Apple researchers tested how well large language models agree with human annotators when judging machine translation quality under two schemes, MQM and ESA. The work matters because LLMs are increasingly seen as a cheaper stand-in for human evaluators, yet they need checking across languages, domains and levels of detail first.
An Apple research paper describes Minions, a system that pools the mostly idle GPUs of Apple Silicon Macs into a distributed cloud for machine learning inference. The idea targets the cost, rigidity and waiting times of centralized GPU infrastructure.
Apple researchers propose SELFCOMPACT, a way for language model agents to decide for themselves when to compress their long traces of reasoning and tool calls. Current scaffolds compact at a fixed token threshold, which can throw away partial results in the middle of a task.
Vals AI found that teams of AI agents barely beat single agents while costing up to 5.1 times more, with only one of four tests using GPT-6 Sol and Claude Opus 5.5 showing a measurable gain. Anthropic's own data shows quality plateaus beyond ten agents while token costs keep rising.
A study finds that safety reminders make healthcare AI less likely to make dangerous choices. The trade-off is that the systems ask for human verification more often.
A study in Science reports that DeepMind's AlphaProof Nexus solved dozens of open problems, including nine Erdős problems, two of them open for 56 years. DeepMind had the work reviewed before publication, in contrast with OpenAI, which faces a boycott for posting unreviewed AI proofs on GitHub.