8B reasoning model beats Opus 4.8 on diverse scientific idea generation
omarsar0 · x · 2026-10-18
Salesforce AI Research with NYU and Stony Brook published 'Training Reasoning Models to Generate Diverse Scientific Ideas,' tackling the well-known homogeneity problem in AI-scientist pipelines.
- Method: an 8B reasoning model is trained to write multiple research idea cards grounded in source papers in a single generation; a set-level RL reward scores the whole set, weighting each idea by its distance from earlier ones so the model learns to diverge.
- Results: on a contamination-free benchmark, the 8B model beats Claude Opus 4.8 and GPT-6 Astra on creative utility (a quality+diversity measure).
- Human eval: in a pilot with four experts, its idea sets were judged as promising as Opus 4.8's and marginally more unique; success@k on proof-of-concept runs matches Opus 4.8, so diversity doesn't cost executability.
- Code and data are open-sourced (github.com/ManyaAdhwa/divergeideation).
More from AGI Musings
- French math PhD students despair: a €20 ChatGPT could solve their thesis problems — FlorianGallwitz · 2026-10-18
- Ethan Mollick: agentic flooding of government services forces a worse-or-better choice — emollick · 2026-10-18
- Robin Hanson: China's AI+ Plan omits competition and the US, focusing on diffusion — Afinetheorem · 2026-10-18
- Gaming breaks dreams; AI will finally let anyone see their game idea come alive — illscience · 2026-10-18
- "I built my career on my era's best computer, so you can't use yours" — timigod · 2026-10-18
- Sparks Brain Preservation Founder on Whole Brain Emulation and Aldehyde Fixation — matiroy · 2026-10-18