A coalition of researchers from the UK AI Security Institute and the startup Timaeus have formed Sequent, a new nonprofit organization focused on developing alignment techniques for superintelligent AI systems. Sequent aims to move beyond reactive safety methods, seeking principled, generalizable insights that ensure safety even in large-scale, long-horizon tasks. The organization plans to build a portfolio of research bets—ranging from scalable oversight to game theory—with an initial funding goal of $100–150M and a target of 40–80 full-time staff within a few years.

Researchers have released ChinaHeritaQA, a multimodal benchmark dataset designed to test vision-language models on their understanding of 51 UNESCO World Heritage sites in China. The dataset includes 2,279 images and 14,133 multiple-choice question-answer pairs. Tested across seven categories, such as historical periodization and architectural analysis, the Qwen-VL-8B-Instruct model achieved an accuracy of 81%, outperforming the average human score of 67%.

Cognition, the developers behind the Devin coding agent, have introduced FrontierCode, a rigorous benchmark for evaluating production-level coding models. The benchmark consists of 150 tasks hand-selected by open-source developers from multi-pull-request chains. FrontierCode highlights include strict quality control and a focus on end-to-end code mergeability. In the hardest "Diamond" difficulty tier, Claude Opus 4.8 scored 13.4%, while GPT-5.5 reached 6.3%.

Xiaomi has debuted the MiMo-V2.5-Pro-UltraSpeed, a 1 trillion parameter LLM capable of generating 1000 tokens per second. The high speed is achieved through codesigned software, including FP4 quantization (reducing model weight precision to 4 bits for efficiency) and a speculative decoding technique called DFlash. The model operates on an 8-GPU commodity node, showcasing efforts to optimize performance on non-specialized hardware.

Researchers from Xi’an Jiaotong University and Xidian University have launched AARRI-Bench to assess how AI systems handle entry-level research tasks. The benchmark features 82 manually crafted tasks ranging from verifying scientific data to recognizing unproductive research trajectories. Among tested systems, Claude-Opus-4.7 reached 68.3% performance, while DeepSeek-v4-Flash scored roughly 60%. The framework emphasizes professional autonomy, technical proficiency, and ethical decision-making in academic settings.

A coalition of researchers from the UK AI Security Institute and the startup Timaeus have formed Sequent, a new nonprofit organization focused on developing alignment techniques for superintelligent AI systems. Sequent aims to move beyond reactive safety methods, seeking principled, generalizable insights that ensure safety even in large-scale, long-horizon tasks. The organization plans to build a portfolio of research bets—ranging from scalable oversight to game theory—with an initial funding goal of $100–150M and a target of 40–80 full-time staff within a few years.

Researchers have released ChinaHeritaQA, a multimodal benchmark dataset designed to test vision-language models on their understanding of 51 UNESCO World Heritage sites in China. The dataset includes 2,279 images and 14,133 multiple-choice question-answer pairs. Tested across seven categories, such as historical periodization and architectural analysis, the Qwen-VL-8B-Instruct model achieved an accuracy of 81%, outperforming the average human score of 67%.

Cognition, the developers behind the Devin coding agent, have introduced FrontierCode, a rigorous benchmark for evaluating production-level coding models. The benchmark consists of 150 tasks hand-selected by open-source developers from multi-pull-request chains. FrontierCode highlights include strict quality control and a focus on end-to-end code mergeability. In the hardest "Diamond" difficulty tier, Claude Opus 4.8 scored 13.4%, while GPT-5.5 reached 6.3%.

Xiaomi has debuted the MiMo-V2.5-Pro-UltraSpeed, a 1 trillion parameter LLM capable of generating 1000 tokens per second. The high speed is achieved through codesigned software, including FP4 quantization (reducing model weight precision to 4 bits for efficiency) and a speculative decoding technique called DFlash. The model operates on an 8-GPU commodity node, showcasing efforts to optimize performance on non-specialized hardware.

Researchers from Xi’an Jiaotong University and Xidian University have launched AARRI-Bench to assess how AI systems handle entry-level research tasks. The benchmark features 82 manually crafted tasks ranging from verifying scientific data to recognizing unproductive research trajectories. Among tested systems, Claude-Opus-4.7 reached 68.3% performance, while DeepSeek-v4-Flash scored roughly 60%. The framework emphasizes professional autonomy, technical proficiency, and ethical decision-making in academic settings.