Salesforce Trains Reasoning Models to Batch-Generate Diverse Research Ideas
Salesforce · hf · 2026-10-20
Salesforce researchers propose a set-level reinforcement learning objective that trains LLMs to generate K diverse, high-quality, document-grounded scientific ideas in a single pass.
Unlike prior work that optimizes one idea toward a gold target, the method explicitly reasons over candidate ideas, avoids directions already taken, and targets open problems — making diversity built-in. The authors report stronger creative utility than larger proprietary models.
More from Research
- Mirror video planner shows emergent in-context learning, boosting robot lava-lake escape rate from 46% to 76% — vincesitzmann · 2026-10-20
- Training a small downscaler in 1 hour on one GPU cuts weather forecast error by up to 21% — yb2698 · 2026-10-20
- Progress-Aware Distillation Targets Stagnation in Small Language Model Agents — Yuanpu Cao · 2026-10-20
- HERA: co-evolving agent harnesses and environments for reliable abstention — UW · 2026-10-20
- HouseholdBench tests 13 LLMs on predicting household economic behavior; GBDT still wins — daveholtz · 2026-10-20
- Mirror author predicts an 'arm-farm' revival as video models make robot learning sample-efficient — vincesitzmann · 2026-10-20