The paper introduces Structured Thoughts, a reasoning format that alternates between <try> blocks containing exploratory work and <outcome> blocks containing distilled conclusions. The authors build a reformatted training dataset by segmenting existing reasoning traces and prompting an LLM to summarize each step. Fine-tuning pretrained foundation models reportedly improves reasoning benchmark performance by up to 8.08% over standard SFT. A proof-of-concept pruning method removes prior <try> blocks after each conclusion, achieving an average 85% memory or context reduction on mathematical tasks, with an 8.67% performance drop.
No heat snapshots are available in the last 24 hours.