An 8B model reportedly outscored Claude Opus 4.8 and GPT-6 Astra on creative utility

A post says Salesforce and colleagues trained an 8B reasoning model to generate several research ideas from source papers in one pass. Salesforce Research says the team rewarded the whole set, scoring each new idea higher the farther it moved from earlier ones. The post says the model outscored Claude Opus 4.8 and GPT-6 Astra on creative utility, a measure combining quality and diversity, on the team’s contamination-free benchmark.