Salesforce Trains 8B Reasoning Model for Diverse Research Ideas
A small open-scale model beating frontier systems on creative diversity shows that reward design, not model size, can be the bottleneck in AI-driven research ideation. Teams building automated research or ideation pipelines can study the set-level reward scheme as a cheaper alternative to relying on frontier APIs. The paper is on arXiv (2610.22324) with a companion Chat with Paper resource from DAIR Academy.
Salesforce and collaborators trained an 8B reasoning model to generate multiple research ideas in a single pass from a set of source papers. The reward function scores the entire idea set, weighting each idea's quality by its distance from earlier ideas so the model learns to diversify. This targets a known weakness of AI-scientist pipelines, which struggle to propose novel and diverse ideas. On the team's contamination-free benchmark, the 8B model outperforms Claude Opus 4.8 and GPT-6 Astra on creative utility, a metric combining quality and diversity. The work is described in arXiv paper 2610.22324.
Great application of small language models from Salesforce and colleagues. AI-scientist pipelines are well known to struggle to propose novel and diverse research ideas. This team trains an 8B reasoning model to write several research ideas in one generation from a set of source papers. The reward scores the whole set, weighting each idea's quality by how far it sits from the ideas written before it. The model learns to move each new idea away from the earlier ones. On their contamination-free benchmark, the 8B model scores higher on creative utility, a measure that combines quality and diversity, than Claude Opus 4.8 and GPT-6 Astra. arxiv.org/abs/2610.22324 Chat with Paper: academy.dair.ai/papers/trainin… 💬 3 🔄 0 ❤️ 5 👀 696 📊 3 ⚡
