Tag
This study introduces section-aware compression for reasoning traces, training models to drop filler narration while preserving compute and verification spans, matching or exceeding original accuracy while using 2-3 times fewer tokens.