CLAWS:Creativity detection for LLM-generated solutions using Attention Window of Sections

TL;DR

This work proposes CLAWS, a method that defines and classifies mathematical solutions into typical, creative, and hallucinated categories without human evaluation, by leveraging attention weights across prompt sections and output.

Abstract

Recent advances in enhancing the reasoning ability of large language models (LLMs) have been remarkably successful. LLMs trained with reinforcement learning (RL) for reasoning demonstrate strong performance in challenging tasks such as mathematics and coding, even with relatively small model sizes. However, despite these improvements in task accuracy, the assessment of creativity in LLM generations has been largely overlooked in reasoning tasks, in contrast to writing tasks. The lack of research on creativity assessment in reasoning primarily stems from two challenges: (1) the difficulty of defining the range of creativity, and (2) the necessity of human evaluation in the assessment process. To address these challenges, we propose CLAWS, a method that defines and classifies mathematical solutions into typical, creative, and hallucinated categories without human evaluation, by leveraging attention weights across prompt sections and output. CLAWS outperforms five existing white-box detection methods (Perplexity, Logit Entropy, Window Entropy, Hidden Score, and Attention Score) on five 7-8B math RL models (DeepSeek, Qwen, Mathstral, OpenMath2, and Oreal). We validate CLAWS on 4545 math problems collected from 181 math contests (AJHSME, AMC, AIME).

Figures (10)

• Figure 1: Overview of the proposed framework. Step 1: LLM generates a solution. Step 2: Features are extracted during generation. Step 3: LLM Evaluator assigns labels (Hallucinated / Creative / Typical). Step 4: Detection methods are evaluated by comparing predictions with the labels.
• Figure 2: Example of an input prompt $X$ used to elicit creative solutions from LLMs. Prompt consists of four sections: Guideline ($G$, yellow), Problem ($P$, green), Reference Solutions ($S$, blue), and Instruction ($I$, purple).
• Figure 3: Architecture of CLAWS. All tokens are segmented into five sections: Guideline, Problem, Solutions, Instruction, and Response. The average attention weight for each section ($\mathrm{AVGA}_\mathcal{U}$) is computed, normalized to obtain section-wise attention ratios ($\mathrm{CLAWS}_\mathcal{U}$), and used as features for detecting Hallucinated, Creative, or Typical solutions.
• Figure 4: Visualization of class-wise average scores for each method, computed on the reference set generated using Qwen2.5-math-7B-inst. To enhance visual clarity, the normalization range is clipped between 0.1 and 0.9 for CLAWS. Visualizations for all models are presented in Figure \ref{['fig:visual4']}.
• Figure 4: Impact of each prompt section. Performance comparison of the DeepSeek-Math-7B-RL using the Prototype strategy through an ablation study, where each section is selectively removed.