CLARO: Controlled Attribute-Driven Reasoning Optimization for Efficient Chain-of-Thought

CLARO reframes efficient Chain-of-Thought reasoning as a byproduct of structural quality rather than forced brevity, using reinforcement learning to internalize attributes like readability, math density, syntactic compression, and low redundancy within a user-defined token budget.

Problem Statement

Verbose, unstructured CoT outputs make LLM reasoning computationally expensive, and existing length-control methods (token penalties, truncation) risk stripping out necessary reasoning steps, degrading accuracy. This paper addresses the false trade-off between reasoning efficiency and reasoning quality by targeting the structure of thought itself rather than its length.

Key Novelty

Attribute-Guided Prompting (AGP): a lightweight, zero-shot prompting strategy that improves reasoning accuracy and reduces inference cost without any training
CLARO: an RL framework that internalizes structural reasoning attributes (readability, math density, syntactic compression, low redundancy) directly into model behavior under a token budget
Conceptual reframing of efficiency as an emergent property of well-structured reasoning rather than a directly optimized compression target
A concrete, measurable taxonomy of 'quality attributes' for Chain-of-Thought traces that can be used as reward signals

Evaluation Highlights

Outperforms state-of-the-art length-control/truncation baselines across diverse reasoning benchmarks
Accuracy gains of up to 63.6% while operating within constrained token budgets
Demonstrated efficiency-performance trade-off improvement over both zero-shot prompting (AGP) and RL-trained variants

Signal Assessment

6/10 The paper offers a well-motivated conceptual reframing (structure-driven emergent efficiency) and strong empirical gains, but builds on established RL-for-reasoning and structured-prompting techniques rather than introducing a fundamentally new paradigm.

Methodology

Identify and validate candidate structural attributes (readability, math density, syntactic compression, low redundancy) via zero-shot Attribute-Guided Prompting
Show these attributes correlate with improved reasoning accuracy and lower token usage across benchmarks
Design an RL reward scheme (CLARO) combining task correctness with attribute-quality signals under a user-defined token budget constraint
Fine-tune models via RL to internalize attribute-driven reasoning structure, removing reliance on explicit prompting
Evaluate against SOTA length-control baselines across diverse reasoning benchmarks

System Components

Attribute-Guided Prompting (AGP)

Zero-shot prompting technique that instructs the model to produce reasoning with specific structural qualities, serving as a cost-free baseline improvement

Structural Attribute Set

Defined qualities of reasoning text (readability, math density, syntactic compression, low redundancy) used as proxies for 'well-structured thought'

CLARO RL Framework

Reinforcement learning optimization pipeline that trains models to internalize the structural attributes as intrinsic generation behavior

Token Budget Controller

User-defined constraint mechanism ensuring reasoning stays within a target length while preserving necessary steps

Reward Signal

Composite reward likely combining answer correctness with attribute-quality scores to guide RL training

Results

Metric/Benchmark	Baseline	This Paper	Delta
Reasoning accuracy (peak improvement)	SOTA length-controlled baselines	CLARO	+63.6% (best case)
Zero-shot reasoning accuracy	Standard prompting	Attribute-Guided Prompting (AGP)	Improved accuracy, qualitative
Inference cost / token usage	Verbose unstructured CoT	CLARO within token budget	Reduced cost, qualitative

Key Takeaways

Optimizing the structural quality of reasoning traces (readability, compactness, low redundancy) can simultaneously improve accuracy and reduce inference cost, rather than trading one for the other
Simple zero-shot attribute-guided prompting is a low-effort first step practitioners can apply before investing in RL fine-tuning
Token budgets can be enforced without sacrificing critical reasoning steps if the model is trained to structure thoughts efficiently rather than truncate them
Treating CoT length control as a truncation/penalty problem may be suboptimal; treating it as a structure/quality problem yields better accuracy-efficiency trade-offs
The released code/models provide a practical starting point for integrating attribute-driven reasoning optimization into existing RLHF-style pipelines

Abstract

Large language models exhibit strong reasoning capabilities but often require significant computational resources due to verbose, un-structured Chain-of-Thought outputs. Recent approaches guide reasoning length through to-ken penalties or truncation, risking the omission of necessary steps. We posit that conciseness should be an emergent property of structured thought, rather than a result of artificially forced brevity. To this end, we first demonstrate that Attribute-Guided Prompting , a lightweight zero-shot strategy, improves reasoning performance while reducing inference cost. Building on this foundation, we introduce C ontro l led A ttribute-Driven R easoning O ptimization ( CLARO ), a reinforcement learning framework designed to internalize these benefits. CLARO guides models to embed high-quality structural attributes, such as read-ability, math density, syntactic compression, and low redundancy, within a user-defined to-ken budget. The proposed method outperforms state-of-the-art baselines across diverse benchmarks, yielding accuracy gains of up to 63.6%, demonstrating that guiding generated output language structure enhances reasoning. Overall, our findings establish that optimizing the thought process structure refines reasoning efficacy, with computational efficiency emerging as a derivative benefit of a clearer thought process. Code and models are available at https://github.com/odedsc/CLARO .