AI Digest — Aug 13, 2026 (Morning)

This paper explores strong-to-weak capability transfer via harnesses at test time, where a stronger builder model helps a weaker target model solve tasks more reliably without parameter updates. The study uses four Theory-of-Mind benchmarks and achieves significant performance gains, nearly doubling the target model's average performance. The gains come from offloading unstable model reasoning into deterministic code and strict answer-format enforcement. The results suggest that inference-time harness design complements conventional training-time distillation, enabling strong models to transfer cognitive structure to weaker models without retraining. The approach allows weaker target models to receive the largest gains.

A study on multi-agent reinforcement learning identifies the issue of simulator collapse, where a single large language model simulator fails to generalize due to mode collapse. The researchers propose two solutions: Verbalized Sampling, which broadens the simulator's behavior at inference time, and Co-Training, which jointly optimizes the policy against a population of trainable simulators during training. Both methods improve held-out success and preserve policy diversity. The study validates these solutions on three benchmarks and releases an open-source framework called SCOPE for further research.

OpenAI has released research on how enterprises are adopting agentic AI, leveraging tools like ChatGPT and Codex. The study highlights the ways in which frontier firms are advancing in AI adoption, showcasing the integration of AI into core business operations. This adoption is significant as it demonstrates the practical application of AI technologies in enhancing business processes and decision-making. The research provides insights into the strategies and outcomes of early adopters, offering lessons for other enterprises looking to implement AI solutions.

Google DeepMind has introduced a sign-language-to-text (SL2T) model, a breakthrough in AI technology that can interpret sign language and translate it into text. This innovation is aimed at assisting Deaf and hard of hearing users. The model powers new sign language features, enhancing accessibility and communication for these users. The development of SL2T demonstrates the potential of AI in improving inclusivity and facilitating interaction across different language barriers.

Google researchers have identified recall as a significant bottleneck for achieving parametric factuality in generative AI models. Parametric factuality refers to the ability of models to accurately recall and generate factual information. The study suggests that improving recall is crucial for enhancing the overall performance and reliability of generative AI systems. This research has implications for the development of more accurate and informative generative models. The findings are based on experiments and analyses conducted by the Google research team.

Microsoft Research has introduced MindTopo, a new benchmark for evaluating the spatial reasoning and planning capabilities of Visual-Language Models (VLMs). MindTopo assesses how well AI models understand topological relationships, such as those between paths, fences, and knots. This benchmark highlights areas where VLMs can be improved to enhance their spatial reasoning abilities. The development of MindTopo could lead to advancements in AI's ability to understand and interact with complex environments. By identifying the strengths and weaknesses of current VLMs, researchers can develop more effective models for spatial reasoning and planning.

Hugging Face has introduced OlmoEarth embeddings, which are custom embedding exports from OlmoEarth Studio. These embeddings can be used for downstream analysis. OlmoEarth Studio is a platform that allows users to create and customize their own embeddings. The introduction of OlmoEarth embeddings provides users with more flexibility and control over their embedding exports. This development is relevant to those working with embeddings in natural language processing and related fields.

Hugging Face has introduced LFM2.5-VL-3B, a model designed to enhance vision capabilities at the edge. This model aims to provide better and faster performance for vision tasks. The release is significant for edge AI applications, where efficient and accurate vision processing is crucial. LFM2.5-VL-3B is part of the efforts to make AI more accessible and efficient for various use cases.

Anthropic is reportedly in discussions to acquire AI startup Decart for $6 billion. Decart is known for its work on world models, which are AI systems designed to understand and simulate complex real-world environments. This potential acquisition could significantly expand Anthropic's capabilities in AI research and development. The deal, if completed, would be one of the largest in the AI sector. Anthropic's interest in Decart underscores the growing importance of world models in advancing AI technology.

DeepSeek V4 Pro 0813 has been announced on OpenRouter.ai. The update is part of the DeepSeek series, which focuses on advancements in AI routing technologies. This version may include improvements to existing features or the introduction of new capabilities. The release is discussed on a popular forum with 831 points and 327 comments, indicating significant interest within the community. The technical details of the update are available on the OpenRouter.ai website.