Higher-Order Cognitive Chain-of-Thought Is Enough: Evaluation, Analysis, and Optimization

ICLR 2026 Conference Submission22431 Authors

Keywords: Chain of thought, Bloom's Taxonomy， Cognitive assessment, data distribution

TL;DR: We analyze chain-of-thought data through Bloom’s taxonomy and find most CoTs focus on mid-level reasoning, with higher-order skills underrepresented; adding metacognitive CoTs improves LLM performance.

Abstract:

Chain-of-Thought (CoT) data has become essential for advancing large language models' reasoning capabilities, yet current quality assessment methods neglect the quality of underlying reasoning processes and thus undermine their effectiveness.
To address these challenges, we propose a CoT data quality assessment framework from a cognitive perspective, grounded in Bloom's Taxonomy as our core theoretical foundation.
Through systematic analysis of existing CoT datasets, we reveal that current CoT data exhibits significant distributional biases toward intermediate-order cognitive operations, failing to adequately represent the full spectrum of human-level cognitive capabilities. These findings demonstrate systematic inadequacies in reasoning quality across multiple benchmarks, with models struggling to reproduce sophisticated cognitive processes essential for complex problem-solving. Based on these insights, we propose a simple-yet-effective cognitive-guided CoT data enhancement approach that supplements datasets with minimal higher-order cognitive CoT data.
Consequently, we introduce a simple-yet-effective CoT data enhancement method that rapidly enhances model performance using minimal additional high-order cognitive CoT data, experiments demonstrates the effectiveness of cognitive-aware CoT dataset construction and evaluation.

Primary Area: interpretability and explainable AI

Submission Number: 22431