# Chain-of-Thought Compression

Large reasoning models often keep deliberating after enough information is available, raising decoding latency and memory cost. LoCo moves the compression decision to individual reasoning states: it finds candidate overthinking states in the model's own traces, freezes that prefix, regenerates shorter continuations whose answers verify as correct, and learns a continue-or-conclude preference with DPO. Across five math benchmarks, completion length drops 41.7%–45.4% while accuracy holds or improves; trained on math only, the same model cuts length by 23.4%–36.3% on three code benchmarks.

Large reasoning models often keep deliberating after enough information is available, raising decoding latency and memory cost. LoCo moves the compression decision to individual reasoning states: it finds candidate overthinking states in the model's own traces, freezes that prefix, regenerates shorter continuations whose answers verify as correct, and learns a continue-or-conclude preference with DPO. Across five math benchmarks, completion length drops 41.7%–45.4% while accuracy holds or improves; trained on math only, the same model cuts length by 23.4%–36.3% on three code benchmarks.

SynLat reframes chain-of-thought compression as a representation-fidelity allocation problem: constituency syntax partitions a trace into ordered, non-overlapping Syntax-Aligned Units (SAUs), an answer-conditioned Teacher builds progressive KEEP/LATENT targets for a single compression-conditioned Student, and latent units map to a shared discrete codebook, so at inference the Student generates mixed text-and-latent reasoning from only the question and requested compression level. Across two Qwen3 Student scales, Standard- and Long-CoT groups, and three compression levels, SynLat matches or exceeds the strongest evaluated baseline in all 12 task-group aggregates and strictly leads in 11, with overall gains of 3.6/2.6 points at MEDIUM and 7.0/5.5 points at HIGH.

SynLat reframes chain-of-thought compression as a representation-fidelity allocation problem: constituency syntax partitions a trace into ordered, non-overlapping Syntax-Aligned Units (SAUs), an answer-conditioned Teacher builds progressive KEEP/LATENT targets for a single compression-conditioned Student, and latent units map to a shared discrete codebook, so at inference the Student generates mixed text-and-latent reasoning from only the question and requested compression level. Across two Qwen3 Student scales, Standard- and Long-CoT groups, and three compression levels, SynLat matches or exceeds the strongest evaluated baseline in all 12 task-group aggregates and strictly leads in 11, with overall gains of 3.6/2.6 points at MEDIUM and 7.0/5.5 points at HIGH.