compact-llms

ClueWeaver is a reward-guided dual-agent framework that improves long-narrative question answering with compact local language models by decomposing evidence selection and reasoning into separate agents optimized via reinforcement learning.