Extra-CoT tackles one of the most pressing practical problems in deploying large reasoning models: Chain-of-Thought traces are computationally expensive at inference time. While models like DeepSeek-R1 and o1 have demonstrated that longer thinking leads to better answers, the cost scales linearly with token count — making real-world deployment expensive. This paper proposes a two-stage training pipeline: first a dedicated "semantically-preserved compressor" is trained on annotated mathematical reasoning data, then the base LLM is fine-tuned on compressed CoT pairs using a mixed-ratio SFT regime followed by a novel RL algorithm called CHRPO (Constrained and Hierarchical Ratio Policy Optimization).
The key innovation here is achieving 73%+ token reduction on MATH-500 with Qwen3-1.7B while actually improving accuracy by 0.6% — a result that sounds paradoxical but reflects a real phenomenon: much of the token budget in standard CoT traces is redundant self-repetition, false starts, and verbose restatements that don't contribute to correct answers. By training the model to internalize a compressed reasoning style, Extra-CoT essentially teaches the model to "think tighter." This is a fundamentally different approach from post-hoc pruning or summarization, since the compression is baked into the generation process itself via RL reward shaping.
The inference efficiency angle here has direct commercial implications. If reasoning models can maintain answer quality at 73% fewer tokens, that translates to roughly 3x throughput improvement and proportional cost reduction — critical for any API provider serving reasoning-heavy workloads at scale. This feeds directly into the cost-collapse narrative currently reshaping the AI industry. Furthermore, CHRPO's hierarchical reward structure — which explicitly penalizes verbosity under tighter budgets — is a clean formulation that could generalize beyond mathematical reasoning to code generation, multi-step planning, and scientific Q&A.
The benchmarks are limited to mathematical reasoning tasks (MATH-500 and two others), which may not generalize to domains where chain-of-thought diversity and exploration are more critical — e.g. open-ended scientific reasoning or adversarial tasks. The compressor model itself adds a training complexity that smaller labs may struggle to replicate. And the paper doesn't address whether compressed CoT affects interpretability or auditability — shorter traces may be harder for humans to verify. Still, as a step toward making reasoning models practical, Extra-CoT is a meaningful contribution to the efficiency frontier.