💡 The paper introduces D-OPSD, a new training approach for diffusion models that enables efficient supervised fine-tuning while preserving few-step inference capabilities. The current landscape of high-performance image generation models is shifting from inefficient multi-step models to efficient few-step models, but these models are challenging to fine-tune using traditional techniques. The problem with traditional fine-tuning methods is that they compromise the model's inherent few-step inference capability.

To address this issue, the authors propose D-OPSD, which leverages on-policy self-distillation with text and multimodal features. The method works by making the model act as both the teacher and the student, where the student is conditioned only on the text feature, and the teacher is conditioned on the multimodal feature of both the text prompt and the target image. The training process minimizes the difference between the predicted distributions over the student's own roll-outs, allowing the model to learn new concepts and styles without sacrificing its original few-step capacity.

The key contribution of D-OPSD is that it enables on-policy learning during supervised fine-tuning, which allows the model to learn from its own trajectory and under its own supervision. This approach enables the model to inherit the in-context capabilities of its encoder, making it possible to fine-tune the model continuously without compromising its few-step inference capability. The results show that D-OPSD enables efficient supervised fine-tuning for diffusion models, making it a promising approach for high-performance image generation models.

💡 The paper presents a domain specific audit framework called MedSkillAudit for assessing the quality of medical research agent skills. The problem addressed is the need for reliable evaluation of these skills to ensure scientific integrity, methodological validity, reproducibility, and boundary safety in healthcare applications.

The authors developed MedSkillAudit, a layered framework that evaluates skill release readiness before deployment. They applied this framework to 75 skills across five medical research categories and compared the results with expert reviews. Two experts independently assigned quality scores and release dispositions to each skill, and the system expert agreement was quantified using statistical measures.

The results show that MedSkillAudit achieved a high level of agreement with expert reviews, with a mean consensus quality score of 72.4. The framework was able to identify skills that fell below the Limited Release threshold, with 57.3 percent of skills requiring further development. The system expert agreement was higher than the human inter-rater agreement, indicating that MedSkillAudit can provide a reliable and consistent evaluation of medical research agent skills.

The study found that MedSkillAudit performed well in certain categories, such as Protocol Design, but had a negative agreement in Academic Writing, suggesting a structural mismatch between the rubric and expert reviews. Overall, the paper concludes that domain specific pre-deployment audit frameworks like MedSkillAudit can provide a practical foundation for governing medical research agent skills, complementing general purpose quality checks with structured audit workflows tailored to scientific use cases.

The contributions of the paper are the development and evaluation of a domain specific audit framework for medical research agent skills, and the demonstration of its reliability and consistency in assessing the quality of these skills. The study provides a foundation for further research in this area and highlights the importance of domain specific evaluation frameworks for ensuring the quality and safety of AI systems in healthcare applications.

💡 The paper presents JoyAI-Image, a unified multimodal foundation model that integrates visual understanding, text-to-image generation, and instruction-guided image editing. The model combines a spatially enhanced Multimodal Large Language Model with a Multimodal Diffusion Transformer, allowing for a shared multimodal interface between perception and generation. The authors propose a scalable training recipe that incorporates unified instruction tuning, long-text rendering supervision, spatially grounded data, and general and spatial editing signals. This design enables the model to achieve broad multimodal capabilities while strengthening geometry-aware reasoning and controllable visual synthesis. The experiments demonstrate that JoyAI-Image achieves state-of-the-art or highly competitive performance across various benchmarks, including understanding, generation, long-text rendering, and editing tasks. The model's bidirectional loop between enhanced understanding, controllable spatial editing, and novel-view-assisted reasoning enables it to move beyond general visual competence toward stronger spatial intelligence. The results suggest a promising path for unified visual models in downstream applications such as vision-language-action systems and world models. Overall, the paper contributes to the development of a unified multimodal model that can effectively understand and generate visual content with enhanced spatial intelligence.

💡 The paper introduces OpenSearch-VL, an open-source framework for training advanced multimodal search agents using reinforcement learning. The problem addressed is that current top-tier multimodal search agents are difficult to reproduce due to the lack of open high-quality training data, transparent trajectory synthesis pipelines, and detailed training recipes. To solve this, the authors propose a fully open-source recipe for training frontier multimodal deep search agents.

The method involves curating high-quality training data through a dedicated pipeline that includes Wikipedia path sampling, fuzzy entity rewriting, and source-anchor visual grounding. This pipeline is used to create two training datasets, SearchVL-SFT-36k and SearchVL-RL-8k. The authors also design a diverse tool environment that combines text search, image search, and other tools to enable agents to acquire external knowledge.

A new training algorithm, multi-turn fatal-aware GRPO, is proposed to handle cascading tool failures by masking post-failure tokens while preserving useful pre-failure reasoning. The results show that OpenSearch-VL delivers substantial performance gains, with over 10-point average improvements across seven benchmarks, and achieves results comparable to proprietary commercial models on several tasks. The authors will release all data, code, and models to support open research on multimodal deep search agents.

Overall, the paper contributes to the development of multimodal search agents by providing an open-source framework, high-quality training data, and a novel training algorithm, making it easier for researchers to reproduce and improve upon the results.

💡 The paper introduces Continuous-Time Distribution Matching, a new method for accelerating diffusion models through distillation. The existing Distribution Matching Distillation method has limitations, as it relies on sparse supervision at discrete timesteps, leading to visual artifacts and over-smoothed outputs. To address this, the authors propose a continuous-time optimization approach that replaces the fixed discrete schedule with a dynamic continuous schedule, allowing distribution matching to be enforced at arbitrary points along sampling trajectories. Additionally, they introduce a continuous-time alignment objective that performs active off-trajectory matching on latents extrapolated via the student's velocity field, improving generalization and preserving fine visual details. The results show that this new method, called Continuous-Time Distribution Matching, provides highly competitive visual fidelity for few-step image generation without relying on complex auxiliary objectives, outperforming existing methods on various architectures. The code for the method is made available, demonstrating the effectiveness of this new approach for diffusion model distillation.

💡 The paper introduces MARBLE, a novel gradient-space optimization framework for multi-reward reinforcement learning fine-tuning of diffusion models. The problem addressed is that existing methods for handling multiple rewards either train separate models for each reward or use a weighted-sum reward aggregation, which can lead to poor performance due to sample-level mismatch. This mismatch occurs because most rollouts are highly informative for certain reward dimensions but irrelevant for others, causing the weighted summation to dilute their supervision.

To address this issue, MARBLE maintains independent advantage estimators for each reward and computes per-reward policy gradients. These gradients are then harmonized into a single update direction without manual reward weighting, by solving a quadratic programming problem. This approach allows for a unified model that can be jointly trained on all rewards, eliminating the need for heavy manual tuning and sequential training.

The authors also propose an amortized formulation that reduces the computational cost of MARBLE, making it more efficient. Additionally, they use exponential moving average smoothing on the balancing coefficients to stabilize updates against transient fluctuations.

The results show that MARBLE improves all five reward dimensions simultaneously on the SD3.5 Medium dataset, outperforming the baseline method. Specifically, MARBLE turns the worst-aligned reward's gradient cosine from negative to consistently positive, indicating better alignment with human preferences. Furthermore, MARBLE runs at nearly the same training speed as the baseline method, with only a 3% slowdown. Overall, MARBLE provides a more effective and efficient approach to multi-reward reinforcement learning fine-tuning of diffusion models.

💡 The paper proposes SwiftI2V, an efficient framework for high resolution image to video generation. The problem addressed is that existing methods for generating high resolution videos from images are either computationally expensive or lack fidelity to the input image. High resolution image to video generation aims to synthesize realistic temporal dynamics while preserving fine grained appearance details of the input image, but at 2K resolution, this becomes extremely challenging. Existing solutions suffer from weaknesses such as high memory and latency costs, or hallucinating details and drifting from input specific local structures.

The proposed method, SwiftI2V, addresses these weaknesses by using a two stage design. First, it generates a low resolution motion reference to reduce token costs and ease the modeling burden. Then, it performs a strongly image conditioned 2K synthesis guided by the motion to recover input faithful details with controlled overhead. To make generation more scalable, SwiftI2V introduces Conditional Segment wise Generation, which synthesizes videos segment by segment with a bounded per step token budget. It also adopts bidirectional contextual interaction within each segment to improve cross segment coherence and input fidelity.

The results show that SwiftI2V achieves performance comparable to end to end baselines while reducing total GPU time by 202 times on the VBench I2V benchmark at 2K resolution. This enables practical 2K image to video generation on a single datacenter GPU or consumer GPU, making it a significant contribution to the field of image to video generation. Overall, SwiftI2V provides an efficient and scalable solution for high resolution image to video generation, addressing the efficiency fidelity dilemma and achieving state of the art results.

💡 The paper introduces DeepCode, a fully autonomous framework that addresses the challenge of converting documents into codebases, such as turning scientific papers into code. The existing methods for doing this have significant limitations due to the large amount of information in documents and the limited context that large language models can handle. DeepCode solves this problem by optimizing the flow of information through four key operations: source compression, structured indexing, knowledge injection, and error correction.

The source compression operation uses blueprint distillation to reduce the amount of information in the document. The structured indexing operation uses stateful code memory to organize the information in a way that makes it easier to access and use. The knowledge injection operation uses retrieval-augmented generation to add relevant knowledge to the code. The error correction operation uses closed-loop error correction to ensure that the code is accurate and reliable.

The paper evaluates DeepCode on a benchmark called PaperBench and finds that it achieves state-of-the-art performance, outperforming leading commercial agents and even surpassing PhD-level human experts. This means that DeepCode can take a scientific paper and turn it into code that is comparable in quality to code written by a human expert. The results of this paper have significant implications for the field of autonomous scientific reproduction, as they demonstrate the potential for AI systems to accelerate research evaluation and discovery by automating the process of converting scientific papers into code. Overall, the paper presents a major breakthrough in the field of document-to-codebase synthesis and has the potential to revolutionize the way that scientific research is conducted.

💡 This paper provides a comprehensive survey of the design, evaluation, and improvement of intelligent agents based on modular, brain-inspired architectures. The authors discuss the challenges and opportunities in developing advanced intelligent agents capable of sophisticated reasoning, robust perception, and versatile action across diverse domains. The survey is structured into four parts, starting with the modular foundation of intelligent agents, where the authors map cognitive, perceptual, and operational modules onto analogous human brain functionalities, including memory, world modeling, reward processing, and emotion-like systems.

The second part of the survey focuses on self-enhancement and adaptive evolution mechanisms, exploring how agents can autonomously refine their capabilities, adapt to dynamic environments, and achieve continual learning through automated optimization paradigms, including emerging AutoML and LLM-driven optimization strategies. The authors also examine collaborative and evolutionary multi-agent systems, investigating the collective intelligence emerging from agent interactions, cooperation, and societal structures, and highlighting parallels to human social dynamics.

Finally, the paper addresses the critical imperative of building safe, secure, and beneficial AI systems, emphasizing intrinsic and extrinsic security threats, ethical alignment, robustness, and practical mitigation strategies necessary for trustworthy real-world deployment. The authors provide a comprehensive overview of the current state of intelligent agents, framing their design, evaluation, and improvement within a modular, brain-inspired architecture that integrates principles from cognitive science, neuroscience, and computational research.

The paper's contributions include a systematic mapping of intelligent agent modules onto human brain functionalities, an exploration of self-enhancement and adaptive evolution mechanisms, an examination of collaborative and evolutionary multi-agent systems, and an emphasis on building safe, secure, and beneficial AI systems. The authors provide a comprehensive framework for understanding the design, evaluation, and improvement of intelligent agents, and highlight the importance of addressing the challenges and opportunities in developing advanced intelligent agents capable of sophisticated reasoning, robust perception, and versatile action across diverse domains.

💡 The paper discusses the limitations of traditional retrieval methods in complex tasks, particularly in agentic search where agents need to query raw text directly. Conventional retrieval systems use a fixed similarity interface that compresses access into a single step, making it difficult to implement tasks that require multiple steps, such as discovering intermediate entities, combining weak clues, and revising plans. To address this limitation, the authors propose direct corpus interaction, where an agent searches the raw corpus directly using general-purpose tools without any embedding model or retrieval API. This approach requires no offline indexing and adapts naturally to evolving local corpora. The authors evaluate this approach on several benchmarks and tasks, including IR benchmarks and end-to-end agentic search tasks, and find that it substantially outperforms strong baselines, including sparse, dense, and reranking methods. The results indicate that direct corpus interaction is a simple yet effective approach that opens a broader interface-design space for agentic search, and that retrieval quality depends not only on reasoning ability but also on the resolution of the interface through which the model interacts with the corpus. Overall, the paper contributes to the development of more effective agentic search systems by introducing a new approach to retrieval that allows for more flexible and direct interaction with the corpus.

💡 The paper presents UI-TARS-2, a native GUI-centered agent model that addresses challenges in data scalability, multi-turn reinforcement learning, and environment stability. The development of autonomous agents for graphical user interfaces is a major challenge in artificial intelligence, with open problems in data scalability, multi-turn reinforcement learning, and environment stability. To address these challenges, the authors propose a systematic training methodology that includes a data flywheel for scalable data generation, a stabilized multi-turn reinforcement learning framework, a hybrid GUI environment that integrates file systems and terminals, and a unified sandbox platform for large-scale rollouts.

The authors evaluate UI-TARS-2 on various benchmarks, including GUI benchmarks such as Online-Mind2Web, OSWorld, WindowsAgentArena, and AndroidWorld, as well as game environments and software engineering benchmarks. The results show that UI-TARS-2 achieves significant improvements over its predecessor UI-TARS-1.5 and strong baselines such as Claude and OpenAI agents. Specifically, UI-TARS-2 reaches high scores on GUI benchmarks, attains a mean normalized score of 59.8 across a 15-game suite, and remains competitive with frontier proprietary models on LMGame-Bench.

The model also generalizes to long-horizon information-seeking tasks and software engineering benchmarks, highlighting its robustness across diverse agent tasks. The authors provide detailed analyses of training dynamics, which provide insights into achieving stability and efficiency in large-scale agent reinforcement learning. Overall, the paper demonstrates UI-TARS-2's potential to advance the state of GUI agents and exhibit strong generalization to real-world interactive scenarios. The contributions of the paper include the development of a systematic training methodology, the evaluation of UI-TARS-2 on various benchmarks, and the analysis of training dynamics, which provide insights into achieving stability and efficiency in large-scale agent reinforcement learning.