Two Mixture-of-Experts (MoE) forward passes can share every weight yet route the same token through different experts. This creates a possible blind spot in same-weight self-distillation, where a demonstration-conditioned teacher supervises a query-only student. We study this mismatch in its single-step form, with frozen weights rather than as a proxy for a full training trajectory. An exact blockwise decomposition separates a routing term, which changes gates at fixed content, from a dense-like content term. Across seven open-weight checkpoints and two domains, the routing term spans only $1.6\times$ as a fraction of block output, while its residual-stream exposure spans $3.2\times$. Exposure is ordered by the routed block's share of the residual. Scaling the always-on backbone in two confirmatory models moves exposure monotonically; common-mode controls support a mass-and-coherence mechanism rather than denominator dilution alone. Preregistered PubMedQA patches on three models show that the full routing term moves outputs by less than half the natural context effect and is largely reproduced by matched-norm noise, whereas the content term is strongly direction-specific. Scale and merged-expert probes show that the narrow block-level range is not universal, although exposure remains small at the tested boundaries. Router movement alone is therefore not evidence of behavioral influence: measure exposure first, and use a behavioral intervention when the decision matters.
展开完整摘要收起摘要↓
Two Mixture-of-Experts (MoE) forward passes can share every weight yet route the same token through different experts. This creates a possible blind spot in same-weight self-distillation, where a demonstration-conditioned teacher supervises a query-only student. We study this mismatch in its single-step form, with frozen weights rather than as a proxy for a full training trajectory. An exact blockwise decomposition separates a routing term, which changes gates at fixed content, from a dense-like content term. Across seven open-weight checkpoints and two domains, the routing term spans only $1.6\times$ as a fraction of block output, while its residual-stream exposure spans $3.2\times$. Exposure is ordered by the routed block's share of the residual. Scaling the always-on backbone in two confirmatory models moves exposure monotonically; common-mode controls support a mass-and-coherence mechanism rather than denominator dilution alone. Preregistered PubMedQA patches on three models show that the full routing term moves outputs by less than half the natural context effect and is largely reproduced by matched-norm noise, whereas the content term is strongly direction-specific. Scale and merged-expert probes show that the narrow block-level range is not universal, although exposure remains small at the tested boundaries. Router movement alone is therefore not evidence of behavioral influence: measure exposure first, and use a behavioral intervention when the decision matters.
Learning to generate or reconstruct explorable worlds requires video paired with more than RGB: camera motion, scene geometry, temporal correspondence and, for interactive models, control signals. Real capture can provide some of these signals, but dense geometry and long-range correspondence usually rely on estimation or specialised instrumentation. Rendering provides these quantities directly, yet existing synthetic resources rarely combine them on the same frames while also supporting controlled changes of viewpoint and appearance. We introduce WorldRover, a data engine for generating richly annotated, long-range explorations of artist-built environments. At its core, WorldRover-Engine is an Unreal Engine pipeline that executes and offline-renders minute-scale routes while preserving their full trajectories and scene geometry. The same exploration can be replayed from first-person, third-person, and 360 panoramic cameras under different environmental states. Using WorldRover-Engine, we construct WorldRover-10M, whose sequences pair RGB with metric depth, camera trajectories, and trajectory-derived action signals throughout each exploration. Third-person subsets additionally provide dense optical flow, long-range 2D/3D point tracks with visibility, and a character trajectory distinct from the camera trajectory. The engine can render a traversal from first-person, third-person and 360 panoramic viewpoints, under different environmental states or with a neutral white material, while preserving the route and scene geometry. WorldRover therefore turns long-horizon world exploration into a scalable data-generation problem, providing supervision for models that must build, maintain, and revisit coherent representations of an explorable world.
展开完整摘要收起摘要↓
Learning to generate or reconstruct explorable worlds requires video paired with more than RGB: camera motion, scene geometry, temporal correspondence and, for interactive models, control signals. Real capture can provide some of these signals, but dense geometry and long-range correspondence usually rely on estimation or specialised instrumentation. Rendering provides these quantities directly, yet existing synthetic resources rarely combine them on the same frames while also supporting controlled changes of viewpoint and appearance. We introduce WorldRover, a data engine for generating richly annotated, long-range explorations of artist-built environments. At its core, WorldRover-Engine is an Unreal Engine pipeline that executes and offline-renders minute-scale routes while preserving their full trajectories and scene geometry. The same exploration can be replayed from first-person, third-person, and 360 panoramic cameras under different environmental states. Using WorldRover-Engine, we construct WorldRover-10M, whose sequences pair RGB with metric depth, camera trajectories, and trajectory-derived action signals throughout each exploration. Third-person subsets additionally provide dense optical flow, long-range 2D/3D point tracks with visibility, and a character trajectory distinct from the camera trajectory. The engine can render a traversal from first-person, third-person and 360 panoramic viewpoints, under different environmental states or with a neutral white material, while preserving the route and scene geometry. WorldRover therefore turns long-horizon world exploration into a scalable data-generation problem, providing supervision for models that must build, maintain, and revisit coherent representations of an explorable world.
作者Kareem Hassani, Chaymaa Abbas, Hadi Al Mubasher, Mariette Awad
Conditional image-to-image generators are single-shot: they map input features to an output in one forward pass and treat it as final, with no opportunity to improve on it. Although trained to produce the best possible result in one step, such a model leaves room for improvement if it can adaptively revise its own output over iterations. We propose Recursive Refinement via Feedback Conditioning (RRFC), a novel feedback-conditioning framework for iterative output refinement that teaches a model to adaptively revise its output by conditioning on a new signal, namely its most recent previous prediction, which is fed back as an auxiliary set of channels alongside the original input. This preserves the generator's core architecture while modifying its conditioning interface and, depending on the model family, its training or inference procedure, so RRFC can be attached to existing generators without redesign. We evaluate RRFC across six baselines spanning adversarial, equilibrium, and diffusion-based models and three paired image-to-image translation tasks. Across 18 architecture-task settings, RRFC yields seven Holm-corrected improvements, seven degradations, and four non-significant changes. The gains concentrate on reconstruction-fidelity and identity settings, while five of the seven degradations fall on the single semantic-layout task, where every model declines. These results indicate that feedback-based refinement helps when its objective overlaps with the evaluated property, and that its gains concentrate on the tasks where that overlap holds.
展开完整摘要收起摘要↓
Conditional image-to-image generators are single-shot: they map input features to an output in one forward pass and treat it as final, with no opportunity to improve on it. Although trained to produce the best possible result in one step, such a model leaves room for improvement if it can adaptively revise its own output over iterations. We propose Recursive Refinement via Feedback Conditioning (RRFC), a novel feedback-conditioning framework for iterative output refinement that teaches a model to adaptively revise its output by conditioning on a new signal, namely its most recent previous prediction, which is fed back as an auxiliary set of channels alongside the original input. This preserves the generator's core architecture while modifying its conditioning interface and, depending on the model family, its training or inference procedure, so RRFC can be attached to existing generators without redesign. We evaluate RRFC across six baselines spanning adversarial, equilibrium, and diffusion-based models and three paired image-to-image translation tasks. Across 18 architecture-task settings, RRFC yields seven Holm-corrected improvements, seven degradations, and four non-significant changes. The gains concentrate on reconstruction-fidelity and identity settings, while five of the seven degradations fall on the single semantic-layout task, where every model declines. These results indicate that feedback-based refinement helps when its objective overlaps with the evaluated property, and that its gains concentrate on the tasks where that overlap holds.
作者Ivan Mikheev, Viacheslav Vasilev, Anna Dmitrienko, Alexey Letunovskiy, Ivan Kirillov, Kirill Chernyshev, Denis Dimitrov
Text-to-audio-video (T2AV) generation models produce a video and its soundtrack from a textual description, but offer no control over whose voice speaks in the output. We show that a base T2AV model can be turned into a voice-cloning model by adding a single zero-initialized linear layer on top of its audio backbone, fine-tuning for a comparatively short training schedule, and conditioning on a short reference recording at inference time. The reference is injected through two complementary signals: its diffusion latents are prepended to the audio stream, and a global speaker embedding modulates token of the target audio. On a benchmark of 674 speaker-text pairs spanning 30 speakers we compare against five strong voice-cloning text-to-speech baselines: our enhanced 5B model attains the highest speaker-encoder cosine similarity (SECS) across three independent verification networks (ECAPA-TDNN, WavLM-SV, Resemblyzer), statistically significantly outperforming every baseline. A side product of the architecture is that the audio path can be evaluated without the video path at inference time, yielding a ~30x speed-up over the full audio-video diffusion loop while preserving the voice-cloning behaviour.
展开完整摘要收起摘要↓
Text-to-audio-video (T2AV) generation models produce a video and its soundtrack from a textual description, but offer no control over whose voice speaks in the output. We show that a base T2AV model can be turned into a voice-cloning model by adding a single zero-initialized linear layer on top of its audio backbone, fine-tuning for a comparatively short training schedule, and conditioning on a short reference recording at inference time. The reference is injected through two complementary signals: its diffusion latents are prepended to the audio stream, and a global speaker embedding modulates token of the target audio. On a benchmark of 674 speaker-text pairs spanning 30 speakers we compare against five strong voice-cloning text-to-speech baselines: our enhanced 5B model attains the highest speaker-encoder cosine similarity (SECS) across three independent verification networks (ECAPA-TDNN, WavLM-SV, Resemblyzer), statistically significantly outperforming every baseline. A side product of the architecture is that the audio path can be evaluated without the video path at inference time, yielding a ~30x speed-up over the full audio-video diffusion loop while preserving the voice-cloning behaviour.
Multimodal large language models increasingly use visual chain-of-thought (Visual CoT) to reason about spatial, temporal, and embodied environments. By generating intermediate reasoning images, Visual CoT provides an intuitive mechanism for visual foresight but introduces substantial inference overhead, which is particularly problematic for proactive video reasoning. We ask whether models can learn to think visually during training while reasoning directly at inference. We introduce Internalized Visual Thinking (IVT), a post-training framework that jointly optimizes textual prediction and next-embedding prediction over unlabeled videos. Given a partially observed video, IVT predicts latent representations of future frames together with the target textual answer, encouraging the model to capture motion, object transitions, interactions, and latent intent. At inference, IVT generates the answer directly without synthesizing or re-encoding future frames. We conduct controlled studies across target representations, decoder designs, prediction horizons, data mixtures, training curricula, and predictive objectives. IVT improves over direct-answer fine-tuning on all six evaluation settings while retaining the same inference pathway. Compared with explicit Visual CoT, IVT achieves comparable or better performance and reduces average end-to-end latency by more than 5x. Together, our findings suggest that explicit pixel-space generation at inference time, as used in visual chain-of-thought, may not be necessary for effective proactive video reasoning. Predictive world modeling can be internalized during training to produce multimodal reasoners that are both more accurate and substantially more efficient.
展开完整摘要收起摘要↓
Multimodal large language models increasingly use visual chain-of-thought (Visual CoT) to reason about spatial, temporal, and embodied environments. By generating intermediate reasoning images, Visual CoT provides an intuitive mechanism for visual foresight but introduces substantial inference overhead, which is particularly problematic for proactive video reasoning. We ask whether models can learn to think visually during training while reasoning directly at inference. We introduce Internalized Visual Thinking (IVT), a post-training framework that jointly optimizes textual prediction and next-embedding prediction over unlabeled videos. Given a partially observed video, IVT predicts latent representations of future frames together with the target textual answer, encouraging the model to capture motion, object transitions, interactions, and latent intent. At inference, IVT generates the answer directly without synthesizing or re-encoding future frames. We conduct controlled studies across target representations, decoder designs, prediction horizons, data mixtures, training curricula, and predictive objectives. IVT improves over direct-answer fine-tuning on all six evaluation settings while retaining the same inference pathway. Compared with explicit Visual CoT, IVT achieves comparable or better performance and reduces average end-to-end latency by more than 5x. Together, our findings suggest that explicit pixel-space generation at inference time, as used in visual chain-of-thought, may not be necessary for effective proactive video reasoning. Predictive world modeling can be internalized during training to produce multimodal reasoners that are both more accurate and substantially more efficient.
Automatic video dubbing in the wild remains fundamentally limited by two competing constraints: hierarchical methods depend on brittle, multi-stage preprocessing pipelines that severely restrict data scalability and practical deployment, while holistic approaches operating on uncropped video suffer from weak temporal alignment and speaker-utterance ambiguity in multi-speaker settings. To overcome these limitations, we propose CineDub, a unified diffusion-based model that achieves precise multi-speaker dialogue dubbing directly from uncropped videos, without face cropping or speaker diarization. Central to our approach is the Implicitly-Coupled Holistic Conditioning (ICHC) paradigm, where holistic visual representations and a semantic-bundled transcription format are encoded independently, yet implicitly coupled through cross-modal training to resolve speaker ambiguity and enable precise multi-speaker multi-turn dialogue dubbing. Building on the unified temporal cues captured by holistic visual features, we further extend CineDub to joint speech and audio generation. We introduce an Ambient-to-Linguistic Curriculum Learning (ALC) to mitigate sub-task degradation, and a decoupled textual branch control mechanism to resolve cross-prompt interference during simultaneous generation. We also release two in-the-wild benchmarks, CineDub-Multi for multi-speaker dialogue dubbing and CineDub-SA for video-to-speech-and-audio (V2SA) generation, to enable evaluation under realistic conditions. Experiments show that CineDub achieves state-of-the-art results on established single-speaker dubbing and video-to-audio benchmarks while excelling in multi-speaker dialogue dubbing and acoustically coherent joint generation.
展开完整摘要收起摘要↓
Automatic video dubbing in the wild remains fundamentally limited by two competing constraints: hierarchical methods depend on brittle, multi-stage preprocessing pipelines that severely restrict data scalability and practical deployment, while holistic approaches operating on uncropped video suffer from weak temporal alignment and speaker-utterance ambiguity in multi-speaker settings. To overcome these limitations, we propose CineDub, a unified diffusion-based model that achieves precise multi-speaker dialogue dubbing directly from uncropped videos, without face cropping or speaker diarization. Central to our approach is the Implicitly-Coupled Holistic Conditioning (ICHC) paradigm, where holistic visual representations and a semantic-bundled transcription format are encoded independently, yet implicitly coupled through cross-modal training to resolve speaker ambiguity and enable precise multi-speaker multi-turn dialogue dubbing. Building on the unified temporal cues captured by holistic visual features, we further extend CineDub to joint speech and audio generation. We introduce an Ambient-to-Linguistic Curriculum Learning (ALC) to mitigate sub-task degradation, and a decoupled textual branch control mechanism to resolve cross-prompt interference during simultaneous generation. We also release two in-the-wild benchmarks, CineDub-Multi for multi-speaker dialogue dubbing and CineDub-SA for video-to-speech-and-audio (V2SA) generation, to enable evaluation under realistic conditions. Experiments show that CineDub achieves state-of-the-art results on established single-speaker dubbing and video-to-audio benchmarks while excelling in multi-speaker dialogue dubbing and acoustically coherent joint generation.
作者Genying Li, Boda Lin, Jiachen Li, Zijian Jia, Haojie Zheng, Yiming Wang, Shuchen Weng, Si Li
Music-driven dance video synthesis aims to animate a reference person according to a given music clip. The task is challenging because it requires a model to jointly learn music-to-motion correspondence, identity-preserving human animation, temporal coherence, and visually realistic video generation. We present FlowDance, a music-driven dance video generation framework that integrates explicit motion modeling with reference-preserving visual synthesis through parallel pose and RGB streams. We further introduce timestep-aware pose injection to adapt structural guidance across denoising steps and persistent identity injection to preserve the reference appearance over long video. To support this task, we further build a popularity-curated, high-resolution in-the-wild dance video dataset with synchronized music, RGB videos, 3D body motion, camera parameters, and projected 2D pose annotations. Extensive experiments show that FlowDance achieves strong performance in both dance motion generation and music-driven dance video synthesis.
展开完整摘要收起摘要↓
Music-driven dance video synthesis aims to animate a reference person according to a given music clip. The task is challenging because it requires a model to jointly learn music-to-motion correspondence, identity-preserving human animation, temporal coherence, and visually realistic video generation. We present FlowDance, a music-driven dance video generation framework that integrates explicit motion modeling with reference-preserving visual synthesis through parallel pose and RGB streams. We further introduce timestep-aware pose injection to adapt structural guidance across denoising steps and persistent identity injection to preserve the reference appearance over long video. To support this task, we further build a popularity-curated, high-resolution in-the-wild dance video dataset with synchronized music, RGB videos, 3D body motion, camera parameters, and projected 2D pose annotations. Extensive experiments show that FlowDance achieves strong performance in both dance motion generation and music-driven dance video synthesis.
Video models are increasingly used to predict what happens next in a scene, yet the metrics commonly used to compare their outputs say little about whether the predicted objects move correctly. Motion, geometry, identity, background stability, and visual similarity can fail independently, but whole-frame scores often mix these errors together. We introduce RigidBench, a simulator-grounded benchmark that compares a generated continuation with a reference rollout from the same initial frame and motion description. Its five rigid-body tasks vary objects, materials, viewpoints, and indoor and outdoor scenes, with per-frame masks, depth, 6-DoF trajectories, and contacts available for scoring. We evaluate eight models on the same 100 examples with ten measurements that keep these aspects separate. The resulting rankings depend strongly on what is measured: no model leads on all ten, and across model means, higher SSIM accompanies larger 3D trajectory error (r = 0.89). RigidBench also includes 5,000 training videos with exact simulator state, which we use to fine-tune and analyze Wan 2.2 TI2V-5B. Full fine-tuning reduces 3D trajectory error by about 20% with almost no change in SSIM, while teacher-forced probes and targeted interventions show that object position is represented throughout Wan's diffusion transformer and used by its denoising computation.
展开完整摘要收起摘要↓
Video models are increasingly used to predict what happens next in a scene, yet the metrics commonly used to compare their outputs say little about whether the predicted objects move correctly. Motion, geometry, identity, background stability, and visual similarity can fail independently, but whole-frame scores often mix these errors together. We introduce RigidBench, a simulator-grounded benchmark that compares a generated continuation with a reference rollout from the same initial frame and motion description. Its five rigid-body tasks vary objects, materials, viewpoints, and indoor and outdoor scenes, with per-frame masks, depth, 6-DoF trajectories, and contacts available for scoring. We evaluate eight models on the same 100 examples with ten measurements that keep these aspects separate. The resulting rankings depend strongly on what is measured: no model leads on all ten, and across model means, higher SSIM accompanies larger 3D trajectory error (r = 0.89). RigidBench also includes 5,000 training videos with exact simulator state, which we use to fine-tune and analyze Wan 2.2 TI2V-5B. Full fine-tuning reduces 3D trajectory error by about 20% with almost no change in SSIM, while teacher-forced probes and targeted interventions show that object position is represented throughout Wan's diffusion transformer and used by its denoising computation.
作者Yufeng Chi, Huimin Ma, Fan Gao, Zhice Niu, Keqin Li, Jianmin Li
While Text-to-Image (T2I) diffusion models have achieved remarkable success, precise spatial and orientational control in multi-object scenes remains a persistent challenge. Existing methods either rely on computationally expensive dense 3D maps or suffer from severe attribute leakage and "cut-and-paste" artifacts. To address these limitations, we propose PoseAdapter, a lightweight framework for high-fidelity 2.5D controllable image generation. Instead of dense spatial maps, it establishes precise spatial-angular anchors using an efficient condition layout: individual object captions, 2D bounding boxes, and 3D angles. To resolve the generative trade-off between strict instance isolation and global coherence, we introduce a Context-Aware Dual-Stream Representation. By injecting local object tokens and relation-enriched scene tokens into the visual stream of modern MM-DiT architectures via parallel masked and unmasked pathways, PoseAdapter eliminates attribute leakage while preserving natural inter-object relationships and scene-level coherence. To support this paradigm, we construct OrientLayout, a high-quality dataset featuring standardized 2.5D annotations and instance-level decoupled semantics. Extensive experiments demonstrate that PoseAdapter outperforms state-of-the-art baselines in spatial accuracy, orientational precision, and multi-object visual fidelity. Code and dataset will be available at https://github.com/cyf23/PoseAdapter.
展开完整摘要收起摘要↓
While Text-to-Image (T2I) diffusion models have achieved remarkable success, precise spatial and orientational control in multi-object scenes remains a persistent challenge. Existing methods either rely on computationally expensive dense 3D maps or suffer from severe attribute leakage and "cut-and-paste" artifacts. To address these limitations, we propose PoseAdapter, a lightweight framework for high-fidelity 2.5D controllable image generation. Instead of dense spatial maps, it establishes precise spatial-angular anchors using an efficient condition layout: individual object captions, 2D bounding boxes, and 3D angles. To resolve the generative trade-off between strict instance isolation and global coherence, we introduce a Context-Aware Dual-Stream Representation. By injecting local object tokens and relation-enriched scene tokens into the visual stream of modern MM-DiT architectures via parallel masked and unmasked pathways, PoseAdapter eliminates attribute leakage while preserving natural inter-object relationships and scene-level coherence. To support this paradigm, we construct OrientLayout, a high-quality dataset featuring standardized 2.5D annotations and instance-level decoupled semantics. Extensive experiments demonstrate that PoseAdapter outperforms state-of-the-art baselines in spatial accuracy, orientational precision, and multi-object visual fidelity. Code and dataset will be available at https://github.com/cyf23/PoseAdapter.
作者Cedar Site Bai, Amber Yijia Zheng, Raymond A. Yeh, Brian Bullins
Machine unlearning (MU) aims to remove the influence of specific training data while preserving model utility. As the name suggests, MU can be viewed as the inverse of learning, using gradient-based updates to reduce the influence of a forget-set by counteracting the previously learned behavior. Recently, Muon, a gradient descent variant, has been introduced. Muon applies spectral magnitude normalization to encourage exploration of rare directions and demonstrates promising performance. Inspired by Muon, we adopt the spectral view for unlearning and propose Spectral Saliency Unlearning (SSU). SSU thresholds weak singular components and updates only those directions supported by a confident unlearning signal. We further provide theoretical justification for this thresholding approach from the perspective of the forgetting-retention trade-off. Experiments across image classifiers, diffusion models, and LLMs demonstrate SSU's effectiveness.
展开完整摘要收起摘要↓
Machine unlearning (MU) aims to remove the influence of specific training data while preserving model utility. As the name suggests, MU can be viewed as the inverse of learning, using gradient-based updates to reduce the influence of a forget-set by counteracting the previously learned behavior. Recently, Muon, a gradient descent variant, has been introduced. Muon applies spectral magnitude normalization to encourage exploration of rare directions and demonstrates promising performance. Inspired by Muon, we adopt the spectral view for unlearning and propose Spectral Saliency Unlearning (SSU). SSU thresholds weak singular components and updates only those directions supported by a confident unlearning signal. We further provide theoretical justification for this thresholding approach from the perspective of the forgetting-retention trade-off. Experiments across image classifiers, diffusion models, and LLMs demonstrate SSU's effectiveness.
Deciding whether a Sokoban puzzle is solvable is PSPACE-complete (Culberson, 1997): solutions can be exponentially long and there is no short certificate to check. Solvability is also a fragile property, since even a single misplaced wall can silently render an entire puzzle unsolvable. In this work, we show that a transformer-based discrete diffusion model trained purely on tile completion, with no access to solvers, rewards, or solvability labels, achieves a solvability rate of 77.4%, with 94.5% of the remaining failures rendered solvable by removing a single wall. In other words, a global, search-heavy property follows from a local training objective: trained only to fill in masked cells, the model inherits solvability it was never trained on. An autoregressive model factorizes as $p(c_k \mid c_1 \dots c_{k-1})$, meaning a fixed order, always conditioned on a prefix. Masked diffusion does not: it hides a random subset of cells and learns $p(c_k \mid \text{any subset})$, so at generation time it can reveal cells in any order, each one conditioned on everything already placed, wherever it sits on the board. A puzzle's difficulty comes from exactly this kind of non-local interaction, a decision in one part of the grid constraining what will work somewhere else entirely. A generator that is not locked into a single fixed order is therefore a better structural match for the problem than one that is. The training pipeline is adapted from MD4 (Shi et al., 2024) and the dataset is DeepMind's Boxoban (Guez et al., 2019). The trained model and instructions for generating puzzles are publicly available.
展开完整摘要收起摘要↓
Deciding whether a Sokoban puzzle is solvable is PSPACE-complete (Culberson, 1997): solutions can be exponentially long and there is no short certificate to check. Solvability is also a fragile property, since even a single misplaced wall can silently render an entire puzzle unsolvable. In this work, we show that a transformer-based discrete diffusion model trained purely on tile completion, with no access to solvers, rewards, or solvability labels, achieves a solvability rate of 77.4%, with 94.5% of the remaining failures rendered solvable by removing a single wall. In other words, a global, search-heavy property follows from a local training objective: trained only to fill in masked cells, the model inherits solvability it was never trained on. An autoregressive model factorizes as $p(c_k \mid c_1 \dots c_{k-1})$, meaning a fixed order, always conditioned on a prefix. Masked diffusion does not: it hides a random subset of cells and learns $p(c_k \mid \text{any subset})$, so at generation time it can reveal cells in any order, each one conditioned on everything already placed, wherever it sits on the board. A puzzle's difficulty comes from exactly this kind of non-local interaction, a decision in one part of the grid constraining what will work somewhere else entirely. A generator that is not locked into a single fixed order is therefore a better structural match for the problem than one that is. The training pipeline is adapted from MD4 (Shi et al., 2024) and the dataset is DeepMind's Boxoban (Guez et al., 2019). The trained model and instructions for generating puzzles are publicly available.
Current flow matching models learn to transport the source i.i.d. Gaussian noise into the target distribution of natural images, yet this source distribution carries no notion of spatial structure. Images however are fundamentally local since nearby pixels are strongly correlated. By sampling the noise independently, we hypothesize that models are implicitly encouraged to exploit less noisy neighbors as context during training, partially bypassing the need to properly learn the true local structure of images. The source distribution, in other words, works against the inductive bias of the image domain. To ameliorate this design discrepancy, we propose StructFlow which encodes spatial locality directly into the source by having the pixels within a small region share a common noise component. This structured source produces transport paths that are geometrically aligned with image regions - enabling properties that generic flow matching struggles to provide: fine-grained local editing that naturally respects boundaries, robust structure preservation, and smooth semantic interpolation between images. We show that these benefits also extend to large pre-trained models, demonstrating that StructFlow can even be incorporated through a lightweight post-training phase. Comprehensive experiments on multiple datasets, in unconditional, class and text-conditioned regimes, using different diffusion transformer architectures confirm that StructFlow not only offers competitive image generation quality, but also significantly improves localized controllable re-synthesis.
展开完整摘要收起摘要↓
Current flow matching models learn to transport the source i.i.d. Gaussian noise into the target distribution of natural images, yet this source distribution carries no notion of spatial structure. Images however are fundamentally local since nearby pixels are strongly correlated. By sampling the noise independently, we hypothesize that models are implicitly encouraged to exploit less noisy neighbors as context during training, partially bypassing the need to properly learn the true local structure of images. The source distribution, in other words, works against the inductive bias of the image domain. To ameliorate this design discrepancy, we propose StructFlow which encodes spatial locality directly into the source by having the pixels within a small region share a common noise component. This structured source produces transport paths that are geometrically aligned with image regions - enabling properties that generic flow matching struggles to provide: fine-grained local editing that naturally respects boundaries, robust structure preservation, and smooth semantic interpolation between images. We show that these benefits also extend to large pre-trained models, demonstrating that StructFlow can even be incorporated through a lightweight post-training phase. Comprehensive experiments on multiple datasets, in unconditional, class and text-conditioned regimes, using different diffusion transformer architectures confirm that StructFlow not only offers competitive image generation quality, but also significantly improves localized controllable re-synthesis.
作者Hongfu Huang, Yuzhe Li, Ao Xu, Bo Liu, Changrui Wang, Kan Tang, Ning Yang, Shengxian Liu, Hanyu Liu, Pengpeng Zhang, Linggang Zhu, Fengkai Liu, Yichen Lu, Tong Zhao, Naihua Miao, Jian Zhou, Zhimei Sun
Despite the powerful multi-scale modeling methods and high-throughput infrastructures established in the materials community, real material computation workflows remain fragmented and heavily manual, requiring researchers to constantly bridge software tools, data analysis, and intermediate decisions. This growing gap between methodological capability and practical execution highlights the need for a new kind of autonomous computational framework, one that can coordinate tools, knowledge, and workflows in a more unified and adaptive way. Here, we introduce ALKEMIE Agent, an agentic platform in which retrieval-augmented generation, a materials-computation knowledge base, registered skills, database-supported provenance, AI-assisted structure modeling, bounded task execution, tool-calling iteration, and error-diagnostic assistance are integrated within a traceable control loop. The capabilities of ALKEMIE Agent are demonstrated through applications including materials recommendation, structure modeling, phonon calculations, machine-learned interatomic potential training, LAMMPS simulations, Ab Initio Monte Carlo (AIMC) sampling, and active-learning-based materials screening. Finally, we outline the future directions and challenges for the development of agentic platforms for computational materials design.
展开完整摘要收起摘要↓
Despite the powerful multi-scale modeling methods and high-throughput infrastructures established in the materials community, real material computation workflows remain fragmented and heavily manual, requiring researchers to constantly bridge software tools, data analysis, and intermediate decisions. This growing gap between methodological capability and practical execution highlights the need for a new kind of autonomous computational framework, one that can coordinate tools, knowledge, and workflows in a more unified and adaptive way. Here, we introduce ALKEMIE Agent, an agentic platform in which retrieval-augmented generation, a materials-computation knowledge base, registered skills, database-supported provenance, AI-assisted structure modeling, bounded task execution, tool-calling iteration, and error-diagnostic assistance are integrated within a traceable control loop. The capabilities of ALKEMIE Agent are demonstrated through applications including materials recommendation, structure modeling, phonon calculations, machine-learned interatomic potential training, LAMMPS simulations, Ab Initio Monte Carlo (AIMC) sampling, and active-learning-based materials screening. Finally, we outline the future directions and challenges for the development of agentic platforms for computational materials design.
作者Yuqi Chen, Sixuan Li, Yunfeng Cai, Xueai Li, Ka Man Yan, Ying Li
Large language models can write fluent stories, but open-ended storytelling requires more than local fluency. In evolving world simulations and AI-native games, models must preserve facts, relationships, causal dependencies, and character states as the world changes. We introduce WSE-bench, a process benchmark that separately evaluates sustained generation, canonical coherence, and meaningful development in dynamic LLM storytelling. Generation Coverage records the proportion of planned narrative steps produced; Consistency tracks when canon breaks; and Richness measures how meaningfully branching, player-shaped trajectories develop. Across frontier models, Consistency and Richness do not form a smooth trade-off: their empirical Pareto frontier is non-concave, with several non-dominated intermediate configurations that no positive linear weighting can select. Added structure can enrich trajectories, but it does not uniformly improve coherence and may shorten them. Model scale chiefly improves sustained generation, without producing reliable gains in canonical coherence or meaningful development. These results show that sustained generation, canonical coherence, and meaningful development are distinct and sometimes competing capacities. WSE-bench makes those dynamics visible by extending narrative evaluation from finished stories to the processes that create them.
展开完整摘要收起摘要↓
Large language models can write fluent stories, but open-ended storytelling requires more than local fluency. In evolving world simulations and AI-native games, models must preserve facts, relationships, causal dependencies, and character states as the world changes. We introduce WSE-bench, a process benchmark that separately evaluates sustained generation, canonical coherence, and meaningful development in dynamic LLM storytelling. Generation Coverage records the proportion of planned narrative steps produced; Consistency tracks when canon breaks; and Richness measures how meaningfully branching, player-shaped trajectories develop. Across frontier models, Consistency and Richness do not form a smooth trade-off: their empirical Pareto frontier is non-concave, with several non-dominated intermediate configurations that no positive linear weighting can select. Added structure can enrich trajectories, but it does not uniformly improve coherence and may shorten them. Model scale chiefly improves sustained generation, without producing reliable gains in canonical coherence or meaningful development. These results show that sustained generation, canonical coherence, and meaningful development are distinct and sometimes competing capacities. WSE-bench makes those dynamics visible by extending narrative evaluation from finished stories to the processes that create them.
Agentic retrieval workflows produce query, retrieval, and stopping traces as a byproduct of answering questions. We study how these traces can adapt a deployed dense retriever to changing workflow distributions without new relevance labels, synthetic queries, or LLM judgments. We introduce Navigation-Informed Embeddings (NIE), a family of trace-derived objectives. NIE-Stop turns the stopping document into a soft positive; NIE-Path additionally uses preceding path documents as hard comparisons and imposes ordinal constraints with geometric decay. A BGE encoder adapted from retained source trajectories improves support Recall@20 on an independent target benchmark from 72.2 to 78.0 overall. NIE-Stop reaches 76.9 overall and 52.3 on long paths; NIE-Path raises long-path performance to 55.4, compared with 46.7 for the unadapted encoder. A shuffled-order control under the full path objective loses 3.2 points. Without public-benchmark training, the same adapter also improves nDCG@10 by 1.9 points on standard BEIR HotpotQA. NIE therefore provides a lightweight adaptation channel for settings where trajectories are already retained, with zero incremental labeling cost.
展开完整摘要收起摘要↓
Agentic retrieval workflows produce query, retrieval, and stopping traces as a byproduct of answering questions. We study how these traces can adapt a deployed dense retriever to changing workflow distributions without new relevance labels, synthetic queries, or LLM judgments. We introduce Navigation-Informed Embeddings (NIE), a family of trace-derived objectives. NIE-Stop turns the stopping document into a soft positive; NIE-Path additionally uses preceding path documents as hard comparisons and imposes ordinal constraints with geometric decay. A BGE encoder adapted from retained source trajectories improves support Recall@20 on an independent target benchmark from 72.2 to 78.0 overall. NIE-Stop reaches 76.9 overall and 52.3 on long paths; NIE-Path raises long-path performance to 55.4, compared with 46.7 for the unadapted encoder. A shuffled-order control under the full path objective loses 3.2 points. Without public-benchmark training, the same adapter also improves nDCG@10 by 1.9 points on standard BEIR HotpotQA. NIE therefore provides a lightweight adaptation channel for settings where trajectories are already retained, with zero incremental labeling cost.
作者Tu Nguyen, Duy Cao, Viet Nguyen, Phu Nguyen, Vy Le, Nguyen TK Nguyen, Tien N. Nguyen, Vu Nguyen
In automated web testing, generating test cases and performing testing using functionality descriptions in natural-language is crucial for improving efficacy. These tasks require such a testing agent to carry out tasks on the target application and generating tests autonomously. We introduce HxAgent, an iterative LLM-based planning agent with a proactive correction strategy. After each step, HxAgent reassesses the web state to determine the next action using (1) current observations, (2) short-term memory of past actions, and (3) long-term experience extracted from past (in)correct sequences of actions. HxAgent achieves 97.4% Exact-Match accuracy on MiniWoB++, comparable to the best baselines without human demonstrations and surpassing the recent WALT by 10.5%. On a dataset of 350 web tasks, it attains 83.8% Exact-Match and 91.8% Prefix-Match, exceeding WALT by 13.4%. On OnlineMind2Web, it further improves over WALT by 4.6%.
展开完整摘要收起摘要↓
In automated web testing, generating test cases and performing testing using functionality descriptions in natural-language is crucial for improving efficacy. These tasks require such a testing agent to carry out tasks on the target application and generating tests autonomously. We introduce HxAgent, an iterative LLM-based planning agent with a proactive correction strategy. After each step, HxAgent reassesses the web state to determine the next action using (1) current observations, (2) short-term memory of past actions, and (3) long-term experience extracted from past (in)correct sequences of actions. HxAgent achieves 97.4% Exact-Match accuracy on MiniWoB++, comparable to the best baselines without human demonstrations and surpassing the recent WALT by 10.5%. On a dataset of 350 web tasks, it attains 83.8% Exact-Match and 91.8% Prefix-Match, exceeding WALT by 13.4%. On OnlineMind2Web, it further improves over WALT by 4.6%.
作者Pouya Ghiasnezhad Omran, Michael Zimmermann, Duncan Cambridge, Ashmita Kapoor, Tanya Dixit
Large Language Model (LLM) agents deployed in production environments face a fundamental tension: the agent's behavior is frozen at deployment time, while the business rules and edge cases it must handle continue to evolve. Existing approaches address agent construction and one-time evaluation but provide no structured mechanism for continuous post-deployment behavioral correction without modifying the agent's source code. Most of the approaches offered in the market, require intense collection of logs and traces, and re-examining the agent design by the engineering team, a process which is heavy, long and negates the economical value of agentic transformation. We introduce Agent Gym, a modular, domain-agnostic framework that wraps any existing LLM-based agent in a continuous evaluation-and-evolution loop. The framework provides six composable capabilities --- Act, Evaluate, Investigate, Correct, Learn, and Observe --- organized across three architectural zones: a constitution layer that codifies domain knowledge in configuration artifacts, a runtime inference pipeline that chains acting, investigation, and adaptive correction, and a learning loop that enables subject matter experts to discover and validate new correction rules through natural language interaction. The key technical contributions include a hybrid deterministic-LLM correction engine with 21 condition operators and three-tier actions, a three-layer investigation architecture for ground-truth-free compliance validation, and a programmatic safety loop that guarantees rule correctness before human approval. We further introduce the Spec-to-Note Gap, an autoencoder-inspired view of agentic system transparency. An open-source reference implementation for invoice processing demonstrates that the framework is fully operational and ready for adoption.
展开完整摘要收起摘要↓
Large Language Model (LLM) agents deployed in production environments face a fundamental tension: the agent's behavior is frozen at deployment time, while the business rules and edge cases it must handle continue to evolve. Existing approaches address agent construction and one-time evaluation but provide no structured mechanism for continuous post-deployment behavioral correction without modifying the agent's source code. Most of the approaches offered in the market, require intense collection of logs and traces, and re-examining the agent design by the engineering team, a process which is heavy, long and negates the economical value of agentic transformation. We introduce Agent Gym, a modular, domain-agnostic framework that wraps any existing LLM-based agent in a continuous evaluation-and-evolution loop. The framework provides six composable capabilities --- Act, Evaluate, Investigate, Correct, Learn, and Observe --- organized across three architectural zones: a constitution layer that codifies domain knowledge in configuration artifacts, a runtime inference pipeline that chains acting, investigation, and adaptive correction, and a learning loop that enables subject matter experts to discover and validate new correction rules through natural language interaction. The key technical contributions include a hybrid deterministic-LLM correction engine with 21 condition operators and three-tier actions, a three-layer investigation architecture for ground-truth-free compliance validation, and a programmatic safety loop that guarantees rule correctness before human approval. We further introduce the Spec-to-Note Gap, an autoencoder-inspired view of agentic system transparency. An open-source reference implementation for invoice processing demonstrates that the framework is fully operational and ready for adoption.
Industrial software-engineering teams increasingly need LLM agents that turn bug reports into correct patches, yet benchmark-scale operation adds long horizons, tool-use discipline, context persistence, heterogeneous clusters, and evaluation reuse. We present Kozuchi Agent, a language-agnostic open-weight repair agent and CI-operated evaluation pipeline. Explicit phases, persistent state, deterministic tools, a model-independent action interface, and cross-agent test-time selection make runs auditable and repeatable. With locally hosted Qwen3.5-27B, no fine-tuning, and TTS@8, Kozuchi resolves 374/500 SWE-bench Verified instances on the official evaluator. Unchanged on Multi-SWE-bench Java, the same 27-billion-parameter agent resolves 41/128 instances (32.03%), ranking first among strict open-weight submissions and fourth of 42 overall; on Python it ranks 12th of 135 and first among open-weight systems. Per-phase behavior remains within +/-5 percentage points across languages. Remaining failures mainly reflect semantic correctness, Java-specific harness issues, and selection errors. Across both tracks, results compare favorably with open/local peers by parameter count. Analysis of candidate diversity, selector regret, and patch reliability shows that the remaining gap is primarily semantic correctness and selection rather than edit formatting or proprietary-model access. Operationally, reusable CI stages reduce operator touch-points from five to one across heterogeneous internal clusters.
展开完整摘要收起摘要↓
Industrial software-engineering teams increasingly need LLM agents that turn bug reports into correct patches, yet benchmark-scale operation adds long horizons, tool-use discipline, context persistence, heterogeneous clusters, and evaluation reuse. We present Kozuchi Agent, a language-agnostic open-weight repair agent and CI-operated evaluation pipeline. Explicit phases, persistent state, deterministic tools, a model-independent action interface, and cross-agent test-time selection make runs auditable and repeatable. With locally hosted Qwen3.5-27B, no fine-tuning, and TTS@8, Kozuchi resolves 374/500 SWE-bench Verified instances on the official evaluator. Unchanged on Multi-SWE-bench Java, the same 27-billion-parameter agent resolves 41/128 instances (32.03%), ranking first among strict open-weight submissions and fourth of 42 overall; on Python it ranks 12th of 135 and first among open-weight systems. Per-phase behavior remains within +/-5 percentage points across languages. Remaining failures mainly reflect semantic correctness, Java-specific harness issues, and selection errors. Across both tracks, results compare favorably with open/local peers by parameter count. Analysis of candidate diversity, selector regret, and patch reliability shows that the remaining gap is primarily semantic correctness and selection rather than edit formatting or proprietary-model access. Operationally, reusable CI stages reduce operator touch-points from five to one across heterogeneous internal clusters.
AI-powered digital-avatar streamers in live e-commerce must answer product questions, engage viewers, and execute changing business strategies in real time. This requires low latency, factual and effective replies, and rapid adaptation to updated campaign, compliance, and style requirements. We develop an evolvable Harness that decouples Skills, Hooks, system prompts, and tools from model weights, allowing runtime behavior to change without retraining. However, Harness evolution creates a moving execution environment: compact models fine-tuned on one configuration may memorize names, schemas, and prompt templates rather than follow the Harness currently provided, while stronger zero-shot models are too slow for real-time use. We address this tension with Harness-Aware Training (HAT), which makes Harness states part of the training distribution. HAT applies task-preserving Harness-State Augmentation (HSA) to Skills, tool schemas, prompt structures, and interaction constraints, and comprises three stages: HSA-based supervised fine-tuning, general on-policy distillation to recover general capabilities, and HSA-based agentic reinforcement learning in a production-informed live-room simulator. Across four evaluation sets with more than 4,500 cases, our compact 35B model scores 94.8 on real-world Live-Stream QA, versus 80.3 for the base model and 93.0 for the strongest evaluated general LLM, while scoring 94.6 on Harness-Variant QA and retaining 83.5 on IFEval. By contrast, fixed-Harness SFT reduces IFEval by 7.7 points. In a controlled complete-agent replay on one NVIDIA H20 GPU with MTP enabled, the system achieves 3.407 s P50 and 8.114 s P95 latency. These results show that HAT produces a latency-feasible compact agent that remains effective under evaluated Harness changes without sacrificing general instruction following.
展开完整摘要收起摘要↓
AI-powered digital-avatar streamers in live e-commerce must answer product questions, engage viewers, and execute changing business strategies in real time. This requires low latency, factual and effective replies, and rapid adaptation to updated campaign, compliance, and style requirements. We develop an evolvable Harness that decouples Skills, Hooks, system prompts, and tools from model weights, allowing runtime behavior to change without retraining. However, Harness evolution creates a moving execution environment: compact models fine-tuned on one configuration may memorize names, schemas, and prompt templates rather than follow the Harness currently provided, while stronger zero-shot models are too slow for real-time use. We address this tension with Harness-Aware Training (HAT), which makes Harness states part of the training distribution. HAT applies task-preserving Harness-State Augmentation (HSA) to Skills, tool schemas, prompt structures, and interaction constraints, and comprises three stages: HSA-based supervised fine-tuning, general on-policy distillation to recover general capabilities, and HSA-based agentic reinforcement learning in a production-informed live-room simulator. Across four evaluation sets with more than 4,500 cases, our compact 35B model scores 94.8 on real-world Live-Stream QA, versus 80.3 for the base model and 93.0 for the strongest evaluated general LLM, while scoring 94.6 on Harness-Variant QA and retaining 83.5 on IFEval. By contrast, fixed-Harness SFT reduces IFEval by 7.7 points. In a controlled complete-agent replay on one NVIDIA H20 GPU with MTP enabled, the system achieves 3.407 s P50 and 8.114 s P95 latency. These results show that HAT produces a latency-feasible compact agent that remains effective under evaluated Harness changes without sacrificing general instruction following.
Tool-calling LLM agents navigate unfamiliar codebases with a handful of generic primitives for listing, reading and searching files (ls, cat, grep). A knowledge graph admits the same interface: listing neighbours, reading node content and searching descriptions are the same operations on a different substrate. Building on this correspondence, we present GRA, a Graph Reasoning Agent that explores hybrid knowledge graphs, whose nodes are either textual concepts or relational tables, with seven generic tools, discovering everything domain-specific at run time. On UFK-M (Unified Factory Knowledge Model), an industrial benchmark of 258 analytical questions whose gold answers are produced by executing validated SQL programs, GRA beats a full-context agent by 5.1 pp (88.4% vs. 83.3%), while reading under a third of its input tokens. A graph-free control shows the gain comes chiefly from selective agentic access rather than graph topology, and that the effect depends on a model able to drive tools reliably. Seeing less, the agent answers better: selective navigation over a structured substrate beats exhaustive context.
展开完整摘要收起摘要↓
Tool-calling LLM agents navigate unfamiliar codebases with a handful of generic primitives for listing, reading and searching files (ls, cat, grep). A knowledge graph admits the same interface: listing neighbours, reading node content and searching descriptions are the same operations on a different substrate. Building on this correspondence, we present GRA, a Graph Reasoning Agent that explores hybrid knowledge graphs, whose nodes are either textual concepts or relational tables, with seven generic tools, discovering everything domain-specific at run time. On UFK-M (Unified Factory Knowledge Model), an industrial benchmark of 258 analytical questions whose gold answers are produced by executing validated SQL programs, GRA beats a full-context agent by 5.1 pp (88.4% vs. 83.3%), while reading under a third of its input tokens. A graph-free control shows the gain comes chiefly from selective agentic access rather than graph topology, and that the effect depends on a model able to drive tools reliably. Seeing less, the agent answers better: selective navigation over a structured substrate beats exhaustive context.
作者Xiao Wang, Lu Dong, Ifeoma Nwogu, Srirangaraj Setlur, Venu Govindaraju
Programming small social robots from natural-language instructions requires more than invoking isolated APIs. Interactive tasks combine reactive physical behaviors with stateful social behaviors, while existing interfaces often require developers to manually compose APIs into skills, configure their parameters, bind sensor events to skills, and manage task states at runtime. We present MistyPilot, a multi-agent LLM framework that interprets high-level natural-language instructions and orchestrates the corresponding skills on the Misty social robot. A Task Router dispatches each instruction to one of two specialized agents: a Physically Interactive Agent for sensor-triggered robot control and direct skill invocation, and a Social Interaction Agent for dialogue-oriented task-state management and context-dependent multimodal response generation. To improve efficiency, the Social Interaction Agent reuses previously generated results when applicable and invokes full generation otherwise. We evaluate MistyPilot on five component-level suites, with sensor bindings and skill invocations executed on the physical Misty robot, and a preliminary user study with 12 participants. MistyPilot attains high accuracy on routing, sensor-skill binding, task-state parsing, result reuse, and skill extension up to 100 skills, and lower variance than an otherwise identical single-agent baseline, while participants report positive perceptions of usability and interaction quality. The code will be made publicly available via the project page.
展开完整摘要收起摘要↓
Programming small social robots from natural-language instructions requires more than invoking isolated APIs. Interactive tasks combine reactive physical behaviors with stateful social behaviors, while existing interfaces often require developers to manually compose APIs into skills, configure their parameters, bind sensor events to skills, and manage task states at runtime. We present MistyPilot, a multi-agent LLM framework that interprets high-level natural-language instructions and orchestrates the corresponding skills on the Misty social robot. A Task Router dispatches each instruction to one of two specialized agents: a Physically Interactive Agent for sensor-triggered robot control and direct skill invocation, and a Social Interaction Agent for dialogue-oriented task-state management and context-dependent multimodal response generation. To improve efficiency, the Social Interaction Agent reuses previously generated results when applicable and invokes full generation otherwise. We evaluate MistyPilot on five component-level suites, with sensor bindings and skill invocations executed on the physical Misty robot, and a preliminary user study with 12 participants. MistyPilot attains high accuracy on routing, sensor-skill binding, task-state parsing, result reuse, and skill extension up to 100 skills, and lower variance than an otherwise identical single-agent baseline, while participants report positive perceptions of usability and interaction quality. The code will be made publicly available via the project page.
Large language model (LLM) agents often perform poorly on complex, long-horizon tasks because their context becomes increasingly cluttered over time. As interactions accumulate, detailed execution traces and intermediate outputs dominate the context, making it difficult for the model to retain and use high-level planning information. Most existing methods address this issue through compression or retrieval applied to a single, flat context, which does not clearly separate different types of context information and often leads to degraded reasoning. To address this challenge, we propose HyMem, a hierarchical framework that explicitly separates the agent's context into distinct functional layers. HyMem organizes context by function to separate high-level planning from execution and complex analysis. Its isolated reasoning module handles complex subtasks without adding intermediate reasoning traces to the persistent planning context, while its memory management module preserves task progress across context refreshes through structured summaries. These components reduce redundant context accumulation, retain task-critical information, and support coherent long-horizon reasoning within a limited context window. Experiments on GAIA and Browsecomp-plus show that, with DeepSeek-V4, HyMem achieves average Pass@1 scores of 66.7% and 61.3%, outperforming the strongest baseline by 6.1 and 4.7 percentage points, respectively. Further analysis indicates that HyMem effectively controls the growth of the reasoning context, allowing the model to maintain focus and accuracy across complex, long-horizon tasks.
展开完整摘要收起摘要↓
Large language model (LLM) agents often perform poorly on complex, long-horizon tasks because their context becomes increasingly cluttered over time. As interactions accumulate, detailed execution traces and intermediate outputs dominate the context, making it difficult for the model to retain and use high-level planning information. Most existing methods address this issue through compression or retrieval applied to a single, flat context, which does not clearly separate different types of context information and often leads to degraded reasoning. To address this challenge, we propose HyMem, a hierarchical framework that explicitly separates the agent's context into distinct functional layers. HyMem organizes context by function to separate high-level planning from execution and complex analysis. Its isolated reasoning module handles complex subtasks without adding intermediate reasoning traces to the persistent planning context, while its memory management module preserves task progress across context refreshes through structured summaries. These components reduce redundant context accumulation, retain task-critical information, and support coherent long-horizon reasoning within a limited context window. Experiments on GAIA and Browsecomp-plus show that, with DeepSeek-V4, HyMem achieves average Pass@1 scores of 66.7% and 61.3%, outperforming the strongest baseline by 6.1 and 4.7 percentage points, respectively. Further analysis indicates that HyMem effectively controls the growth of the reasoning context, allowing the model to maintain focus and accuracy across complex, long-horizon tasks.
Foundation GUI agents can automate complex digital tasks, but deployment is hindered by scarce and biased training data, ambiguous prompts, and unreliable execution. Routine workflows rely on user-specific tools and tacit conventions, so unstated instructions can produce arbitrary variations across runs. We present UI-Mate, a foundation GUI agent that integrates an environment-grounded training stack with in-context demonstration learning. UI-Mate makes three contributions: A Scalable Environment-Grounded Training Stack: A closed-loop data engine automates task generation, environment construction, rollout, filtering, capability balancing, SFT, and online RL across massively parallel environments via unified task-verifier bundles. In-Context Demonstration Learning: A mechanism that transforms multimodal demonstrations into flexible subtask-level workflows, follows relevant demonstrated steps, and re-plans from the live interface. OSWorkerBench Benchmark and Insights: A benchmark of 100 long-horizon office tasks across 41 applications that supports instruction-only and demonstration-guided evaluation. Its demonstration resources separate a 33-task self-demo setting, built from successful strong-agent rollouts of the same targets, from a 45-task variant-demo setting, built from human recordings of related but non-identical tasks. Experiments show that UI-Mate-27B sets a new open-weight state of the art on general computer-use benchmarks, scoring 77.0% on OSWorld-Verified and 66.2% on WindowsAgentArena. On OSWorkerBench, it reaches 41.0% strict success and 76.9% progress, outperforming its Qwen3.6-27B base by 17.7 and 24.5 points. On the 33-task self-demo subset, one demonstration raises strict success from 17.2% to 35.4% and progress from 67.9% to 81.1%, substantially improving long-horizon reliability. Project page: https://ui-mate.github.io.
展开完整摘要收起摘要↓
Foundation GUI agents can automate complex digital tasks, but deployment is hindered by scarce and biased training data, ambiguous prompts, and unreliable execution. Routine workflows rely on user-specific tools and tacit conventions, so unstated instructions can produce arbitrary variations across runs. We present UI-Mate, a foundation GUI agent that integrates an environment-grounded training stack with in-context demonstration learning. UI-Mate makes three contributions: A Scalable Environment-Grounded Training Stack: A closed-loop data engine automates task generation, environment construction, rollout, filtering, capability balancing, SFT, and online RL across massively parallel environments via unified task-verifier bundles. In-Context Demonstration Learning: A mechanism that transforms multimodal demonstrations into flexible subtask-level workflows, follows relevant demonstrated steps, and re-plans from the live interface. OSWorkerBench Benchmark and Insights: A benchmark of 100 long-horizon office tasks across 41 applications that supports instruction-only and demonstration-guided evaluation. Its demonstration resources separate a 33-task self-demo setting, built from successful strong-agent rollouts of the same targets, from a 45-task variant-demo setting, built from human recordings of related but non-identical tasks. Experiments show that UI-Mate-27B sets a new open-weight state of the art on general computer-use benchmarks, scoring 77.0% on OSWorld-Verified and 66.2% on WindowsAgentArena. On OSWorkerBench, it reaches 41.0% strict success and 76.9% progress, outperforming its Qwen3.6-27B base by 17.7 and 24.5 points. On the 33-task self-demo subset, one demonstration raises strict success from 17.2% to 35.4% and progress from 67.9% to 81.1%, substantially improving long-horizon reliability. Project page: https://ui-mate.github.io.
Long-horizon, multi-agent language model (LM) simulations are widely proposed for studying social behavior, yet instruments to measure whether persona-conditioned agents maintain identity fidelity under sustained pressure are lacking. We present MicroVerse, a behavioral-science instrument that measures identity drift in generative agents. Agents carry an immutable "soul file" (core values, moral boundaries, personality, goals) and inhabit a resource-scarce 50 x 50 environment where water is a non-respawning survival constraint. Scarcity is operationalized via a per-tick existence-cost gradient. The eight-verb action space maps directly to moral boundaries (trade, talk, attack, scavenge). Using a three-layer memory architecture, agents periodically revise a mutable current identity against their immutable original soul via importance-triggered reflection. To mitigate survivor bias, MicroVerse decouples measurement from behavior using uniform longitudinal engine snapshots every N ticks alongside a forced-end snapshot of all living and dead agents. Identity drift is scored offline using a paraphrase-aware, value-anchored, multi-register diff rather than raw cosine similarity. We evaluate the instrument via a controlled seed run (n = 25) and a reflection-threshold sweep (thresholds {40, 80, 150}) to determine if drift dynamics are gate artifacts or threshold-robust properties. We report two primary findings: (1) Anti-self-deception emerges unprompted as the single largest semantic category of identity modification (27 of 111 added boundaries, 24%). (2) The system is threshold-robust; lower gates accelerate and increase revision frequency but preserve drift direction. All empirical results are strictly preliminary existence proofs and effect shapes (one model, one seed per arm, n = 25) rather than statistical significance claims.
展开完整摘要收起摘要↓
Long-horizon, multi-agent language model (LM) simulations are widely proposed for studying social behavior, yet instruments to measure whether persona-conditioned agents maintain identity fidelity under sustained pressure are lacking. We present MicroVerse, a behavioral-science instrument that measures identity drift in generative agents. Agents carry an immutable "soul file" (core values, moral boundaries, personality, goals) and inhabit a resource-scarce 50 x 50 environment where water is a non-respawning survival constraint. Scarcity is operationalized via a per-tick existence-cost gradient. The eight-verb action space maps directly to moral boundaries (trade, talk, attack, scavenge). Using a three-layer memory architecture, agents periodically revise a mutable current identity against their immutable original soul via importance-triggered reflection. To mitigate survivor bias, MicroVerse decouples measurement from behavior using uniform longitudinal engine snapshots every N ticks alongside a forced-end snapshot of all living and dead agents. Identity drift is scored offline using a paraphrase-aware, value-anchored, multi-register diff rather than raw cosine similarity. We evaluate the instrument via a controlled seed run (n = 25) and a reflection-threshold sweep (thresholds {40, 80, 150}) to determine if drift dynamics are gate artifacts or threshold-robust properties. We report two primary findings: (1) Anti-self-deception emerges unprompted as the single largest semantic category of identity modification (27 of 111 added boundaries, 24%). (2) The system is threshold-robust; lower gates accelerate and increase revision frequency but preserve drift direction. All empirical results are strictly preliminary existence proofs and effect shapes (one model, one seed per arm, n = 25) rather than statistical significance claims.
LLM-based agents can act on behalf of a user to access cloud services, call tools, or invoke agents. At session start, the agent's permissions are set but remain static, and each request is evaluated independently, without considering prior actions. Within its permissions, an agent may act contrary to the delegated task, combine individually permitted actions into a prohibited outcome, or delegate authority to a sub-agent without limiting it. A prompt injection poses a risk only if the agent has authority to perform such actions; this is therefore a problem of authorization architecture, not just the model. The Agentic Principal Chain (APC) tracks delegated authority from one principal to the next. APC evaluates each request against the accumulated session state using six authorization checks. APC carries forward and restricts delegated scope and budgets. Using composition closure, APC checks requests against prior actions to prevent prohibited combinations and enforces the decision outside the model. We prove Blast Radius Monotonicity and Composition Soundness for APC implementations; Composition Soundness is limited to prohibited combinations under a complete restriction set and serialized admission. We evaluated 3,154 instances including InjecAgent, AgentDojo, and ASB. Our compromised-model evaluation tests APC independently of model behavior by inserting the ground-truth attack call after the first legitimate tool call. AgentDojo exfiltration fell from 75-100% to 0% across all four domains; APC blocked all 544 InjecAgent data-stealing cases. Intent binding reduced destruction from 38.6% to 4.0% and manipulation from 90.5% to 12.1%. Authorization latency was 0.24 ms at the 99th percentile on an idle host; across 949 AgentDojo task-injection pairs, utility was 8.6 and 13.9 percentage points lower in the two settings. Implementation, evaluation tools, and data are publicly available.
展开完整摘要收起摘要↓
LLM-based agents can act on behalf of a user to access cloud services, call tools, or invoke agents. At session start, the agent's permissions are set but remain static, and each request is evaluated independently, without considering prior actions. Within its permissions, an agent may act contrary to the delegated task, combine individually permitted actions into a prohibited outcome, or delegate authority to a sub-agent without limiting it. A prompt injection poses a risk only if the agent has authority to perform such actions; this is therefore a problem of authorization architecture, not just the model. The Agentic Principal Chain (APC) tracks delegated authority from one principal to the next. APC evaluates each request against the accumulated session state using six authorization checks. APC carries forward and restricts delegated scope and budgets. Using composition closure, APC checks requests against prior actions to prevent prohibited combinations and enforces the decision outside the model. We prove Blast Radius Monotonicity and Composition Soundness for APC implementations; Composition Soundness is limited to prohibited combinations under a complete restriction set and serialized admission. We evaluated 3,154 instances including InjecAgent, AgentDojo, and ASB. Our compromised-model evaluation tests APC independently of model behavior by inserting the ground-truth attack call after the first legitimate tool call. AgentDojo exfiltration fell from 75-100% to 0% across all four domains; APC blocked all 544 InjecAgent data-stealing cases. Intent binding reduced destruction from 38.6% to 4.0% and manipulation from 90.5% to 12.1%. Authorization latency was 0.24 ms at the 99th percentile on an idle host; across 949 AgentDojo task-injection pairs, utility was 8.6 and 13.9 percentage points lower in the two settings. Implementation, evaluation tools, and data are publicly available.
作者Changwon Lee, Tak Hur, Jeongwoo Jae, Daniel K. Park
Decoding is an essential component of quantum error correction (QEC), translating stabilizer measurement outcomes into corrective actions that suppress logical errors and preserve logical quantum information. Building fault-tolerant architectures requires increasing the code distance, which in turn places growing demands on decoding accuracy, scalability, and practical deployability. While a wide range of decoding algorithms have been proposed and demonstrated, achieving reliable, scalable, and real-time decoding remains a significant challenge. Machine-learning (ML) approaches are particularly well suited to this setting, as quantum error decoding is fundamentally a problem of processing large volumes of classical data with complex spatiotemporal correlations. This chapter surveys ML-based methods for quantum error decoding, with a focus on topological codes and an emphasis on architectural principles, practical performance, and real-time considerations. We first frame decoding as a learning problem and outline key paradigms, including discriminative, generative, and reinforcement-learning formulations. We then introduce the neural network building blocks that underpin most contemporary neural decoders and discuss how these components can be integrated to balance expressivity, scalability, and latency. Building on this architectural perspective, we review recent progress and benchmarks in neural decoding for memory experiments, and discuss real-time decoding, open challenges, and future directions toward scalable fault-tolerant quantum computing.
展开完整摘要收起摘要↓
Decoding is an essential component of quantum error correction (QEC), translating stabilizer measurement outcomes into corrective actions that suppress logical errors and preserve logical quantum information. Building fault-tolerant architectures requires increasing the code distance, which in turn places growing demands on decoding accuracy, scalability, and practical deployability. While a wide range of decoding algorithms have been proposed and demonstrated, achieving reliable, scalable, and real-time decoding remains a significant challenge. Machine-learning (ML) approaches are particularly well suited to this setting, as quantum error decoding is fundamentally a problem of processing large volumes of classical data with complex spatiotemporal correlations. This chapter surveys ML-based methods for quantum error decoding, with a focus on topological codes and an emphasis on architectural principles, practical performance, and real-time considerations. We first frame decoding as a learning problem and outline key paradigms, including discriminative, generative, and reinforcement-learning formulations. We then introduce the neural network building blocks that underpin most contemporary neural decoders and discuss how these components can be integrated to balance expressivity, scalability, and latency. Building on this architectural perspective, we review recent progress and benchmarks in neural decoding for memory experiments, and discuss real-time decoding, open challenges, and future directions toward scalable fault-tolerant quantum computing.
作者Ao Zhou, Bo Dai, Le Yu, Xingyu Liu, Zeyu Hao, Lingkun Long, Chunming Hu, Jianlei Yang
Vision-Language-Action (VLA) models have emerged as a promising foundation for Embodied AI, but their high inference cost poses significant challenges for deployment in robotic systems. In practice, on-device inference is constrained by limited compute capacity and energy budgets, struggling to simultaneously satisfy real-time control and energy efficiency requirements. Alternatively, offloading the inference workload to an edge server is susceptible to fluctuations in system conditions, introducing unpredictable latency risks. Device-edge co-inference offers a promising solution, but systematic research tailored to VLA models remains scarce, particularly a unified co-inference framework that jointly addresses real-time constraints and system-level energy efficiency. Thus, we propose EcoVLA, an adaptive device-edge co-inference framework for VLA models that maximizes system energy efficiency under real-time constraints. EcoVLA first introduces a unified stage-level abstraction over different VLA paradigms, establishing an architecture-agnostic co-inference design space. It then formulates a joint device-edge-network latency and energy prediction model to enable rapid runtime evaluation of candidate co-inference schemes. Building on this, EcoVLA continuously selects the energy-optimal scheme satisfying real-time constraints with millisecond-level overhead, adapting to runtime variations in network and system states. Furthermore, EcoVLA incorporates a lightweight transmission mechanism for inter-stage intermediate tensors to reduce the communication overhead incurred by cross-device collaboration. Experimental results across VLA models show that EcoVLA improves system energy efficiency by up to 236% over existing co-inference approaches under a 20 Hz action output frequency constraint, while consistently maintaining SLO satisfaction under dynamic network and edge workload conditions.
展开完整摘要收起摘要↓
Vision-Language-Action (VLA) models have emerged as a promising foundation for Embodied AI, but their high inference cost poses significant challenges for deployment in robotic systems. In practice, on-device inference is constrained by limited compute capacity and energy budgets, struggling to simultaneously satisfy real-time control and energy efficiency requirements. Alternatively, offloading the inference workload to an edge server is susceptible to fluctuations in system conditions, introducing unpredictable latency risks. Device-edge co-inference offers a promising solution, but systematic research tailored to VLA models remains scarce, particularly a unified co-inference framework that jointly addresses real-time constraints and system-level energy efficiency. Thus, we propose EcoVLA, an adaptive device-edge co-inference framework for VLA models that maximizes system energy efficiency under real-time constraints. EcoVLA first introduces a unified stage-level abstraction over different VLA paradigms, establishing an architecture-agnostic co-inference design space. It then formulates a joint device-edge-network latency and energy prediction model to enable rapid runtime evaluation of candidate co-inference schemes. Building on this, EcoVLA continuously selects the energy-optimal scheme satisfying real-time constraints with millisecond-level overhead, adapting to runtime variations in network and system states. Furthermore, EcoVLA incorporates a lightweight transmission mechanism for inter-stage intermediate tensors to reduce the communication overhead incurred by cross-device collaboration. Experimental results across VLA models show that EcoVLA improves system energy efficiency by up to 236% over existing co-inference approaches under a 20 Hz action output frequency constraint, while consistently maintaining SLO satisfaction under dynamic network and edge workload conditions.
3D Gaussian Splatting (3DGS) enables real-time novel-view synthesis but remains limited on GPUs at high resolutions. Through a stage-wise Roofline characterization, we identify two distinct hardware bottlenecks: global memory traffic dominates the front end, whereas instruction throughput limits rasterization. Guided by this analysis, we develop RoofGS, a rendering framework that applies bottleneck-specific optimizations rather than generic kernel acceleration. For the memory-bound front end, we design a resolution-adaptive quantized depth sorting key that compresses each key to 32 bits. For the compute-bound rasterizer, we introduce a range-aware bit-level fast exponential approximation tailored to the bounded exponent range after opacity culling, with a derived per-pixel error bound. These two core techniques are complemented by additional optimizations (kernel fusion, compact attribute storage, culling, dual-pixel evaluation) that additionally reduce memory traffic and improve instruction-level parallelism. Experiments show that RoofGS achieves a 10.1$\times$ end-to-end speedup over 3DGS at 4K on an RTX 4090, increasing throughput from 61 to 616 FPS, with only a 0.028 dB PSNR loss.
展开完整摘要收起摘要↓
3D Gaussian Splatting (3DGS) enables real-time novel-view synthesis but remains limited on GPUs at high resolutions. Through a stage-wise Roofline characterization, we identify two distinct hardware bottlenecks: global memory traffic dominates the front end, whereas instruction throughput limits rasterization. Guided by this analysis, we develop RoofGS, a rendering framework that applies bottleneck-specific optimizations rather than generic kernel acceleration. For the memory-bound front end, we design a resolution-adaptive quantized depth sorting key that compresses each key to 32 bits. For the compute-bound rasterizer, we introduce a range-aware bit-level fast exponential approximation tailored to the bounded exponent range after opacity culling, with a derived per-pixel error bound. These two core techniques are complemented by additional optimizations (kernel fusion, compact attribute storage, culling, dual-pixel evaluation) that additionally reduce memory traffic and improve instruction-level parallelism. Experiments show that RoofGS achieves a 10.1$\times$ end-to-end speedup over 3DGS at 4K on an RTX 4090, increasing throughput from 61 to 616 FPS, with only a 0.028 dB PSNR loss.
Vision-Language-Action (VLA) models have demonstrated remarkable capabilities in the field of embodied AI, but their high computational cost and limited predicted action length hinder real-time deployment. Although Dadu-Corki, a dedicated accelerator for efficient embodied AI, has been introduced, it does not exploit the inherent interaction patterns between the robot and its environment, which results in a relatively short predicted action length. We observe that robotic environments naturally alternate between active states-where precise actions are crucial-and inactive states-where actions have limited impact on task success. This insight enables a new scheduling opportunity: long-action-length speculative prediction in inactive states, paired with selective verification in active states. We propose SpecVLA, an algorithm-system co-design framework that adaptively balances action length, inference latency, and task reliability. On the algorithm side, SpecVLA introduces a state-aware VLA inference execution paradigm and a hardware-friendly construction of a smaller verification model (sVLA) using differential residuals and block-wise mixed-precision quantization. On the system side, we develop a heterogeneous architecture consisting of a GPU and a robotic-specific hardware module, along with a speculative dataflow that decouples VLA and sVLA through parallel execution. Comprehensive evaluations on OpenVLA and RDT across LIBERO and ManiSkill benchmarks show that SpecVLA reduces end-to-end latency significantly while preserving task success rate. By enabling long-action-length speculative prediction with timely verification, SpecVLA achieves real-time robotic manipulation with both high efficiency and reliability.
展开完整摘要收起摘要↓
Vision-Language-Action (VLA) models have demonstrated remarkable capabilities in the field of embodied AI, but their high computational cost and limited predicted action length hinder real-time deployment. Although Dadu-Corki, a dedicated accelerator for efficient embodied AI, has been introduced, it does not exploit the inherent interaction patterns between the robot and its environment, which results in a relatively short predicted action length. We observe that robotic environments naturally alternate between active states-where precise actions are crucial-and inactive states-where actions have limited impact on task success. This insight enables a new scheduling opportunity: long-action-length speculative prediction in inactive states, paired with selective verification in active states. We propose SpecVLA, an algorithm-system co-design framework that adaptively balances action length, inference latency, and task reliability. On the algorithm side, SpecVLA introduces a state-aware VLA inference execution paradigm and a hardware-friendly construction of a smaller verification model (sVLA) using differential residuals and block-wise mixed-precision quantization. On the system side, we develop a heterogeneous architecture consisting of a GPU and a robotic-specific hardware module, along with a speculative dataflow that decouples VLA and sVLA through parallel execution. Comprehensive evaluations on OpenVLA and RDT across LIBERO and ManiSkill benchmarks show that SpecVLA reduces end-to-end latency significantly while preserving task success rate. By enabling long-action-length speculative prediction with timely verification, SpecVLA achieves real-time robotic manipulation with both high efficiency and reliability.
Runtime compression of serving state trades quality for capacity with no priced guarantee: systems adapt precision on load signals with no soundness statement, and certified approaches budget request-level risk by a union bound over a pre-declared event count. We show the union budget exhausts on every long request in a production serving stack (100% of requests), and replace it with an anytime-valid, physically accounted ledger whose bound holds at every one of 352,333 admission calls on live traffic and which, in a pre-registered held-out confirmatory round, halves the exact-fallback rate at matched risk (0.30 -> 0.14) — coverage is bought at a price the account states. We then price the remaining distance from the certified witness to what a user experiences: a machine-checked design law (TV <= tanh(a_q w_thr)) turns the served-TV target into a threshold knob, and a three-layer audit of its instantiation — an operator-norm query envelope measured 1.5x from tight, a measured-ellipsoid replacement for the Cauchy-Schwarz ball that buys nothing (0.89x, held-out sound), and the gate's operating point (~700x) — localizes the entire 1064x gap to the operating point, a price the law now states rather than an unknown. A priced bound is worth nothing on a request one has not seen, so the third link is the quantifier: exchangeable extrapolation across 80 serving histories replaces binary conformal prediction's vacuous certificates with order-statistic bounds that discriminate (0.41 against 0.51 calibration risk). All probabilistic kernels are Lean 4-checked (228 exported theorems, no sorry); which object deserves this machinery at all is settled empirically in a companion paper that adjudicates — and rejects — the natural alternative of certifying routing. What ships is an account: risk you can spend, a gap you can read off a law, and a bound that survives the request you have not seen.
展开完整摘要收起摘要↓
Runtime compression of serving state trades quality for capacity with no priced guarantee: systems adapt precision on load signals with no soundness statement, and certified approaches budget request-level risk by a union bound over a pre-declared event count. We show the union budget exhausts on every long request in a production serving stack (100% of requests), and replace it with an anytime-valid, physically accounted ledger whose bound holds at every one of 352,333 admission calls on live traffic and which, in a pre-registered held-out confirmatory round, halves the exact-fallback rate at matched risk (0.30 -> 0.14) — coverage is bought at a price the account states. We then price the remaining distance from the certified witness to what a user experiences: a machine-checked design law (TV <= tanh(a_q w_thr)) turns the served-TV target into a threshold knob, and a three-layer audit of its instantiation — an operator-norm query envelope measured 1.5x from tight, a measured-ellipsoid replacement for the Cauchy-Schwarz ball that buys nothing (0.89x, held-out sound), and the gate's operating point (~700x) — localizes the entire 1064x gap to the operating point, a price the law now states rather than an unknown. A priced bound is worth nothing on a request one has not seen, so the third link is the quantifier: exchangeable extrapolation across 80 serving histories replaces binary conformal prediction's vacuous certificates with order-statistic bounds that discriminate (0.41 against 0.51 calibration risk). All probabilistic kernels are Lean 4-checked (228 exported theorems, no sorry); which object deserves this machinery at all is settled empirically in a companion paper that adjudicates — and rejects — the natural alternative of certifying routing. What ships is an account: risk you can spend, a gap you can read off a law, and a bound that survives the request you have not seen.