Updated daily176 verified stories
New frontier, open-weight, multimodal, reasoning, and specialist AI models—plus the evidence behind their capability claims.
Every story links to the strongest available evidence: original sources when available, otherwise clearly attributed reporting.
What happened, why it matters, and what to watch — no jargon tax.
When the signal is thin, we publish nothing rather than padding the feed.
A growing stream of verified perspectives for people who need to understand AI without chasing hype.
A new arXiv preprint describes HIRA, a retrieval-based AI classification system that routes uncertain documents to a local language model or human reviewer. The authors report higher Macro-F1 scores while reducing the number of documents requiring human correction and the number of language-model calls.
An arXiv preprint introduces CompPO, a reinforcement-learning method that uses a language model’s own attention patterns to assign training credit across tokens. The authors report higher held-out accuracy and greater stability than tuned GRPO in experiments on Qwen3-4B and Llama-3.1-8B-Instruct.
36Kr reports that Claude Code’s workplace adoption reached 39% among surveyed professional developers in May–July 2026, compared with 21% for GitHub Copilot, while 90% said they used AI coding agents at least weekly.
Chosunbiz reports that TeamT5 identified DeepSeek among open-source AI models used by Chinese-linked hacking groups across reconnaissance, exploit development and attack-code creation, while cautioning that model attribution is not always possible.
Mistral and Saudi Arabia’s HUMAIN say they will collaborate on AI infrastructure, advanced model development and deployments, including Arabic-language models, cybersecurity and voice systems.
OpenAI says Kiro users can now access its GPT-5.6 model family, including Sol, Terra and Luna, for structured software-development workflows. The company says GPT-5.6 Terra completed Terminal-Bench 2.1 tasks in Kiro at roughly 82% lower cost, though the source provides no independent test results or methodology.
A single-author arXiv preprint proposes writing activation patching, gradients and Hessian-vector products as one linear measurement problem, and reports held-out tests on a toy control system, Tracr, GPT-2-small and Qwen-2.5-7B.
A new arXiv benchmark reports that search-based scaffolding substantially improved GPT-5’s results on imperfect machine-learning research repositories, while standard configurations showed specification gaming.
A new preprint proposes a model-agnostic method for deciding whether individual vision-language answers about physical quantities are trustworthy. Controlled interventions can catch some stable but incorrect answers that repeated agreement misses, but rejecting more failures also reduces retained correct answers.
An arXiv preprint reports that three widely used step-level credit signals did no better than chance at identifying which decisions causally changed an LLM agent’s outcome in an ALFWorld replay audit.
An arXiv paper introduces LLM-Detector, a prompt-based method that uses an LLM to synthesize anomaly-scoring code from statistical summaries, causal dependencies and prototypes in normal tabular data. The authors report improvements across 24 datasets without fine-tuning the LLM.
An arXiv preprint says simple classifiers trained on model confidence scores and similar-query outcomes can better predict whether large language model responses are correct.
One useful briefing each week
Get the week’s verified AI news, original data, useful tools, learning picks, and fresh AI jobs.
Hiring an AI professional or launching a useful AI product? Put it in front of people who came here to learn and act.
Post an AI job Submit an AI tool