Tag
This paper measures the in-context binding capacity of language models, finding that recall scales with parameter count according to a power law and analyzing factors like pretraining recipes and interference.
The AI model has doubled its parameters, marking a substantial increase in its computational scale and potential capabilities.
Extended reasoning and post-training are key techniques for enhancing AI model performance, with speculation that future models like Qwen 4 could match or surpass large parameter models on specific tasks.