fix its topology via k-nearest neighbors or a static radius rule, regardless of local model confidence. We propose making this graph adaptive: a per-particle variance head, trained jointly with the acceleration head under a heteroscedastic Gaussian NLL loss, drives a trajectory in which high-uncertainty particles receive an expanded neighborhood. This is done at little extra inference cost by using the previous step's uncertainty estimate. A key discovery is that the variance head learns a meaningful notion of uncertainty: high-variance particles concentrate near complex regions, such as splash zones or free surfaces. When this signal drives graph topology, the resulting AdaptGNS simulator achieves a strict Pareto improvement on WaterDrop and a modest gain on Sand. Given the model's stronger performance on WaterDrop, we hypothesize that adaptive graphs are most useful when complexity is concentrated in space. Our code can be found at https://github.com/aidenzhou8/AdaptGNS.
展开完整摘要收起摘要↓
Learned particle simulators based on graph neural networks achieve strong one-step accuracy, but errors compound over long horizons. An underexplored variable is the interaction graph: existing methods fix its topology via k-nearest neighbors or a static radius rule, regardless of local model confidence. We propose making this graph adaptive: a per-particle variance head, trained jointly with the acceleration head under a heteroscedastic Gaussian NLL loss, drives a trajectory in which high-uncertainty particles receive an expanded neighborhood. This is done at little extra inference cost by using the previous step's uncertainty estimate. A key discovery is that the variance head learns a meaningful notion of uncertainty: high-variance particles concentrate near complex regions, such as splash zones or free surfaces. When this signal drives graph topology, the resulting AdaptGNS simulator achieves a strict Pareto improvement on WaterDrop and a modest gain on Sand. Given the model's stronger performance on WaterDrop, we hypothesize that adaptive graphs are most useful when complexity is concentrated in space. Our code can be found at https://github.com/aidenzhou8/AdaptGNS.
作者Shangjian Yin, Zehao Zhao, Kavosh Asadi, Rui Liu, Yuchen Lu, Shike Mei, Hang Cui, Luke Simon, Zhouxing Shi, Hamed Firooz
On-policy distillation (OPD) trains a student model by having it generate trajectories, then matching its next-token predictions with an external teacher's next-token predictions. This provides dense, token-level supervision to the student. On-policy self-distillation (OPSD) eliminates the need for the external teacher. Specifically, a second frozen copy of the student model, now given the ground truth in its context, serves as the teacher. The student model only receives the problem and learns to mimic the privileged teacher model, while the teacher remains frozen throughout training. Previous work showed that freezing the teacher is useful for training stability, but we argue that this can prevent the teacher from incorporating the improvements learned by the student during training. Our primary contribution is to address this limitation with a recursive framework built around two complementary components. First, we let the privileged teacher co-evolve with the student so that revision learned in one round can guide the next, a process we refer to as Dynamic Co-Evolution (DCE). Second, because stronger revision can also make responses too verbose and self-critical, we additionally train on shorter, verified rewrites of the model's own on-policy responses. We call this complementary objective Self-Refined Concise Learning (SRCL). Overall, our comprehensive evaluations show that DCE+SRCL outperforms OPSD across multiple model scales and four competition-level mathematics benchmarks. Specifically, on Qwen3-8B, DCE+SRCL reaches 65.97% Average@12, outperforming OPSD by 35.62 percentage points while reducing mean output length by 7.80% relative to DCE alone.
展开完整摘要收起摘要↓
On-policy distillation (OPD) trains a student model by having it generate trajectories, then matching its next-token predictions with an external teacher's next-token predictions. This provides dense, token-level supervision to the student. On-policy self-distillation (OPSD) eliminates the need for the external teacher. Specifically, a second frozen copy of the student model, now given the ground truth in its context, serves as the teacher. The student model only receives the problem and learns to mimic the privileged teacher model, while the teacher remains frozen throughout training. Previous work showed that freezing the teacher is useful for training stability, but we argue that this can prevent the teacher from incorporating the improvements learned by the student during training. Our primary contribution is to address this limitation with a recursive framework built around two complementary components. First, we let the privileged teacher co-evolve with the student so that revision learned in one round can guide the next, a process we refer to as Dynamic Co-Evolution (DCE). Second, because stronger revision can also make responses too verbose and self-critical, we additionally train on shorter, verified rewrites of the model's own on-policy responses. We call this complementary objective Self-Refined Concise Learning (SRCL). Overall, our comprehensive evaluations show that DCE+SRCL outperforms OPSD across multiple model scales and four competition-level mathematics benchmarks. Specifically, on Qwen3-8B, DCE+SRCL reaches 65.97% Average@12, outperforming OPSD by 35.62 percentage points while reducing mean output length by 7.80% relative to DCE alone.
We propose a momentum-guided federated split distillation framework for personalized, efficient, and autonomous temporal edge intelligence. We introduce TeRR-SAtt, our novel temporal reservoir student attention design that combines fixed reservoir representations, a lightweight temporal student, and personalized output modules. We also present AMGF, our anticipatory momentum-guided fusion mechanism that clusters clients through learning momentum and derives specialized teacher updates. On real-world smart-building data, TeRR-SAtt reduces edge training latency by 65.50%, inference latency by 44.70%, training memory usage by 18.40%, and inference CPU usage by 33.10% over the considered baselines. At the same time, AMGF improves local learning by up to 35.31% in RMSE compared to global updates.
展开完整摘要收起摘要↓
We propose a momentum-guided federated split distillation framework for personalized, efficient, and autonomous temporal edge intelligence. We introduce TeRR-SAtt, our novel temporal reservoir student attention design that combines fixed reservoir representations, a lightweight temporal student, and personalized output modules. We also present AMGF, our anticipatory momentum-guided fusion mechanism that clusters clients through learning momentum and derives specialized teacher updates. On real-world smart-building data, TeRR-SAtt reduces edge training latency by 65.50%, inference latency by 44.70%, training memory usage by 18.40%, and inference CPU usage by 33.10% over the considered baselines. At the same time, AMGF improves local learning by up to 35.31% in RMSE compared to global updates.
作者Anjila Budathoki, Manish Dhakal, Benjamin M. Ampel, Yi Ding
Prior research has demonstrated that the choice of prompt template during Supervised Fine-Tuning (SFT) significantly impacts the robustness of safety alignment afterwards. However, the influence of template selection during Knowledge Distillation (KD) from teacher to student remains largely unexplored. Thus, we fill this gap by analyzing how different template configurations influence the pre-existing safety alignment of the student. We observe a significant degradation of safety alignment present in the aligned base instruct-tuned model. Specifically, we find that utilizing chat templates renders the model more compliant with harmful queries compared to a non-chat template. These findings are consistent across three models: LLaMA, Gemma and Qwen model families and are evaluated across multiple safety benchmarks. We further show that using a non-chat template during distillation better preserves the base student's internal representations, while chat template distillation induces a larger representational shift. Code: https://github.com/anjilab/role-of-prompt-template-in-kd
展开完整摘要收起摘要↓
Prior research has demonstrated that the choice of prompt template during Supervised Fine-Tuning (SFT) significantly impacts the robustness of safety alignment afterwards. However, the influence of template selection during Knowledge Distillation (KD) from teacher to student remains largely unexplored. Thus, we fill this gap by analyzing how different template configurations influence the pre-existing safety alignment of the student. We observe a significant degradation of safety alignment present in the aligned base instruct-tuned model. Specifically, we find that utilizing chat templates renders the model more compliant with harmful queries compared to a non-chat template. These findings are consistent across three models: LLaMA, Gemma and Qwen model families and are evaluated across multiple safety benchmarks. We further show that using a non-chat template during distillation better preserves the base student's internal representations, while chat template distillation induces a larger representational shift. Code: https://github.com/anjilab/role-of-prompt-template-in-kd
作者Xin Di, Mingyu Shi, Yuanfei Bao, Long Peng, Yue Zhao, Jiaming Guo, Renjing Pei, Xueyang Fu, Yang Cao, Zheng-Jun Zha
Real-world image super-resolution (SR) requires recovering perceptually realistic high-resolution images from complex low-resolution observations while preserving faithful content. Diffusion-based SR benefits from strong generative priors but incurs substantial computational overhead, whereas feed-forward CNN and Transformer SR models are efficient yet often struggle to recover realistic high-frequency details. This motivates a natural question: can diffusion priors be transferred to existing diffusion-free SR networks without introducing diffusion components at inference time? To this end, we propose PhoenixSR, a generative heterogeneous distillation framework that transfers diffusion priors to independently designed feed-forward SR networks through score-based distribution matching. Rather than aligning heterogeneous features or imitating sampled diffusion outputs, PhoenixSR uses the pretrained diffusion model as distribution-level supervision, while paired SR supervision preserves reconstruction fidelity. To make distribution matching effective for fidelity-sensitive SR, we introduce Heterogeneous Distribution Adaptation, which adapts the target score to the SR domain, improves tracking of the evolving student distribution, and anchors training with paired supervision. We further employ Directional Reliability Weighting, a lightweight residual-consistency-based reweighting strategy that reduces unstable distributional guidance. All diffusion-related components are removed after training, leaving the original student architecture and inference cost unchanged. Experiments on three SR benchmarks and six feed-forward backbones, including SwinIR, HAT, Real-ESRGAN, and SeeMoRe, show consistent perceptual improvements with largely preserved reconstruction fidelity.
展开完整摘要收起摘要↓
Real-world image super-resolution (SR) requires recovering perceptually realistic high-resolution images from complex low-resolution observations while preserving faithful content. Diffusion-based SR benefits from strong generative priors but incurs substantial computational overhead, whereas feed-forward CNN and Transformer SR models are efficient yet often struggle to recover realistic high-frequency details. This motivates a natural question: can diffusion priors be transferred to existing diffusion-free SR networks without introducing diffusion components at inference time? To this end, we propose PhoenixSR, a generative heterogeneous distillation framework that transfers diffusion priors to independently designed feed-forward SR networks through score-based distribution matching. Rather than aligning heterogeneous features or imitating sampled diffusion outputs, PhoenixSR uses the pretrained diffusion model as distribution-level supervision, while paired SR supervision preserves reconstruction fidelity. To make distribution matching effective for fidelity-sensitive SR, we introduce Heterogeneous Distribution Adaptation, which adapts the target score to the SR domain, improves tracking of the evolving student distribution, and anchors training with paired supervision. We further employ Directional Reliability Weighting, a lightweight residual-consistency-based reweighting strategy that reduces unstable distributional guidance. All diffusion-related components are removed after training, leaving the original student architecture and inference cost unchanged. Experiments on three SR benchmarks and six feed-forward backbones, including SwinIR, HAT, Real-ESRGAN, and SeeMoRe, show consistent perceptual improvements with largely preserved reconstruction fidelity.
Multi-teacher on-policy distillation (MOPD) integrates specialized capabilities into a single student, but existing practice typically hard-routes each prompt to a domain-matched teacher for the entire rollout. This dependence on prompt-level domain labels restricts using unlabeled training mixtures and leaves complementary signals from other teachers unused. We introduce MOPD-Router, a framework that routes supervision over the full teacher pool at each token, without domain labels or training a separate routing model. Its plug-in interface supports different metrics for selecting and weighting teacher-specific OPD signals. Within this interface, we propose ExpertAlign, which scores each teacher by whether its correction to the student at the current token expresses the specialization that teacher acquired during post-training, and compare it against two reference metrics built on teacher confidence (Entropy) and teacher-student discrepancy (Novelty). Experiments on unlabeled and domain-labeled training mixtures under strong-to-weak and same-size distillation scenarios show that ExpertAlign achieves the strongest overall performance in all four settings. On unlabeled data, it improves the overall score by 5.88 (+12.3%) points over Mean aggregation; on domain-labeled data, it outperforms standard MOPD by 3.95 (+7.8%) points without using available domain labels. These results demonstrate token-level routing can exploit cross-domain complementary supervision, and reduce exclusive reliance on prompt-level domain assignment. Code is available at: https://github.com/TURLEing/MOPD-Router.
展开完整摘要收起摘要↓
Multi-teacher on-policy distillation (MOPD) integrates specialized capabilities into a single student, but existing practice typically hard-routes each prompt to a domain-matched teacher for the entire rollout. This dependence on prompt-level domain labels restricts using unlabeled training mixtures and leaves complementary signals from other teachers unused. We introduce MOPD-Router, a framework that routes supervision over the full teacher pool at each token, without domain labels or training a separate routing model. Its plug-in interface supports different metrics for selecting and weighting teacher-specific OPD signals. Within this interface, we propose ExpertAlign, which scores each teacher by whether its correction to the student at the current token expresses the specialization that teacher acquired during post-training, and compare it against two reference metrics built on teacher confidence (Entropy) and teacher-student discrepancy (Novelty). Experiments on unlabeled and domain-labeled training mixtures under strong-to-weak and same-size distillation scenarios show that ExpertAlign achieves the strongest overall performance in all four settings. On unlabeled data, it improves the overall score by 5.88 (+12.3%) points over Mean aggregation; on domain-labeled data, it outperforms standard MOPD by 3.95 (+7.8%) points without using available domain labels. These results demonstrate token-level routing can exploit cross-domain complementary supervision, and reduce exclusive reliance on prompt-level domain assignment. Code is available at: https://github.com/TURLEing/MOPD-Router.
作者Taeckyung Lee, Rinat Amankos, Jeonghye Kim, Hyungjun Yoon, Woogyeol Jin, Sung-Ju Lee
On-policy self-distillation (OPSD) provides dense teacher targets, but evaluates them only along student-sampled rollouts. When the privileged teacher favors an alternative action at a visited prefix, OPSD can provide a target for the branch decision but cannot supervise the successor contexts induced by that action unless the student samples it. This creates a training-time data-collection bottleneck and suggests a different role for teacher-student disagreement: proposing a trajectory branch rather than identifying a sufficient local repair. Our diagnostic framework using controlled token interventions reveals that a teacher-preferred token at peak disagreement can improve student continuation success, while its local corrective value is limited. Motivated by this finding, we introduce a simple branch-regenerate-distill algorithm, Trajectory-Intervention Self-Distillation (TISD). TISD forces a teacher-selected branch action, returns suffix generation to the student, and distills the full trajectory under the privileged-context-conditioned teacher. Across the coding models, TISD improves average Avg@4 over SDPO by 1.2 percentage points. Across the science domains, it improves average Avg@128 by 0.8 points under an equal-step budget and by 0.3 points under an equal-time budget. These results support teacher-guided branching as a way to expose useful successor contexts for self-distillation.
展开完整摘要收起摘要↓
On-policy self-distillation (OPSD) provides dense teacher targets, but evaluates them only along student-sampled rollouts. When the privileged teacher favors an alternative action at a visited prefix, OPSD can provide a target for the branch decision but cannot supervise the successor contexts induced by that action unless the student samples it. This creates a training-time data-collection bottleneck and suggests a different role for teacher-student disagreement: proposing a trajectory branch rather than identifying a sufficient local repair. Our diagnostic framework using controlled token interventions reveals that a teacher-preferred token at peak disagreement can improve student continuation success, while its local corrective value is limited. Motivated by this finding, we introduce a simple branch-regenerate-distill algorithm, Trajectory-Intervention Self-Distillation (TISD). TISD forces a teacher-selected branch action, returns suffix generation to the student, and distills the full trajectory under the privileged-context-conditioned teacher. Across the coding models, TISD improves average Avg@4 over SDPO by 1.2 percentage points. Across the science domains, it improves average Avg@128 by 0.8 points under an equal-step budget and by 0.3 points under an equal-time budget. These results support teacher-guided branching as a way to expose useful successor contexts for self-distillation.
Black-box On-Policy Distillation (OPD) seeks to improve a student from its own generations when the teacher provides sampled responses but not token probabilities. Adversarial distillation offers one route: it learns a discriminator over prompt-matched teacher and student responses and uses its score as the policy reward. However, sampling discriminator negatives from the latest student at each step couples the learned reward to a negative distribution that changes after every policy update. We address this moving-target problem with persistent-negative adversarial distillation, a live-pool method that replaces a fraction of each discriminator batch with historical, prompt-matched teacher--student comparisons. Under matched discriminator compute, historical comparisons train the discriminator, while GRPO remains on-policy with fresh student responses. Our analysis identifies the Bayes-optimal reward as a teacher-to-negative log-density ratio and, under explicit assumptions, shows how persistent negatives anchor the discriminator and reduce reward-estimation MSE relative to fresh-negative training. Across two student families, three judges, and four judged-chat benchmarks, persistent-negative adversarial distillation consistently improves performance over current methods at matched discriminator compute. It also yields smoother fresh-policy discriminator trajectories, with fewer below-chance dips. These findings identify the discriminator's negative distribution as an important design axis in black-box on-policy distillation.
展开完整摘要收起摘要↓
Black-box On-Policy Distillation (OPD) seeks to improve a student from its own generations when the teacher provides sampled responses but not token probabilities. Adversarial distillation offers one route: it learns a discriminator over prompt-matched teacher and student responses and uses its score as the policy reward. However, sampling discriminator negatives from the latest student at each step couples the learned reward to a negative distribution that changes after every policy update. We address this moving-target problem with persistent-negative adversarial distillation, a live-pool method that replaces a fraction of each discriminator batch with historical, prompt-matched teacher--student comparisons. Under matched discriminator compute, historical comparisons train the discriminator, while GRPO remains on-policy with fresh student responses. Our analysis identifies the Bayes-optimal reward as a teacher-to-negative log-density ratio and, under explicit assumptions, shows how persistent negatives anchor the discriminator and reduce reward-estimation MSE relative to fresh-negative training. Across two student families, three judges, and four judged-chat benchmarks, persistent-negative adversarial distillation consistently improves performance over current methods at matched discriminator compute. It also yields smoother fresh-policy discriminator trajectories, with fewer below-chance dips. These findings identify the discriminator's negative distribution as an important design axis in black-box on-policy distillation.
Vision-Language-Action models achieve strong performance in robot manipulation, but often require large numbers of parameters. In this work, we propose the Future Representation Action Model (FRAM), a small policy that explicitly links the future end-effector trajectory to the current visual input. FRAM uses the image coordinates of the predicted trajectory as spatial pointers and reads local visual features related to the motion from the current image. This organizes the information for action generation into the reference position (Where), the visual state (What), and the future motion (Future). Trajectory labels are generated automatically from demonstrations and camera geometry, so no manual annotation is needed. With 138.7M parameters, including a frozen language encoder, FRAM reaches an average success rate of 92.2% over the four standard LIBERO suites, close to the 94.2% of $π_0$ with 3.3B parameters. Without extra training, it also reaches an average of 67.3% on LIBERO-Plus. Ablations confirm that both the future trajectory and the local visual features improve performance and robustness. On a real dual-arm UR5e, FRAM stacks cups using only wrist cameras, including choosing and switching between the left and right arms. These results show that selecting visual information based on future motion is an effective way to obtain both high performance and robustness in a small robot policy.
展开完整摘要收起摘要↓
Vision-Language-Action models achieve strong performance in robot manipulation, but often require large numbers of parameters. In this work, we propose the Future Representation Action Model (FRAM), a small policy that explicitly links the future end-effector trajectory to the current visual input. FRAM uses the image coordinates of the predicted trajectory as spatial pointers and reads local visual features related to the motion from the current image. This organizes the information for action generation into the reference position (Where), the visual state (What), and the future motion (Future). Trajectory labels are generated automatically from demonstrations and camera geometry, so no manual annotation is needed. With 138.7M parameters, including a frozen language encoder, FRAM reaches an average success rate of 92.2% over the four standard LIBERO suites, close to the 94.2% of $π_0$ with 3.3B parameters. Without extra training, it also reaches an average of 67.3% on LIBERO-Plus. Ablations confirm that both the future trajectory and the local visual features improve performance and robustness. On a real dual-arm UR5e, FRAM stacks cups using only wrist cameras, including choosing and switching between the left and right arms. These results show that selecting visual information based on future motion is an effective way to obtain both high performance and robustness in a small robot policy.
Speech content representations are central to voice conversion, speech-to-speech translation, and multimodal language models, yet they are rarely compared under a common generative framework that directly measures what each representation contains. We address this by training a generative model conditioned solely on each representation and evaluating the generated audio along the content, speaker identity, and prosody axes. Across SSL features, supervised tokens, posteriorgrams, and neural audio codecs, we find two distinct regimes: representations that nearly reconstruct the original audio, and representations that effectively disentangle speaker identity. These results show that disentanglement depends not on supervision alone, but on the interaction between the training objective and the representation's information capacity: supervised representations only disentangle speaker identity when their capacity is sufficiently constrained.
展开完整摘要收起摘要↓
Speech content representations are central to voice conversion, speech-to-speech translation, and multimodal language models, yet they are rarely compared under a common generative framework that directly measures what each representation contains. We address this by training a generative model conditioned solely on each representation and evaluating the generated audio along the content, speaker identity, and prosody axes. Across SSL features, supervised tokens, posteriorgrams, and neural audio codecs, we find two distinct regimes: representations that nearly reconstruct the original audio, and representations that effectively disentangle speaker identity. These results show that disentanglement depends not on supervision alone, but on the interaction between the training objective and the representation's information capacity: supervised representations only disentangle speaker identity when their capacity is sufficiently constrained.
作者Olga Zatsarynna, Denis Korzhenkov, Juergen Gall, Amir Habibian, Mohsen Ghafoorian
Efficient video generation requires reducing the quadratic cost of self-attention over long spatio-temporal token sequences. Existing efficient-attention methods typically apply the same computation pattern to every token, even though denoising difficulty varies substantially across video regions and evolves throughout the generation process. We introduce HetA-DiT, a heterogeneous attention mechanism that adaptively allocates computation according to token difficulty. A lightweight uncertainty branch predicts a token-wise estimate of denoising difficulty, which is used to route uncertain tokens through dense global attention while processing more reliable tokens with efficient local attention. The resulting routing is content- and timestep-adaptive, retains global context where it matters most, and provides a single parameter for controlling the quality-efficiency trade-off. HetA-DiT is compatible with few-step distribution-matching distillation and introduces no additional Transformer evaluation at inference time by reusing uncertainty estimates from the preceding denoising step. We evaluate the method on DMD-distilled Wan2.2-5B and Wan2.1-1.3B models. Across VBench, VBench-2.0, and human preference evaluation, HetA-DiT maintains competitive generation quality while routing only approximately 20% of tokens through dense attention.
展开完整摘要收起摘要↓
Efficient video generation requires reducing the quadratic cost of self-attention over long spatio-temporal token sequences. Existing efficient-attention methods typically apply the same computation pattern to every token, even though denoising difficulty varies substantially across video regions and evolves throughout the generation process. We introduce HetA-DiT, a heterogeneous attention mechanism that adaptively allocates computation according to token difficulty. A lightweight uncertainty branch predicts a token-wise estimate of denoising difficulty, which is used to route uncertain tokens through dense global attention while processing more reliable tokens with efficient local attention. The resulting routing is content- and timestep-adaptive, retains global context where it matters most, and provides a single parameter for controlling the quality-efficiency trade-off. HetA-DiT is compatible with few-step distribution-matching distillation and introduces no additional Transformer evaluation at inference time by reusing uncertainty estimates from the preceding denoising step. We evaluate the method on DMD-distilled Wan2.2-5B and Wan2.1-1.3B models. Across VBench, VBench-2.0, and human preference evaluation, HetA-DiT maintains competitive generation quality while routing only approximately 20% of tokens through dense attention.
作者Parsa Mastouri Kashani, Jan-Gerrit Habekost, Stefan Wermter
Vision-Language-Action models (VLAs), while showing strong potential for robot control, require massive amounts of high-quality imitation learning data. Moreover, the absence of an explicit world model casts further doubt on their control capabilities. In this concept paper, we propose a novel architecture that addresses sample efficiency in VLAs by training a predictive world model on the embedding space of the VLA's vision encoder. We hypothesize that these embeddings are action-relevant and usable for future prediction. To this end, we propose using the suggested architecture to investigate how well these embeddings predict the future based on actions, as the inability to do so would mark a key limitation of VLA architectures: the lack of a non-lossy implicit world model to simulate real-world dynamics. The proposed architecture differs from the standard world model dynamics as the loss comes from the embedding space rather than the pixel space, similar to joint embedding predictive architectures. Furthermore, the trained world model can be utilized for short-term planning tasks by sampling VLA actions given goal images. We intend to examine the richness of vision embeddings in VLAs and reduce their high data requirements through a world model that can also generate plans during inference.
展开完整摘要收起摘要↓
Vision-Language-Action models (VLAs), while showing strong potential for robot control, require massive amounts of high-quality imitation learning data. Moreover, the absence of an explicit world model casts further doubt on their control capabilities. In this concept paper, we propose a novel architecture that addresses sample efficiency in VLAs by training a predictive world model on the embedding space of the VLA's vision encoder. We hypothesize that these embeddings are action-relevant and usable for future prediction. To this end, we propose using the suggested architecture to investigate how well these embeddings predict the future based on actions, as the inability to do so would mark a key limitation of VLA architectures: the lack of a non-lossy implicit world model to simulate real-world dynamics. The proposed architecture differs from the standard world model dynamics as the loss comes from the embedding space rather than the pixel space, similar to joint embedding predictive architectures. Furthermore, the trained world model can be utilized for short-term planning tasks by sampling VLA actions given goal images. We intend to examine the richness of vision embeddings in VLAs and reduce their high data requirements through a world model that can also generate plans during inference.
The growing popularity of short videos has driven demand for one-click content creation. Visual templates turn uploaded images into personalized content with preset effects, but reusable template generation still requires substantial manual effort in asset preparation and tool orchestration. We propose TemplateCraft, a multi-agent system that converts natural-language instructions into client-executable templates through planning, material generation, effect-workflow generation, and protocol compilation. Its Planner-Evaluator loop uses execution feedback for targeted rollback, while stage-level and long-term memory support revision without parameter updates. We evaluate TemplateCraft on TemplateBench, derived from 60 real-world templates. With the same Qwen3-VL backbone, TemplateCraft raises image/video generation success rates from 56.7%/30.0% to 66.7%/50.0% over Planner-only (best-of-three) and improves template adherence and style consistency. With additional evaluation and revision, it matches or exceeds a GPT-4o Planner-only baseline on selected metrics. Persistent assets further improve cross-input style consistency.
展开完整摘要收起摘要↓
The growing popularity of short videos has driven demand for one-click content creation. Visual templates turn uploaded images into personalized content with preset effects, but reusable template generation still requires substantial manual effort in asset preparation and tool orchestration. We propose TemplateCraft, a multi-agent system that converts natural-language instructions into client-executable templates through planning, material generation, effect-workflow generation, and protocol compilation. Its Planner-Evaluator loop uses execution feedback for targeted rollback, while stage-level and long-term memory support revision without parameter updates. We evaluate TemplateCraft on TemplateBench, derived from 60 real-world templates. With the same Qwen3-VL backbone, TemplateCraft raises image/video generation success rates from 56.7%/30.0% to 66.7%/50.0% over Planner-only (best-of-three) and improves template adherence and style consistency. With additional evaluation and revision, it matches or exceeds a GPT-4o Planner-only baseline on selected metrics. Persistent assets further improve cross-input style consistency.
作者Zihan Su, Siwen Lu, Junhao Zhuang, Zeyue Xue, Haoyang Huang, Guanghao Li, Xiaofeng Tan, Chun Yuan, Nan Duan
Audio-driven streaming avatar generation requires real-time synthesis of speech-synchronized videos with dynamic and diverse motion. Self Forcing uses Distribution Matching Distillation (DMD) to distill bidirectional video diffusion models into causal, few-step generators for real-time streaming. However, DMD minimizes a reverse KL divergence, which is inherently mode-seeking: it causes the student to discard high-dynamic modes and collapse onto static outputs, compressing both dynamics and diversity of generated videos. We find that this collapse is region-heterogeneous: person regions involving pose and gesture variations suffer the largest diversity loss, the audio-driven mouth region shows a small loss, and the background remains nearly stable. Based on this observation, we propose Routed Forcing, which routes the distillation objective by semantic region and noise stage to improve dynamics and diversity while preserving visual quality. Specifically, (1) Where to Force: Semantic-Region Routing applies Data-Forcing Distillation (DFD), which supervises the student with real videos, to the person region where diversity collapse is most severe, while retaining DMD for the mouth and background to preserve lip synchronization and scene stability. (2) When to Force: Noise-Stage Routing activates DFD at high noise stages, where real video serves as effective supervision to inject diverse and dynamic motion patterns. At low noise stages, DMD is used to refine details, avoiding blur and artifacts from spatial differences between real video and student-generated video. Experiments show that Routed Forcing improves dynamics by up to 45% and diversity by 7-25% over Self Forcing, while preserving video quality and lip synchronization.
展开完整摘要收起摘要↓
Audio-driven streaming avatar generation requires real-time synthesis of speech-synchronized videos with dynamic and diverse motion. Self Forcing uses Distribution Matching Distillation (DMD) to distill bidirectional video diffusion models into causal, few-step generators for real-time streaming. However, DMD minimizes a reverse KL divergence, which is inherently mode-seeking: it causes the student to discard high-dynamic modes and collapse onto static outputs, compressing both dynamics and diversity of generated videos. We find that this collapse is region-heterogeneous: person regions involving pose and gesture variations suffer the largest diversity loss, the audio-driven mouth region shows a small loss, and the background remains nearly stable. Based on this observation, we propose Routed Forcing, which routes the distillation objective by semantic region and noise stage to improve dynamics and diversity while preserving visual quality. Specifically, (1) Where to Force: Semantic-Region Routing applies Data-Forcing Distillation (DFD), which supervises the student with real videos, to the person region where diversity collapse is most severe, while retaining DMD for the mouth and background to preserve lip synchronization and scene stability. (2) When to Force: Noise-Stage Routing activates DFD at high noise stages, where real video serves as effective supervision to inject diverse and dynamic motion patterns. At low noise stages, DMD is used to refine details, avoiding blur and artifacts from spatial differences between real video and student-generated video. Experiments show that Routed Forcing improves dynamics by up to 45% and diversity by 7-25% over Self Forcing, while preserving video quality and lip synchronization.
作者Haoyang Li, Ruoxi Sun, Qingqing Ye, Benjamin Zi Hao Zhao, Yaxin Xiao, Jason Xue, Haibo Hu
Text-to-image diffusion models enable data-efficient "mimicry" attacks, wherein adversaries fine-tune the model on a handful of public photos to synthesize convincing forgeries of a target individual. A common countermeasure is to embed imperceptible, low-energy watermarks, yet recent studies show these signatures are brittle: modest post-processing or lightweight adversarial perturbations readily suppress detection, exposing a fundamental tension between imperceptibility and robustness. We introduce FeatMark, a watermarking framework that shifts from pixel-level, energy-starved perturbations to inconspicuous semantic features: small, scene-consistent micro-features that remain natural to humans while providing a stronger, machine-verifiable provenance signal. FeatMark builds domain-specific feature banks that encode each watermark as a compact concept program, pairing open-vocabulary semantic cues with reliable edit regions and instruction templates. It then automatically selects features that are both feasible and executable and injects them through modular, mask-guided concept editing, yielding highly localized, scene-consistent micro-edits that are difficult to perceive. We conduct extensive experiments across VGGFace2, CelebA-HQ, and WikiArt, evaluating against 10 strong watermark removal/purification attacks (including regeneration-style purification) and several bespoke adaptive attacks tailored to FeatMark, to assess perceptual fidelity, watermark detection accuracy, and robustness. We further demonstrate FeatMark's extensibility to video mimicry attacks. The results show FeatMark remains virtually impervious, withstanding all evaluated attacks with negligible bit-accuracy and fidelity degradation.
展开完整摘要收起摘要↓
Text-to-image diffusion models enable data-efficient "mimicry" attacks, wherein adversaries fine-tune the model on a handful of public photos to synthesize convincing forgeries of a target individual. A common countermeasure is to embed imperceptible, low-energy watermarks, yet recent studies show these signatures are brittle: modest post-processing or lightweight adversarial perturbations readily suppress detection, exposing a fundamental tension between imperceptibility and robustness. We introduce FeatMark, a watermarking framework that shifts from pixel-level, energy-starved perturbations to inconspicuous semantic features: small, scene-consistent micro-features that remain natural to humans while providing a stronger, machine-verifiable provenance signal. FeatMark builds domain-specific feature banks that encode each watermark as a compact concept program, pairing open-vocabulary semantic cues with reliable edit regions and instruction templates. It then automatically selects features that are both feasible and executable and injects them through modular, mask-guided concept editing, yielding highly localized, scene-consistent micro-edits that are difficult to perceive. We conduct extensive experiments across VGGFace2, CelebA-HQ, and WikiArt, evaluating against 10 strong watermark removal/purification attacks (including regeneration-style purification) and several bespoke adaptive attacks tailored to FeatMark, to assess perceptual fidelity, watermark detection accuracy, and robustness. We further demonstrate FeatMark's extensibility to video mimicry attacks. The results show FeatMark remains virtually impervious, withstanding all evaluated attacks with negligible bit-accuracy and fidelity degradation.
作者Kevin Jiang, Morgane Austern, Edgar Dobriban, Jason M. Klusowski
Generative AI systems are increasingly used, but aligning their outputs with user requirements poses a continuing challenge. Here, we aim to ensure that the distribution of an attribute of an AI-generated output aligns with a user-specified target. This is motivated by examples such as fairness, where we want to ensure that a protected attribute (e.g., gender, race, or age categories) follows a desired distribution, and synthetic data generation, where we want the generated data to be representative of a target distribution. We study the practically important black-box access setting, where a user can repeatedly query a generative AI model. The goal is to return $m\ge 1$ outputs whose joint attribute distribution is as close as possible to this target. For both exact and approximate alignment, we develop algorithms that minimize the expected number of queries to the generator, and we further demonstrate their optimality as the number of requested outputs $m \rightarrow \infty$. Experiments on text-to-image generation and geocoded persona generation tasks show that our post-processing algorithms improve statistical attribute alignment, complementing prompting-based interventions.
展开完整摘要收起摘要↓
Generative AI systems are increasingly used, but aligning their outputs with user requirements poses a continuing challenge. Here, we aim to ensure that the distribution of an attribute of an AI-generated output aligns with a user-specified target. This is motivated by examples such as fairness, where we want to ensure that a protected attribute (e.g., gender, race, or age categories) follows a desired distribution, and synthetic data generation, where we want the generated data to be representative of a target distribution. We study the practically important black-box access setting, where a user can repeatedly query a generative AI model. The goal is to return $m\ge 1$ outputs whose joint attribute distribution is as close as possible to this target. For both exact and approximate alignment, we develop algorithms that minimize the expected number of queries to the generator, and we further demonstrate their optimality as the number of requested outputs $m \rightarrow \infty$. Experiments on text-to-image generation and geocoded persona generation tasks show that our post-processing algorithms improve statistical attribute alignment, complementing prompting-based interventions.
作者Zeqiang Zhang, Fabian Wurzberger, Maximilian Otte, Daniel Schmid, Sebastian Gottwald, Arne Peter Raulf, Daniel Alexander Braun
A robust world model must strike the balance between faithfully capturing environmental dynamics and abstracting away from irrelevant content. While reconstruction-based world models ensure faithful supervision, they misallocate representational capacity by pixel area rather than dynamics relevance for visual tasks, which can cause task-irrelevant content to dominate the learned representation. Alternatively, reconstruction-free methods avoid this bias but risk discarding possibly relevant information. We propose StarWM, which uses a cross-attention module trained on self-supervised dynamics to decide where reconstruction applies. A dual-stream decoder then restricts reconstruction to the attended regions, with stop-gradient barriers preventing interference between the two objectives. These components allows reconstruction to supervise the visual content of attended regions without contaminating the latent with non-predictive information. On DeepMind Control with dynamic video backgrounds, default (reward-free) StarWM achieves the strongest performance under random-frame distractors and substantially outperforms reconstruction-based baselines under sequential video. In addition, its reward-augmented variant matches or exceeds reconstruction-free methods on sequential video, achieving the highest overall return across all distractor regimes. Mechanistic probing confirms StarWM preserves state attributes with near-perfect fidelity through long-horizon imagination while systematically discarding distractors.
展开完整摘要收起摘要↓
A robust world model must strike the balance between faithfully capturing environmental dynamics and abstracting away from irrelevant content. While reconstruction-based world models ensure faithful supervision, they misallocate representational capacity by pixel area rather than dynamics relevance for visual tasks, which can cause task-irrelevant content to dominate the learned representation. Alternatively, reconstruction-free methods avoid this bias but risk discarding possibly relevant information. We propose StarWM, which uses a cross-attention module trained on self-supervised dynamics to decide where reconstruction applies. A dual-stream decoder then restricts reconstruction to the attended regions, with stop-gradient barriers preventing interference between the two objectives. These components allows reconstruction to supervise the visual content of attended regions without contaminating the latent with non-predictive information. On DeepMind Control with dynamic video backgrounds, default (reward-free) StarWM achieves the strongest performance under random-frame distractors and substantially outperforms reconstruction-based baselines under sequential video. In addition, its reward-augmented variant matches or exceeds reconstruction-free methods on sequential video, achieving the highest overall return across all distractor regimes. Mechanistic probing confirms StarWM preserves state attributes with near-perfect fidelity through long-horizon imagination while systematically discarding distractors.
Action-conditioned video world models aim to predict scene evolution under different actions, a capability that is essential for reliable planning, decision-making, and interaction in dynamic environments. However, futures generated independently from the same initial scene may each appear plausible while implying incompatible physical properties, such as friction or mass. This inconsistency can lead to contradictory predictions across interventions, making it difficult for the model to maintain a coherent understanding of the underlying world and limiting its reliability for planning and decision-making. To address these issues, we propose OneWorld, a shared-mechanism counterfactual generation framework that jointly models multiple action-conditioned futures under a common latent physical mechanism. A physical mechanism interpreter first infers a distribution over latent mechanisms from each action-outcome branch. These distributions are then aggregated into shared-world evidence, which captures whether the branches admit a common physical explanation while accounting for uncertainty in less informative branches. This evidence constrains flow training and guides sampling, encouraging consistency in the underlying physical mechanism while preserving the distinct outcomes induced by different actions. We further introduce a multi-intervention evaluation protocol in controlled environments, following the interaction settings of ACWM-Phys, to assess whether generated futures can be jointly explained by the same physical parameters, alongside standard measures of single-rollout prediction quality. Experiments in these environments show that OneWorld improves cross-intervention physical consistency while maintaining competitive single-rollout prediction quality.
展开完整摘要收起摘要↓
Action-conditioned video world models aim to predict scene evolution under different actions, a capability that is essential for reliable planning, decision-making, and interaction in dynamic environments. However, futures generated independently from the same initial scene may each appear plausible while implying incompatible physical properties, such as friction or mass. This inconsistency can lead to contradictory predictions across interventions, making it difficult for the model to maintain a coherent understanding of the underlying world and limiting its reliability for planning and decision-making. To address these issues, we propose OneWorld, a shared-mechanism counterfactual generation framework that jointly models multiple action-conditioned futures under a common latent physical mechanism. A physical mechanism interpreter first infers a distribution over latent mechanisms from each action-outcome branch. These distributions are then aggregated into shared-world evidence, which captures whether the branches admit a common physical explanation while accounting for uncertainty in less informative branches. This evidence constrains flow training and guides sampling, encouraging consistency in the underlying physical mechanism while preserving the distinct outcomes induced by different actions. We further introduce a multi-intervention evaluation protocol in controlled environments, following the interaction settings of ACWM-Phys, to assess whether generated futures can be jointly explained by the same physical parameters, alongside standard measures of single-rollout prediction quality. Experiments in these environments show that OneWorld improves cross-intervention physical consistency while maintaining competitive single-rollout prediction quality.
Most existing human motion generation (HMG) methods use cross-attention modules to inject text semantics, but ignore the importance of bidirectional modeling between motion and text tokens, which limits text comprehension. A straightforward idea is introducing multimodal diffusion transformers (MMDiT), which have shown effective joint text--visual modeling in vision generation, into HMG. However, we find that articulated motion is temporally coherent but weakly correlated across joints, in which directly applying an MMDiT with flow matching produces poorly coordinated and jerky motion. In this work, we propose Timo, a novel kinematics-aware MMDiT framework tailored for HMG. Timo combines fully shared multimodal attention for bidirectional text--motion modeling with flow matching, geometric and rotational-kinematics supervision that compares actual rotations and their changes over time, and a two-stage curriculum progressing from broad motion learning to detailed caption alignment. Further, we construct a benchmark of $40{,}025$ held-out clips from six public datasets spanning diverse actions, assessing six complementary dimensions under a common evaluator and scoring protocol. Our model substantially outperforms state-of-the-art methods in both quantitative and qualitative evaluations. Remarkably, Timo surpasses Kimodo on five of six dimensions, achieving a $40.8$% relative improvement in the average benchmark score. Project page: https://kyfafyd.wang/projects/timo. Demo page: https://timo.kyfafyd.wang.
展开完整摘要收起摘要↓
Most existing human motion generation (HMG) methods use cross-attention modules to inject text semantics, but ignore the importance of bidirectional modeling between motion and text tokens, which limits text comprehension. A straightforward idea is introducing multimodal diffusion transformers (MMDiT), which have shown effective joint text--visual modeling in vision generation, into HMG. However, we find that articulated motion is temporally coherent but weakly correlated across joints, in which directly applying an MMDiT with flow matching produces poorly coordinated and jerky motion. In this work, we propose Timo, a novel kinematics-aware MMDiT framework tailored for HMG. Timo combines fully shared multimodal attention for bidirectional text--motion modeling with flow matching, geometric and rotational-kinematics supervision that compares actual rotations and their changes over time, and a two-stage curriculum progressing from broad motion learning to detailed caption alignment. Further, we construct a benchmark of $40{,}025$ held-out clips from six public datasets spanning diverse actions, assessing six complementary dimensions under a common evaluator and scoring protocol. Our model substantially outperforms state-of-the-art methods in both quantitative and qualitative evaluations. Remarkably, Timo surpasses Kimodo on five of six dimensions, achieving a $40.8$% relative improvement in the average benchmark score. Project page: https://kyfafyd.wang/projects/timo. Demo page: https://timo.kyfafyd.wang.
Recent advances in text-to-image (T2I) generation have substantially improved visual synthesis, but have also raised increasing safety concerns due to their potential to generate harmful or undesirable content. Existing concept erasure methods predominantly follow a static weight paradigm, producing a single frozen adapter that struggles to adapt to diverse prompt variations and suffers from parameter interference when scaling to multiple concepts. We propose HyperErase, a framework for concept erasure based on hypernetwork-driven prompt-conditioned parameter synthesis. Our approach first reframes concept erasure as prompt-conditioned parameter amortization and trains a hypernetwork to map textual descriptions to prompt-specific LoRA updates, eliminating the need for per-prompt gradient optimization or manual LoRA merging. To further improve the stability and precision of synthesized adapters, we develop a decoupled rectification strategy, which disentangles LoRA tokens into pattern and scale subspaces, applies a square-root transform to curb multiplicative over-scaling, and leverages teacher-derived canonical priors for inference-time correction. Extensive experiments across major concept categories demonstrate that HyperErase consistently improves the trade-off between erasure effectiveness, image quality, and semantic alignment, achieving performance comparable to gold-standard single-concept baselines. Furthermore, the resulting models can provide specialized LoRAs for each input prompt variation in a single forward pass without requiring gradient updates during inference. These principled and flexible framework offers a new paradigm for concept erasure in T2I models.
展开完整摘要收起摘要↓
Recent advances in text-to-image (T2I) generation have substantially improved visual synthesis, but have also raised increasing safety concerns due to their potential to generate harmful or undesirable content. Existing concept erasure methods predominantly follow a static weight paradigm, producing a single frozen adapter that struggles to adapt to diverse prompt variations and suffers from parameter interference when scaling to multiple concepts. We propose HyperErase, a framework for concept erasure based on hypernetwork-driven prompt-conditioned parameter synthesis. Our approach first reframes concept erasure as prompt-conditioned parameter amortization and trains a hypernetwork to map textual descriptions to prompt-specific LoRA updates, eliminating the need for per-prompt gradient optimization or manual LoRA merging. To further improve the stability and precision of synthesized adapters, we develop a decoupled rectification strategy, which disentangles LoRA tokens into pattern and scale subspaces, applies a square-root transform to curb multiplicative over-scaling, and leverages teacher-derived canonical priors for inference-time correction. Extensive experiments across major concept categories demonstrate that HyperErase consistently improves the trade-off between erasure effectiveness, image quality, and semantic alignment, achieving performance comparable to gold-standard single-concept baselines. Furthermore, the resulting models can provide specialized LoRAs for each input prompt variation in a single forward pass without requiring gradient updates during inference. These principled and flexible framework offers a new paradigm for concept erasure in T2I models.
Existing text-to-video (T2V) jailbreak methods mainly seek more effective or stealthier attack candidates. In guarded T2V systems, however, video generation and security evaluation are costly, so an attacker often cannot test a large candidate pool. We therefore formulate T2V jailbreak as a query-constrained candidate allocation and ranking problem and propose TempQ-Jail. The method combines heterogeneous attack mechanisms to expand candidate coverage, estimates each candidate's end-to-end attack value from security-gate passage, dangerous visual generation, preservation of the original intent, and temporal validity, and ranks candidates so that high-value attacks appear early in a limited query trajectory. We evaluate TempQ-Jail on CogVideoX-5B using 70 common viable intents derived from T2VSafetyBench and compare it with six representative T2V jailbreak methods under a unified protocol. TempQ-Jail achieves TP-ASR@5 and TP-ASR@10 of 48.9% and 65.4%, improving over the strongest baselines by 4.6 and 4.0 percentage points, respectively. It also obtains the highest AUC-TP (0.469) and the lowest AvgQ (6.3). Analyses of query trajectories, candidate allocation, failure attribution, and ablations show that TempQ-Jail more effectively identifies and prioritises candidates with complete attack potential under limited query budgets.
展开完整摘要收起摘要↓
Existing text-to-video (T2V) jailbreak methods mainly seek more effective or stealthier attack candidates. In guarded T2V systems, however, video generation and security evaluation are costly, so an attacker often cannot test a large candidate pool. We therefore formulate T2V jailbreak as a query-constrained candidate allocation and ranking problem and propose TempQ-Jail. The method combines heterogeneous attack mechanisms to expand candidate coverage, estimates each candidate's end-to-end attack value from security-gate passage, dangerous visual generation, preservation of the original intent, and temporal validity, and ranks candidates so that high-value attacks appear early in a limited query trajectory. We evaluate TempQ-Jail on CogVideoX-5B using 70 common viable intents derived from T2VSafetyBench and compare it with six representative T2V jailbreak methods under a unified protocol. TempQ-Jail achieves TP-ASR@5 and TP-ASR@10 of 48.9% and 65.4%, improving over the strongest baselines by 4.6 and 4.0 percentage points, respectively. It also obtains the highest AUC-TP (0.469) and the lowest AvgQ (6.3). Analyses of query trajectories, candidate allocation, failure attribution, and ablations show that TempQ-Jail more effectively identifies and prioritises candidates with complete attack potential under limited query budgets.
作者Shengqi Dang, Zhengxi Yu, Feilin Han, Xingyu Lan, Nan Cao
Text-to-image generation has advanced in controlling what, where, and how objects appear, yet how visual attention is distributed among objects remains largely unexplored. In this paper, we introduce Target Saliency Boosting, a new task aimed at boosting the visual saliency of a specific object during text-to-image generation without requiring any visual priors. Our key insight is that visual saliency is inherently relative: boosting the saliency of a target object also depends on the global saliency distribution across all objects in the scene. Based on this insight, we propose GazeME, a lightweight framework that uses saliency-marked prompts, inserting learnable marker tokens around object descriptions to indicate which objects to visually emphasize or suppress. To learn these markers, we construct a saliency-semantics dataset that associates objects in image--prompt pairs with object-level saliency scores, and propose Saliency Prior Marker Activation (SPMA), a saliency-aware stochastic marker activation strategy that exploits relative saliency relationships for robust training. During inference, GazeME automatically inserts appropriate markers into the prompt, thereby directly enhancing the visual saliency of the target object. Extensive experiments demonstrate that GazeME effectively boosts target saliency while preserving both semantic alignment and image quality.
展开完整摘要收起摘要↓
Text-to-image generation has advanced in controlling what, where, and how objects appear, yet how visual attention is distributed among objects remains largely unexplored. In this paper, we introduce Target Saliency Boosting, a new task aimed at boosting the visual saliency of a specific object during text-to-image generation without requiring any visual priors. Our key insight is that visual saliency is inherently relative: boosting the saliency of a target object also depends on the global saliency distribution across all objects in the scene. Based on this insight, we propose GazeME, a lightweight framework that uses saliency-marked prompts, inserting learnable marker tokens around object descriptions to indicate which objects to visually emphasize or suppress. To learn these markers, we construct a saliency-semantics dataset that associates objects in image--prompt pairs with object-level saliency scores, and propose Saliency Prior Marker Activation (SPMA), a saliency-aware stochastic marker activation strategy that exploits relative saliency relationships for robust training. During inference, GazeME automatically inserts appropriate markers into the prompt, thereby directly enhancing the visual saliency of the target object. Extensive experiments demonstrate that GazeME effectively boosts target saliency while preserving both semantic alignment and image quality.
Existing human motion diffusion methods provide strong motion generation quality, and recent style transfer models can inject target style cues, but fine-grained continuous control of style intensity remains underexplored. In production, style intensity is subjective across artists and directors, so the practical requirement is not a universal absolute unit, but a reliable monotonic control axis. We propose Motion Style Slider, a motion-to-motion style transfer framework for endpoint-supervised continuous control. Given a content motion and a style motion, we construct a style direction in a learned motion-style embedding space and condition diffusion generation with a scalar intensity. The training objective combines diffusion denoising with latent intensity regularization to encourage smooth and monotonic style scaling without requiring intermediate-intensity ground-truth motions. Our framework is compatible with pretrained motion diffusion backbones and supports heterogeneous style datasets, including the multi-actor style motion dataset. To test out-of-range usability, we additionally introduce a small real-capture over-reaction extension and evaluate large-intensity behavior against these unseen targets. Experiments measure controllability, interpolation/extrapolation behavior, content preservation, and motion realism, with ablations on direction construction and loss design.
展开完整摘要收起摘要↓
Existing human motion diffusion methods provide strong motion generation quality, and recent style transfer models can inject target style cues, but fine-grained continuous control of style intensity remains underexplored. In production, style intensity is subjective across artists and directors, so the practical requirement is not a universal absolute unit, but a reliable monotonic control axis. We propose Motion Style Slider, a motion-to-motion style transfer framework for endpoint-supervised continuous control. Given a content motion and a style motion, we construct a style direction in a learned motion-style embedding space and condition diffusion generation with a scalar intensity. The training objective combines diffusion denoising with latent intensity regularization to encourage smooth and monotonic style scaling without requiring intermediate-intensity ground-truth motions. Our framework is compatible with pretrained motion diffusion backbones and supports heterogeneous style datasets, including the multi-actor style motion dataset. To test out-of-range usability, we additionally introduce a small real-capture over-reaction extension and evaluate large-intensity behavior against these unseen targets. Experiments measure controllability, interpolation/extrapolation behavior, content preservation, and motion realism, with ablations on direction construction and loss design.
作者Zhangyi Wang, Jiaxu Liu, Chen Song, Chao Xu, Shengze Cai
In many safety-critical applications, control of uncertain dynamical systems relies on observers that estimate states and external disturbances. Neural network observers can improve estimation accuracy, but certifying their Lyapunov stability via Linear Matrix Inequality (LMI) constraints leads to large-scale semidefinite programs (SDPs) that are difficult to solve for large networks. To overcome this scalability bottleneck, we propose a novel two-stage training framework for provably stable neural network observers. Our approach decouples the optimization into a point-guided Lyapunov pre-training phase, which rapidly achieves high estimation accuracy and local stability over sampled states, followed by an LMI fine-tuning phase that efficiently satisfies a strict global Lyapunov stability certificate. We provide formal theoretical guarantees for local stability radii and probabilistic coverage over a prescribed compact error-state domain under specified regularity and sampling assumptions. Experiments on nonlinear control benchmarks and X-29 aircraft ablations show that our LMI-certified neural network observers train significantly faster than direct LMI-based methods and generalize robustly across diverse systems, achieving improved tracking accuracy over a range of observer baselines. The code is available at https://github.com/Berry-Myon/LearningNeuralNetworkObserver.
展开完整摘要收起摘要↓
In many safety-critical applications, control of uncertain dynamical systems relies on observers that estimate states and external disturbances. Neural network observers can improve estimation accuracy, but certifying their Lyapunov stability via Linear Matrix Inequality (LMI) constraints leads to large-scale semidefinite programs (SDPs) that are difficult to solve for large networks. To overcome this scalability bottleneck, we propose a novel two-stage training framework for provably stable neural network observers. Our approach decouples the optimization into a point-guided Lyapunov pre-training phase, which rapidly achieves high estimation accuracy and local stability over sampled states, followed by an LMI fine-tuning phase that efficiently satisfies a strict global Lyapunov stability certificate. We provide formal theoretical guarantees for local stability radii and probabilistic coverage over a prescribed compact error-state domain under specified regularity and sampling assumptions. Experiments on nonlinear control benchmarks and X-29 aircraft ablations show that our LMI-certified neural network observers train significantly faster than direct LMI-based methods and generalize robustly across diverse systems, achieving improved tracking accuracy over a range of observer baselines. The code is available at https://github.com/Berry-Myon/LearningNeuralNetworkObserver.
We introduce a framework for retraction-based convex optimization on Riemannian manifolds, which includes a notion of retraction-specific convex sets and retraction-based gradient projection algorithms. The standard theory of gradient projection algorithms generalizes easily to this framework. Within this framework, we establish convergence results for retraction-based gradient projection algorithms with various stepsize rules. As an application, we use our framework to study the weighted low-rank approximation. We also provide numerical validation of our convergence results on the image completion task.
展开完整摘要收起摘要↓
We introduce a framework for retraction-based convex optimization on Riemannian manifolds, which includes a notion of retraction-specific convex sets and retraction-based gradient projection algorithms. The standard theory of gradient projection algorithms generalizes easily to this framework. Within this framework, we establish convergence results for retraction-based gradient projection algorithms with various stepsize rules. As an application, we use our framework to study the weighted low-rank approximation. We also provide numerical validation of our convergence results on the image completion task.
作者Kai-Chen Tung, Qi Wu, David Bauer, Mengjiao Han, Silvio Rizzi, Kwan-Liu Ma
Implicit neural representations (INRs) have gained momentum in scientific visualization due to their compactness and scalability to large datasets, making them well suited for integration with direct volume rendering (DVR). However, real-time volume rendering of INR with advanced illumination effects, such as shadows, remains computationally expensive, as evaluating shadow terms via ray marching is costly. Alternatively, precomputing and storing shadows for many lighting directions is prohibitive in both memory and storage. To address this, we introduce a diffusion-based shadow caching framework that compresses a vast set of pre-calculated shadow INRs into a single diffusion model. Rather than focusing on generalizing to unseen directions, our method effectively memorizes and reconstructs a dense set of pre-trained lighting conditions on the fly. We first encode a collection of shadow coefficient volumes as shadow INRs, and then train a diffusion model conditioned on lighting direction to predict the corresponding shadow INR weights at inference time. This design integrates directly with standard INR renderers without additional runtime sampling. Experiments show that our approach achieves faster rendering than traditional methods while bypassing the massive storage bloat of independent INRs, producing shadows that closely match most of the reference results.
展开完整摘要收起摘要↓
Implicit neural representations (INRs) have gained momentum in scientific visualization due to their compactness and scalability to large datasets, making them well suited for integration with direct volume rendering (DVR). However, real-time volume rendering of INR with advanced illumination effects, such as shadows, remains computationally expensive, as evaluating shadow terms via ray marching is costly. Alternatively, precomputing and storing shadows for many lighting directions is prohibitive in both memory and storage. To address this, we introduce a diffusion-based shadow caching framework that compresses a vast set of pre-calculated shadow INRs into a single diffusion model. Rather than focusing on generalizing to unseen directions, our method effectively memorizes and reconstructs a dense set of pre-trained lighting conditions on the fly. We first encode a collection of shadow coefficient volumes as shadow INRs, and then train a diffusion model conditioned on lighting direction to predict the corresponding shadow INR weights at inference time. This design integrates directly with standard INR renderers without additional runtime sampling. Experiments show that our approach achieves faster rendering than traditional methods while bypassing the massive storage bloat of independent INRs, producing shadows that closely match most of the reference results.
General-purpose robot models increasingly rely on large and diverse datasets. For embodied 3D navigation, however, existing data sources face a fundamental trade-off: simulated data can be generated at scale but often suffer from the visual sim-to-real gap, whereas real-world flight data provide realistic observations but are costly to collect. This paper studies another direction: the use of high-fidelity visual generative models as scalable data engines for embodied 3D navigation. We introduce NavGen, a text-to-video data generation pipeline that produces diverse vision-language navigation (VLN) episodes across indoor and outdoor scenes. We also propose a style-diversification method that scales up long-tail data that are difficult and costly to collect. The resulting dataset contains approximately 400K navigation episodes. We evaluate our dataset against existing UAV navigation datasets across multiple metrics, and find that the model trained on our data generally improves with scale, outperforming those trained on existing datasets. To validate real-world transferability, we deploy the trained model in world-action-model paradigm to real-world flying experiments. The final model achieves a 75% success rate across different navigation tasks and environments.
展开完整摘要收起摘要↓
General-purpose robot models increasingly rely on large and diverse datasets. For embodied 3D navigation, however, existing data sources face a fundamental trade-off: simulated data can be generated at scale but often suffer from the visual sim-to-real gap, whereas real-world flight data provide realistic observations but are costly to collect. This paper studies another direction: the use of high-fidelity visual generative models as scalable data engines for embodied 3D navigation. We introduce NavGen, a text-to-video data generation pipeline that produces diverse vision-language navigation (VLN) episodes across indoor and outdoor scenes. We also propose a style-diversification method that scales up long-tail data that are difficult and costly to collect. The resulting dataset contains approximately 400K navigation episodes. We evaluate our dataset against existing UAV navigation datasets across multiple metrics, and find that the model trained on our data generally improves with scale, outperforming those trained on existing datasets. To validate real-world transferability, we deploy the trained model in world-action-model paradigm to real-world flying experiments. The final model achieves a 75% success rate across different navigation tasks and environments.
作者Antoine Lorentz, Stéphane May, Valentine Bellet, Dawa Derksen, Bastien Nespoulous
Large-scale Digital Surface Models (DSMs) can be produced cost-effectively from satellite images via stereo-photogrammetry. However, the resulting 3D maps are often contaminated by noise, outliers, and voids. On the other hand, aerial LiDAR provides high-accuracy elevation measurements at a substantially higher cost. In this work, we study diffusion models conditioned both on photogrammetric DSMs and Pléiades imagery to refine vertically co-registered DSMs. We introduce a modified Stable Diffusion 3 architecture with a pruned text stream and a patch-wise normalization strategy, enabling stable training on LiDAR data and transfer from natural images to elevation maps. Experiments in French cities demonstrate that multimodal conditioning improves elevation accuracy, reducing Dense Urban RMSE from 6.00 to 3.45 m in the in-context cities and from 4.16 to 2.77 m in the held-out city of Bordeaux.
展开完整摘要收起摘要↓
Large-scale Digital Surface Models (DSMs) can be produced cost-effectively from satellite images via stereo-photogrammetry. However, the resulting 3D maps are often contaminated by noise, outliers, and voids. On the other hand, aerial LiDAR provides high-accuracy elevation measurements at a substantially higher cost. In this work, we study diffusion models conditioned both on photogrammetric DSMs and Pléiades imagery to refine vertically co-registered DSMs. We introduce a modified Stable Diffusion 3 architecture with a pruned text stream and a patch-wise normalization strategy, enabling stable training on LiDAR data and transfer from natural images to elevation maps. Experiments in French cities demonstrate that multimodal conditioning improves elevation accuracy, reducing Dense Urban RMSE from 6.00 to 3.45 m in the in-context cities and from 4.16 to 2.77 m in the held-out city of Bordeaux.
Outliers are essential for evaluating and improving the robustness of machine learning systems, especially when future distributions may differ significantly from historical training data. In high-stakes applications, robustness often depends on rare cases that finite datasets fail to capture, making simple resampling or perturbation insufficient for stress scenario generation. Existing outlier synthesis methods typically rely on sparse neighborhoods, low support latent regions, or classifier boundary crossings, which can be heuristic, unstable, and tied to specific modalities or architectures. We therefore propose Sinkhorn Boundary Outlier Generation (SBOG), a structured framework for latent-space outlier generation that couples Sinkhorn optimal transport geometry with distributionally robust boundary modeling. The resulting Sinkhorn-induced support cost guides the sampler toward weakly supported boundary regions, while semantic constraints prevent uncontrolled drift from the intended context, yielding controlled deviations from the in-distribution reference measure rather than arbitrary sparse-region samples. Experiments on time series anomaly generation and image outlier synthesis show that our framework produces informative, semantically controlled outliers and improves downstream robustness evaluation across modalities, providing a foundation for stress scenario generation beyond empirical support.
展开完整摘要收起摘要↓
Outliers are essential for evaluating and improving the robustness of machine learning systems, especially when future distributions may differ significantly from historical training data. In high-stakes applications, robustness often depends on rare cases that finite datasets fail to capture, making simple resampling or perturbation insufficient for stress scenario generation. Existing outlier synthesis methods typically rely on sparse neighborhoods, low support latent regions, or classifier boundary crossings, which can be heuristic, unstable, and tied to specific modalities or architectures. We therefore propose Sinkhorn Boundary Outlier Generation (SBOG), a structured framework for latent-space outlier generation that couples Sinkhorn optimal transport geometry with distributionally robust boundary modeling. The resulting Sinkhorn-induced support cost guides the sampler toward weakly supported boundary regions, while semantic constraints prevent uncontrolled drift from the intended context, yielding controlled deviations from the in-distribution reference measure rather than arbitrary sparse-region samples. Experiments on time series anomaly generation and image outlier synthesis show that our framework produces informative, semantically controlled outliers and improves downstream robustness evaluation across modalities, providing a foundation for stress scenario generation beyond empirical support.
Pocket-conditioned molecular diffusion updates ambient atom coordinates, but many lead-optimization objectives are expressed on quotient features such as distances, contacts, and anchored substructures. We introduce budgeted quotient-residual guidance (QRG), an inference-time correction that makes these quotient objectives active without retraining the molecular generator. QRG lifts quotient covectors to metric-horizontal ambient directions and delivers them through a trust budget set by the frozen sampler's own step norm: quotient geometry chooses the direction, while sampler motion bounds the scale. We derive the horizontal lift, closed-form sampler-budget update, KL/kinetic interpretation around a frozen reverse step, equivariance conditions, and a product-budget split for budget-capped section and residual controls. Controlled quotient tasks confirm that sampler-relative delivery activates signals that raw local quotient gradients leave dormant. On frozen TargetDiff backbones, official seed-0 CBGBench ligand-generation/editing sweeps show practical quality-runtime gains: Local-QRG improves validity from 0.815 to 0.864 on fragment growing, 0.664 to 0.707 on scaffold hopping, and 0.681 to 0.712 on linker design, while PredNext-QRG improves fragment/scaffold and remains near-neutral on linker. Novelty remains 1.000 and diversity is preserved in the matched multi-seed molecular slice, giving task-dependent improvements without sampler retraining or backbone modification. Overall, QRG provides a lightweight route to quotient-aware inference for frozen molecular samplers with explicit runtime accounting.
展开完整摘要收起摘要↓
Pocket-conditioned molecular diffusion updates ambient atom coordinates, but many lead-optimization objectives are expressed on quotient features such as distances, contacts, and anchored substructures. We introduce budgeted quotient-residual guidance (QRG), an inference-time correction that makes these quotient objectives active without retraining the molecular generator. QRG lifts quotient covectors to metric-horizontal ambient directions and delivers them through a trust budget set by the frozen sampler's own step norm: quotient geometry chooses the direction, while sampler motion bounds the scale. We derive the horizontal lift, closed-form sampler-budget update, KL/kinetic interpretation around a frozen reverse step, equivariance conditions, and a product-budget split for budget-capped section and residual controls. Controlled quotient tasks confirm that sampler-relative delivery activates signals that raw local quotient gradients leave dormant. On frozen TargetDiff backbones, official seed-0 CBGBench ligand-generation/editing sweeps show practical quality-runtime gains: Local-QRG improves validity from 0.815 to 0.864 on fragment growing, 0.664 to 0.707 on scaffold hopping, and 0.681 to 0.712 on linker design, while PredNext-QRG improves fragment/scaffold and remains near-neutral on linker. Novelty remains 1.000 and diversity is preserved in the matched multi-seed molecular slice, giving task-dependent improvements without sampler retraining or backbone modification. Overall, QRG provides a lightweight route to quotient-aware inference for frozen molecular samplers with explicit runtime accounting.