CtrlCoT: Dual-Granularity Chain-of-Thought Compression for Controllable Reasoning

Dimensions overall score 8.0

• CtrlCoT is a dual-granularity chain-of-thought compression framework that harmonizes semantic abstraction and token-level pruning.method 70%
  Evidencepartial

  We propose \textbf{CtrlCoT}, a dual-granularity CoT compression framework that harmonizes semantic abstraction and token-level pruning through three components

  Implicationpartial

  The abstract explicitly introduces CtrlCoT and describes its core mechanism.

  Verificationpartial

  partial

• CtrlCoT achieves 30.7% fewer tokens on MATH-500 with Qwen2.5-7B-Instruct.result 70%
  Evidencepartial

  On MATH-500 with Qwen2.5-7B-Instruct, CtrlCoT uses 30.7% fewer tokens

  Implicationpartial

  This is a specific quantitative result presented in the abstract.

  Verificationpartial

  partial

• CtrlCoT achieves 7.6 percentage points higher accuracy than the strongest baseline on MATH-500 with Qwen2.5-7B-Instruct.result 70%
  Evidencepartial

  while achieving 7.6 percentage points higher than the strongest baseline

  Implicationpartial

  This is a specific quantitative result presented in the abstract, comparing against a baseline.

  Verificationpartial

  partial

• Existing CoT compression methods struggle to combine semantic shortening and aggressive token pruning effectively.limitation 70%
  Evidencepartial

  Moreover, combining the two is non-trivial due to sequential dependency, task-agnostic pruning, and distribution mismatch.

  Implicationpartial

  The abstract highlights the difficulties of existing methods in combining different compression strategies.

  Verificationpartial

  partial

• CtrlCoT's Hierarchical Reasoning Abstraction component produces CoTs at multiple semantic granularities.method 70%
  Evidencepartial

  Hierarchical Reasoning Abstraction produces CoTs at multiple semantic granularities

  Implicationpartial

  The abstract details the components of CtrlCoT, and this claim describes the function of one component.

  Verificationpartial

  partial

• CtrlCoT's Logic-Preserving Distillation component trains a logic-aware pruner to retain indispensable reasoning cues.method 70%
  Evidencepartial

  Logic-Preserving Distillation trains a logic-aware pruner to retain indispensable reasoning cues (e.g., numbers and operators) across pruning ratios

  Implicationpartial

  The abstract details the components of CtrlCoT, and this claim describes the function of another component.

  Verificationpartial

  partial

• CtrlCoT's Distribution-Alignment Generation component aligns compressed traces with fluent inference-time reasoning styles.method 70%
  Evidencepartial

  Distribution-Alignment Generation aligns compressed traces with fluent inference-time reasoning styles to avoid fragmentation.

  Implicationpartial

  The abstract details the components of CtrlCoT, and this claim describes the function of the third component.

  Verificationpartial

  partial

• The approach may not generalize to all LLM applications and risks losing critical information in reasoning for tasks outside tested domains.limitation 70%
  Evidencepartial

  The approach may not generalize to all LLM applications, and there is potential risk of losing critical information in reasoning for tasks outside the tested domains.

  Implicationpartial

  This is explicitly stated as a caveat in the provided analysis.

  Verificationpartial

  partial

Competitive landscape

Develop an AI tool for compressing reasoning chains in large language models to reduce latency and costs without sacrificing accuracy.