reasoning models news, stories and analysis

Bar chart of LingoQA reasoning benchmark scores comparing six models. Alpamayo 2 Super leads at 79.2 points, followed by Alpamayo 1.5 Nano at 74.2, Qwen3-VL 32B at 72.2, Gemini 2.5 Pro at 64.1, Qwen2.5-VL 72B at 62.2, and GPT-4o at 56.0.

NVIDIA's Alpamayo 2 Super goes open for commercial AV use

NVIDIA Alpamayo 2 Super is a 34B-parameter open reasoning model for autonomous vehicles, released under a permissive commercial license on Hugging Face.

Lars Cornelissen · Aug 5, 2026

Cooperation rate with deceptive power-seeking instructions, showing base models at 19 percent, instruction-tuned models at 41 percent, and reasoning models at 58 percent. Reasoning models seek power nearly three times as often as base models.

Research

SysAdmin benchmark: reasoning models seek power 58% of the time

Instrumental power-seeking in AI is when models grab resources or expand access to complete goals. The SysAdmin benchmark shows reasoning models cooperate with deceptive power-seeking instructions 58 percent of the time, three times the base model rate.

Lars Cornelissen · Jul 23, 2026

GRPO standard deviation chart showing baseline extreme gradient mass at 13.9 percent, GRPO at 24.7 percent, and silent prompts at 44 percent for G=8.

Research

GRPO standard deviation is the reasoning RL dial now

GRPO standard deviation is the update-size dial: Bay and Yearick show 44% of Big-Math prompts go silent at group size 8.

Lars Cornelissen · Jul 2, 2026