Tongji University

Vision-language-action policies process hundreds of visual tokens per step, which drives up inference latency. APS-VLA keeps the backbone frozen and uses a lightweight scorer conditioned on the current observation, instruction and proprioception to pick tokens, training the retained representation both to reproduce policy outputs and to predict the subsequent visual change. On LIBERO with 128 tokens retained, mean success is 95.75%, 1.25 points below the unpruned 97.00%, and inference latency falls to 68.75 ms.

Vision-language-action policies process hundreds of visual tokens per step, which drives up inference latency. APS-VLA keeps the backbone frozen and uses a lightweight scorer conditioned on the current observation, instruction and proprioception to pick tokens, training the retained representation both to reproduce policy outputs and to predict the subsequent visual change. On LIBERO with 128 tokens retained, mean success is 95.75%, 1.25 points below the unpruned 97.00%, and inference latency falls to 68.75 ms.

Shanghai Jiao Tong University and collaborators assembled a roughly 47,000-hour multisource pool and extended a pretrained video model's language–video priors into language–video–action learning, yielding the world action model SWAM. On the 50 manipulation tasks of RoboTwin 2.0, SWAM reaches 61.56% success under the randomized Hard setting, 12.86 percentage points above the strongest baseline OpenWAM, and 88.10% under the clean Easy setting, slightly below OpenWAM's 89.30%. On a real bimanual platform, its three-task average Hard success rate is 40.00% against OpenWAM's 23.33%.

Shanghai Jiao Tong University and collaborators assembled a roughly 47,000-hour multisource pool and extended a pretrained video model's language–video priors into language–video–action learning, yielding the world action model SWAM. On the 50 manipulation tasks of RoboTwin 2.0, SWAM reaches 61.56% success under the randomized Hard setting, 12.86 percentage points above the strongest baseline OpenWAM, and 88.10% under the clean Easy setting, slightly below OpenWAM's 89.30%. On a real bimanual platform, its three-task average Hard success rate is 40.00% against OpenWAM's 23.33%.

Tool-using LLM agents turn model output into real actions, making indirect prompt injection an execution-security problem. SecureGARD inserts an executable ontology between planning and tool execution, grounding heterogeneous tool calls into shared semantic actions and judging each as ready, waiting, or forbidden against the current task state. Under benign and indirect-injection settings on AgentDojo, ASB, and AgentDyn, the observed attack success rate is 0% on both backbones, with higher task utility under attack than the compared baselines.

Tool-using LLM agents turn model output into real actions, making indirect prompt injection an execution-security problem. SecureGARD inserts an executable ontology between planning and tool execution, grounding heterogeneous tool calls into shared semantic actions and judging each as ready, waiting, or forbidden against the current task state. Under benign and indirect-injection settings on AgentDojo, ASB, and AgentDyn, the observed attack success rate is 0% on both backbones, with higher task utility under attack than the compared baselines.

Long-chain reasoning distillation no longer fixes its training questions up front: BoundaryEvo repeatedly has the student and a fixed expert answer the same candidates, keeping only questions the expert solves reliably and the student has not mastered, and with 1,000 training examples raises Qwen3.5-4B accuracy on IMO-AnswerBench from 62.9 to 75.3.

Long-chain reasoning distillation no longer fixes its training questions up front: BoundaryEvo repeatedly has the student and a fixed expert answer the same candidates, keeping only questions the expert solves reliably and the student has not mastered, and with 1,000 training examples raises Qwen3.5-4B accuracy on IMO-AnswerBench from 62.9 to 75.3.

A world action model can regenerate complete rollouts of previously learned tasks from their instructions alone, and after truncation and physical-consistency filtering, replaying them cuts the forgetting of sequential fine-tuning by over 90% on four LIBERO suites and five real-world tasks without keeping any data from previous tasks.

A world action model can regenerate complete rollouts of previously learned tasks from their instructions alone, and after truncation and physical-consistency filtering, replaying them cuts the forgetting of sequential fine-tuning by over 90% on four LIBERO suites and five real-world tasks without keeping any data from previous tasks.

WebLoop jointly optimizes generation, critique, and refinement as roles of one shared policy for functional Web generation: Generator and Refiner outputs receive direct browser-execution rewards, while the Critic never sees execution results and is trained from requirement-level execution labels (discriminability) plus the execution gain induced by subsequent refinements (helpfulness), using a staged reward that first establishes diagnosis and then adds consequence-aware credit. With Qwen3.5-9B, WebLoop reaches 41.5 Overall on WebRise and 38.9% accuracy on WebGen-Bench, improving the base model by 11.3 and 15.4 points.

WebLoop jointly optimizes generation, critique, and refinement as roles of one shared policy for functional Web generation: Generator and Refiner outputs receive direct browser-execution rewards, while the Critic never sees execution results and is trained from requirement-level execution labels (discriminability) plus the execution gain induced by subsequent refinements (helpfulness), using a staged reward that first establishes diagnosis and then adds consequence-aware credit. With Qwen3.5-9B, WebLoop reaches 41.5 Overall on WebRise and 38.9% accuracy on WebGen-Bench, improving the base model by 11.3 and 15.4 points.

The work introduces the principle of personalization diversity and PerFeCT-VAR, which decomposes each client's transition matrix into shared low-rank dynamics, shared sparse links, and personalized sparse departures, uses a cross-client frequency cap to derive a sharp shared-personalized identification threshold, and designs Frequency-Capped Thresholding (FCT) for federated estimation; the theory establishes joint linear convergence and statistical rates showing that under sufficient personalization diversity the shared dynamics retain total-sample-size gains while personalized components achieve client-level accuracy; simulations and a multi-store retail revenue application demonstrate predictive and interpretive benefits.

The work introduces the principle of personalization diversity and PerFeCT-VAR, which decomposes each client's transition matrix into shared low-rank dynamics, shared sparse links, and personalized sparse departures, uses a cross-client frequency cap to derive a sharp shared-personalized identification threshold, and designs Frequency-Capped Thresholding (FCT) for federated estimation; the theory establishes joint linear convergence and statistical rates showing that under sufficient personalization diversity the shared dynamics retain total-sample-size gains while personalized components achieve client-level accuracy; simulations and a multi-store retail revenue application demonstrate predictive and interpretive benefits.

The work introduces JEVal, a bilingual decision benchmark of 11,257 instances from 36 datasets across 10 domains, evaluates 25 configurations and finds that general decision models approach strong thinking LLMs when evidence is available but weaken on specialist knowledge, calibrated uncertainty, exact computation, and long-context evidence use, with local decision advantages offset by system-level errors in long-horizon agents and social simulation; the authors then build InnerJev-4B and InnerJev-27B from open-weight LLMs via first-token decision readout and Reasoning-to-Readout Self-Distillation, with InnerJev-27B reaching 79.24% on JEVal versus Jev's 78.38% and answering a typical query in about 0.1 s.

The work introduces JEVal, a bilingual decision benchmark of 11,257 instances from 36 datasets across 10 domains, evaluates 25 configurations and finds that general decision models approach strong thinking LLMs when evidence is available but weaken on specialist knowledge, calibrated uncertainty, exact computation, and long-context evidence use, with local decision advantages offset by system-level errors in long-horizon agents and social simulation; the authors then build InnerJev-4B and InnerJev-27B from open-weight LLMs via first-token decision readout and Reasoning-to-Readout Self-Distillation, with InnerJev-27B reaching 79.24% on JEVal versus Jev's 78.38% and answering a typical query in about 0.1 s.