This paper introduces VLAW, an iterative algorithm for co-improving vision-language-action (VLA) policies and action-conditioned video generation world models using real-world rollouts. VLAW leverages real-world data to refine the world model, which is then used to generate synthetic data for further policy improvement, addressing the limitations of world models trained solely on demonstration datasets. Experiments on a real robot demonstrate a 39.2% absolute improvement in success rate over the base policy, highlighting the effectiveness of the iterative co-improvement strategy.

This paper explores test-time verification as a method to improve vision-language-action (VLA) alignment, addressing the "intention-action gap" in embodied instruction following. They demonstrate that scaling both rephrased instructions and generated actions at test time enhances sample diversity and improves action selection. The authors introduce CoVer, a contrastive verifier, and a hierarchical verification inference pipeline, showing that this verification approach outperforms scaling policy pre-training on the SIMPLER and PolaRiS benchmarks.

The paper introduces Fun-DDPS, a generative framework for carbon capture and storage (CCS) modeling that combines function-space diffusion models with differentiable neural operator surrogates for both forward and inverse problems. By decoupling the learning of a prior over geological parameters from the physics-consistent guidance provided by a Local Neural Operator (LNO) surrogate, Fun-DDPS effectively handles data sparsity and ensures physically realistic solutions. Experiments on synthetic CCS datasets demonstrate that Fun-DDPS significantly outperforms standard surrogates in forward modeling with sparse observations and achieves comparable accuracy to rejection sampling in inverse modeling, while also generating physically consistent realizations with improved sample efficiency.

The paper introduces Olmix, a framework designed to address challenges in data mixing for language model training, specifically focusing on understanding the configuration space of mixing methods and efficiently adapting to evolving domain sets. Through an empirical study, the authors identify key design choices for effective mixing methods and propose "mixture reuse," a technique that leverages past mixture ratios to efficiently recompute mixtures after domain set updates. Experiments show that mixture reuse achieves comparable performance to full recomputation with significantly reduced compute (74% less) and outperforms training without mixing by 11.6% on downstream tasks.

The paper introduces CROSS-ALIGN+, a three-stage framework for meme-based social abuse detection that addresses cultural blindness, boundary ambiguity, and lack of interpretability in existing methods. CROSS-ALIGN+ enriches multimodal representations with structured knowledge, reduces boundary ambiguity using LoRA adapters, and enhances interpretability through cascaded explanations. Experiments on five benchmarks and eight LVLMs show that CROSS-ALIGN+ outperforms state-of-the-art methods, achieving up to a 17% relative F1 improvement.

The paper introduces FOFPred, a language-conditioned optical flow forecasting model that combines a Vision-Language Model (VLM) with a Diffusion architecture. This model is trained on web-scale human activity data using specific preprocessing techniques to extract meaningful signals from noisy video-caption pairs. FOFPred demonstrates strong performance in both robotic manipulation and video generation tasks, highlighting the benefits of the unified architecture and scalable web data learning for predicting future motion.

The International AI Safety Report 2025's Second Key Update analyzes the current state of AI risk management and technical mitigations employed by researchers, companies, and governments. It highlights advancements in training safer models and monitoring outputs while acknowledging uncertainties in the effectiveness of these measures and their variability across applications. The report aims to inform policymakers, researchers, and the public about progress and remaining gaps in AI safety.

This paper investigates the key factors influencing the performance of batch online reinforcement learning in robotics, focusing on algorithm class, policy extraction methods, and policy expressivity. Through a systematic empirical study, the authors found that Q-function-based methods outperform imitation learning, implicit policy extraction is crucial, and expressive policy classes are preferred. Based on these findings, they propose a recipe for effective batch online RL, further enhanced by temporally-correlated noise, achieving superior performance compared to existing methods.

The paper introduces $\pi_{0.5}$, a vision-language-action (VLA) model designed for improved generalization in real-world robotic manipulation tasks. The model builds upon $\pi_{0}$ and employs co-training on heterogeneous data sources, including data from multiple robots, web data, and semantic predictions, to enhance its ability to generalize to unseen environments. Experiments demonstrate that $\pi_{0.5}$ can perform long-horizon, dexterous manipulation skills like cleaning a kitchen or bedroom in novel homes, showcasing the effectiveness of knowledge transfer for real-world robotic systems.

This paper critiques the rigid application of the Helpful, Honest, and Harmless (HHH) principle in AI alignment, arguing that its dimensions require adaptive prioritization based on context. The authors introduce the concept of "priority order" to manage trade-offs between HHH dimensions and propose a reference framework incorporating context definition, value prioritization, risk assessment, and benchmarking. Through case studies and analysis of interdependencies, the paper demonstrates how to jointly enhance harmlessness and helpfulness, providing a practical guide for ethically grounded and operationally effective AI deployment.

The study adapted and evaluated open-source (Llama 2-7B, Mistral-7B) and closed-source (GPT-4 Turbo) large language models to extract clinical history elements from imaging orders, aiming to automate the assessment of clinical history completeness, a known problem in radiology. Fine-tuned Mistral-7B achieved performance rivaling GPT-4 Turbo in extracting elements like "past medical history," "what," "when," "where," and "clinical concern," demonstrating substantial agreement with radiologists (mean κ 0.73-0.77). Applying Mistral-7B to a large dataset revealed that only 26.2% of clinical histories contained all five key elements, establishing a benchmark for completeness.