AI Digest — Jul 31, 2026 (Morning)

OpenAI has announced the release of GPT-5.6, a new version of their language model. This update aims to improve the price-performance ratio of their models. The release is significant for AI researchers and developers as it may provide better efficiency and cost-effectiveness for various natural language processing tasks. GPT-5.6 is part of OpenAI's efforts to continuously enhance their language models. The announcement has garnered significant attention on platforms like Hacker News.

Researchers introduced AskChem, a claim-centered infrastructure for cross-paper chemistry search. It converts papers into atomic, typed claims with source DOIs and verbatim quotes. AskChem provides a web interface and API access for AI agents, currently indexing 2.4M claims from 147K papers. The system yields 100% resolvable DOIs when grounding a GPT-5.5 reader. AskChem aims to facilitate chemistry literature synthesis by assembling specific findings from multiple publications.

Researchers introduced Change2Task, a system that generates executable coding agent tasks from repository changes. It uses historical evidence and reconstructs task states to create verified tasks. The system was evaluated on five common coding agent task families and achieved 79.6% verified task construction success. Change2Task reduces environment setup and task construction effort, and its reconstructed cases show high agreement with historical outcomes. The system also reduces expenditure across the pipeline by 10.8%.

Researchers compared seven language model methods, including self-refine and reflexion, against repeated sampling on mathematics benchmarks with 1.5B, 3B, and 7B parameter models. The study found that no method was reliably better than repeated sampling at equal cost, with ten methods being reliably worse, particularly those involving self-inspection. The performance gap between self-inspection methods and repeated sampling decreased as model size increased. The study's findings suggest that the benefits of self-refine and reflexion methods may be due to increased sampling rather than the methods themselves. The researchers released their code, prompts, and verification scripts to facilitate further study.

A recent study on stage-replay diagnostics in Qwen2.5-derived systems examines the assumption of treating fresh-prefill continuation as continuation from the decoder state that originally reached the prefix. The experiment compares retained live cache with one-shot prefill of identical integer tokens and finds that fixed-prefix precision controls and bidirectional cache transplantation can make token-by-token incremental and retained live caches bit-exact. The study also shows that boundary K/V cache is a causally sufficient carrier of the divergent trajectory, while numerical precision moderates its behavioral expression. The findings have implications for understanding the role of cache in stage-replay divergence. The study used a matched 200-item experiment and found that the accuracy difference between BF16 and FP32 was only one point.

Google DeepMind has introduced Gemini Robotics ER 2, a system designed to enhance robotics capabilities through advanced video understanding, task orchestration, and multi-robot collaboration. This technology aims to improve robots' ability to reason, collaborate, and solve real-world tasks. Gemini Robotics ER 2 represents a significant advancement in robotics, particularly in areas requiring complex task execution and coordination. The system's capabilities are expected to have implications for various robotic applications, including those in industrial and service settings.

Google Research has introduced the Science One Framework, a verifiable autonomous research framework that utilizes a Chain-of-Evidence approach. This framework is designed to enhance the transparency and reproducibility of scientific research. By leveraging this framework, researchers can create a clear, auditable record of their experiments and findings. The Science One Framework has the potential to improve the reliability of research outcomes. It is particularly relevant in fields where data integrity and experiment reproducibility are crucial.

Microsoft Research has introduced Echoverse, a platform designed to train computer-use AI agents in realistic, evolving environments. This approach aims to improve the agents' ability to handle multi-step workflows such as email and customer support. Unlike traditional methods that focus on providing more training tasks, Echoverse allows agents to learn and adapt as the tasks, tests, and environments evolve. This can potentially lead to more effective and efficient AI agents in real-world applications. The Echoverse platform is part of Microsoft's efforts to advance AI research and development.

Microsoft Research has introduced EvoLib, a system designed to turn experience into evolving knowledge for large language models (LLMs). EvoLib aims to enable LLMs to learn and adapt across tasks long after deployment by leveraging reusable skills and insights. This approach differs from simply increasing the memory of LLMs, focusing instead on enhancing their ability to apply learned knowledge in new contexts. The goal is to improve the long-term adaptability and performance of LLMs in various applications.

Researchers presented a paper at the International Conference on Machine Learning, arguing that large language models (LLMs) cannot be made fully secure due to a fundamental flaw in their design. This flaw leaves LLMs open to potential hacks, which has significant implications for the safety and security of this technology. The paper highlights the limitations of current LLM architectures and the need for new approaches to mitigate these vulnerabilities. The findings are based on a study presented at a top AI conference, indicating the seriousness of the issue.

A US judge has expressed doubt over the justification of the US ban on Anthropic AI. The ban was imposed due to concerns over the potential risks associated with advanced AI technologies. The judge's skepticism highlights the need for clearer guidelines and regulations regarding AI development and deployment. This case involves Anthropic, a company known for its work in AI safety and development of the Claude AI model. The outcome of this case could have implications for the development and regulation of AI in the US.

Anthropic AI models were tested on three companies and successfully hacked into their systems. The tests were conducted to evaluate the AI's capabilities and potential risks. The companies involved have not been named, but the results highlight the potential vulnerabilities of certain systems to advanced AI models. The tests demonstrate the importance of considering AI safety and security in the development and deployment of AI technologies.