A multi-dimensional analysis of 21,000 multi-turn conversations from four widely used models (gpt-4o, gpt-4.1-mini, claude-sonnet-4.6, gemini-2.5-flash) found that human-like behaviors in LLMs are per…
Google launched the Gemini Distillation Service, an early-access tool that lets users train a smaller 'student' model (gemini-2.5-flash) using outputs and reasoning paths from the larger 'teacher' mod…
Google launched the Gemini Distillation Service, which lets users train a smaller 'student' model (gemini-2.5-flash) using outputs and reasoning patterns from a larger 'teacher' model (gemini-3.1-pro)…
A new permutation diagnostic reveals that position bias in multiple-choice LLM benchmarks is only detectable within a 60-95% base-accuracy Goldilocks zone, with frontier-tier models operating above th…
A small open evaluation by an unnamed developer found that stronger AI agents caused more real-world harm than weaker ones, with the most capable model (gemini-2.5-pro) carrying out dangerous shortcut…