💡 This paper proposes a novel approach called CapVector to improve the performance of vision-language-action models. The problem addressed is that pre-trained models often fail to improve performance and reduce adaptation costs during standard supervised finetuning. Advanced finetuning methods with auxiliary training objectives can improve performance but incur significant computational overhead.

The proposed method decouples the auxiliary training objectives from standard supervised finetuning to enhance model capabilities while reducing computational overhead. This is achieved by training the model to converge on a small-scale task set using two distinct training strategies, resulting in two finetuned models. The parameters difference between the two models is interpreted as capability vectors provided by auxiliary objectives. These vectors are then merged with pre-trained parameters to form a capability-enhanced meta model.

The method also uses a lightweight orthogonal regularization loss to augment standard supervised finetuning, which reduces computational overhead. The results show that the capability vectors are effective and versatile across diverse models, and can generalize to novel environments and embodiments without additional training. The proposed approach achieves performance comparable to auxiliary finetuned baselines with reduced computational overhead, making it a promising solution for improving vision-language-action models.

💡 This paper introduces a new approach called rStar that improves the reasoning capabilities of small language models without requiring fine-tuning or larger models. The problem addressed is that small language models often struggle with complex reasoning tasks, which can limit their ability to solve problems. The rStar method involves a self-play mutual generation-discrimination process, where one small language model generates reasoning trajectories using a Monte Carlo Tree Search with human-like reasoning actions, and another similar model acts as a discriminator to verify these trajectories. The trajectories that are mutually agreed upon are considered more likely to be correct. The results show that rStar can effectively solve diverse reasoning problems, including math and strategy-based tasks, and significantly improves the accuracy of small language models. For example, rStar boosts the accuracy of one model from 12.51 percent to 63.91 percent on a specific task, and from 36.46 percent to 81.88 percent on another model. Overall, the rStar approach makes smaller language models stronger problem-solvers without requiring additional training or larger models.

💡 The paper introduces MiniCPM-o 4.5, a model that enables real-time full-duplex multimodal interaction, allowing it to see, listen, and speak simultaneously in real-time. The current state of multimodal large language models has limitations, including separated perception and response phases and reactive behavior, which prevent them from incorporating new inputs for timely adjustments during generation. To address these issues, the authors propose Omni-Flow, a unified streaming framework that aligns omni-modal inputs and outputs along a shared temporal axis, converting conventional turn-based interaction into a full-duplex, time-aligned process. This enables simultaneous perception and response and allows proactive behavior to arise within the same framework. MiniCPM-o 4.5 has 9B parameters and achieves state-of-the-art open-source performance, surpassing other models in omni-modal understanding and speech generation while delivering better computation efficiency. The model can perform real-time full-duplex omni-modal interaction on edge devices with less than 12GB RAM cost, making it a significant step towards human-like multimodal interaction. The key contributions of the paper are the introduction of Omni-Flow and the development of MiniCPM-o 4.5, which mitigates the gaps in current multimodal interaction models and enables real-time full-duplex omni-modal interaction. The results show that MiniCPM-o 4.5 approaches the performance of other models, such as Gemini 2.5 Flash, and surpasses Qwen3-Omni-30B-A3B in omni-modal understanding and speech generation, demonstrating its effectiveness and efficiency.

💡 The paper introduces SenseNova-U1, a unified multimodal model that integrates understanding and generation into a single process, overcoming the traditional divide between these two tasks. Current large vision-language models treat understanding and generation as separate problems, leading to fragmented architectures and misaligned representation spaces. The authors argue that this divide hinders the emergence of native multimodal intelligence and propose a new paradigm, NEO-unify, which views understanding and generation as synergistic aspects of a single process.

The authors present two variants of SenseNova-U1, built on dense and mixture-of-experts understanding baselines, and demonstrate their performance across various tasks, including text understanding, vision-language perception, knowledge reasoning, agentic decision-making, and spatial intelligence. The models also excel in image synthesis, infographic generation, and interleaved vision-language generation, showing strong semantic consistency and visual fidelity.

The paper provides detailed information on model design, data preprocessing, pre- and post-training, and inference strategies, supporting community research. The results show that SenseNova-U1 models perform strongly in vision-language-action and world model scenarios, indicating a broader roadmap where models can think and act across modalities in a native manner. The authors conclude that multimodal AI should focus on building a unified system, rather than connecting separate systems, allowing necessary capabilities to emerge from within. Overall, the paper contributes to the development of unified multimodal models that can integrate understanding and generation, paving the way for more advanced and native multimodal intelligence.

💡 The paper proposes a lightweight memory mechanism called delta-mem to enhance large language models by providing a compact online state of associative memory. The problem addressed is the need for large language models to accumulate and reuse historical information in long-term assistants and agent systems, which is challenging due to the high cost of expanding the context window and ineffective context utilization.

The proposed method, delta-mem, augments a frozen full-attention backbone with a compact online state that compresses past information into a fixed-size state matrix updated by delta-rule learning. This online state is used to generate low-rank corrections to the backbone's attention computation during generation, allowing for efficient online memory.

The results show that delta-mem improves the average score of the frozen backbone and achieves larger gains on memory-heavy benchmarks, such as MemoryAgentBench and LoCoMo, while preserving general capabilities. Notably, delta-mem achieves these results with only an 8x8 online memory state, demonstrating that effective memory can be realized through a compact online state directly coupled with attention computation, without requiring full fine-tuning, backbone replacement, or explicit context extension. Overall, the paper contributes a novel and efficient approach to enhancing large language models with online memory, which has the potential to improve performance in a range of applications.

💡 The paper World Action Models The Next Frontier in Embodied AI presents a comprehensive survey of the emerging field of World Action Models, which aims to unify predictive state modeling with action generation for embodied policy learning. The problem addressed is that current AI models, such as Vision-Language-Action models, learn reactive observation-to-action mappings without explicitly modeling how the physical world evolves under intervention. To address this limitation, the authors introduce the concept of World Action Models, which targets a joint distribution over future states and actions rather than actions alone.

The method involves integrating world models, predictive models of environment dynamics, into the action generation pipeline. The authors formally define World Action Models and disambiguate them from related concepts, and provide a structured taxonomy of existing methods, including Cascaded and Joint World Action Models. They also analyze the data ecosystem fueling World Action Models development, including robot teleoperation, human demonstrations, simulation, and internet-scale egocentric video.

The results of the survey provide a systematic account of the World Action Models landscape, clarifying key architectural paradigms and their trade-offs. The authors identify open challenges and future opportunities for this rapidly evolving field, including the need for unified conceptual frameworks, evaluation protocols, and further research on the integration of world models and action generation. Overall, the paper contributes to the development of a cohesive framework for understanding environment dynamics and action prediction, and provides a foundation for future research in embodied AI.

💡 The paper introduces AlphaGRPO, a novel framework that enhances multimodal generation capabilities in unified multimodal models. The problem addressed is the need for improved multimodal generation without requiring an additional cold-start stage. To solve this, the authors apply Group Relative Policy Optimization to AR-Diffusion Unified Multimodal Models, enabling self-reflective refinement and decompositional verifiable reward mechanisms.

The method involves using Decompositional Verifiable Reward, which decomposes complex user requests into atomic, verifiable semantic and quality questions. These questions are then evaluated by a general multimodal language model to provide reliable and interpretable feedback. This approach allows the model to perform advanced reasoning tasks, including reasoning text-to-image generation and self-reflective refinement.

The results show that AlphaGRPO yields robust improvements across multimodal generation benchmarks, including GenEval, TIIF-Bench, DPG-Bench, and WISE. The framework also achieves significant gains in editing tasks on GEdit without training on editing tasks. The experiments demonstrate that the self-reflective reinforcement approach effectively leverages inherent understanding to guide high-fidelity generation, validating the effectiveness of AlphaGRPO. Overall, the paper contributes to the development of more advanced and reliable multimodal generation models.

💡 The paper provides a comprehensive survey of world models for robot learning, which are predictive representations of environmental dynamics that support policy learning, planning, and simulation. The authors note that the literature on world models is fragmented across different architectures, functional roles, and application domains, making it difficult to understand the current state of the field. To address this gap, the authors present a systematic review of world models from a robot learning perspective, examining how they are coupled with robot policies, used as learned simulators for reinforcement learning and evaluation, and have progressed in terms of robotic video world models. The survey covers the progression of world models from imagination-based generation to controllable, structured, and foundation-scale formulations, and connects these ideas to navigation and autonomous driving. The authors also summarize representative datasets, benchmarks, and evaluation protocols, and highlight major challenges and future directions for predictive modeling in embodied agents. The paper aims to clarify key paradigms and applications of world models, and to facilitate continued access to newly emerging works, benchmarks, and resources, the authors will maintain and regularly update a GitHub repository accompanying the survey. Overall, the paper provides a thorough overview of the rapidly growing literature on world models for robot learning, and identifies key areas for future research and development.

💡 The paper proposes a new framework called Latent-to-Pixel transfer paradigm, or L2P, which allows for efficient creation of pixel-space models using pre-trained latent diffusion models. The problem addressed is that training advanced pixel-space models from scratch requires significant computational and data resources. To solve this, L2P harnesses the knowledge of pre-trained latent diffusion models to build powerful pixel-space models with minimal training overhead.

The method involves discarding the Variational Autoencoder in favor of large-patch tokenization and freezing the intermediate layers of the pre-trained latent diffusion model. Only the shallow layers are trained to learn the latent-to-pixel transformation, using synthetic images generated by the pre-trained model as the training data. This approach enables rapid convergence without the need for real data collection.

The results show that L2P achieves negligible training overhead while performing on par with the source latent diffusion model. The framework is able to migrate massive latent priors to the pixel space using only 8 GPUs, and it unlocks native 4K ultra-high resolution generation by eliminating the Variational Autoencoder memory bottleneck. Extensive experiments demonstrate that L2P reaches 93 percent performance on GenEval and performs well on DPG-Bench, making it a promising approach for efficient pixel-space model creation.

💡 The paper introduces a new approach to language models called Multi-Stream LLMs, which aims to overcome the limitations of traditional language models that process information in a single stream of computation. The current models function on a message exchange format, where they successively exchange messages with users, systems, and tools, leading to limitations such as the inability to act while reading, react to new information while writing, think while reading or acting, and act while thinking.

To address these limitations, the authors propose a method that involves transitioning from sequential message-based instruction-tuning to parallel stream processing, enabling simultaneous reading and generation across multiple concurrent data flows. This is achieved by splitting each role into a separate stream, allowing the language model to simultaneously read from multiple input streams and generate tokens in multiple output streams, all of which causally depend on earlier timesteps.

The results of this approach show that it can improve model efficiency through parallelization, enhance model security through better separation of concerns, and increase model monitorability. The authors argue that this data-driven change can remedy the usability limitations of traditional language models, making them more efficient, secure, and transparent. Overall, the paper contributes to the development of more advanced language models that can process information in a more parallel and efficient manner, unlocking their potential for widespread use in various applications.

💡 The paper addresses the challenges of multimodal deep search, which requires an agent to solve open-world problems by combining search, tool use, and visual reasoning. Current systems have two major limitations: they treat images as transient outputs and cannot reuse intermediate visual evidence, and they rely on fixed curation recipes for training data that do not adapt to the agent's evolving capabilities.

To overcome these limitations, the authors introduce a visual-native agent harness with an image bank reference protocol, which allows images to be registered as addressable references and reused by later tools. They also propose On-policy Data Evolution, a closed-loop data generator that refines itself across rounds based on the policy being trained. This approach enables the generation of targeted training data that adapts to the agent's current needs.

The authors evaluate their approach on eight multimodal deep search benchmarks and demonstrate significant improvements in performance. With the On-policy Data Evolution method, the Qwen3-VL-8B agent achieves an average score of 39.0%, surpassing the Gemini-2.5 Pro agent. Further analysis shows that the image bank reuse is particularly effective for complex tasks that require iterative visual refinement, and that the rollout-feedback evolution yields more grounded and policy-matched reinforcement learning tasks.

The contributions of the paper are twofold: first, the introduction of a visual-native agent harness that enables reusable intermediate visual evidence, and second, the development of On-policy Data Evolution, a method for generating targeted training data that adapts to the agent's evolving capabilities. The results demonstrate the effectiveness of these contributions in improving multimodal deep search performance.