Large language models could be rote learners

2025-04-11

AI Summary

This work addresses benchmark contamination in LLM evaluation, which conflates rote memorization with genuine capability. We propose TrinEval, a novel trinary evaluation framework that treats contamination as an inherent aspect of the learning process and decouples memory from reasoning via question-type reconstruction. Methodologically, TrinEval integrates performance attribution analysis, MCQ-type re-modeling, and controlled memory-condition experiments. Empirical evaluation on MMLU reveals that ~20.5% of mainstream LLM responses stem from mechanical memorization—and such memorized answers degrade accuracy by an average of −3.7%. TrinEval quantitatively isolates memory effects from authentic reasoning ability, enabling contamination-robust assessment. It establishes a new paradigm for LLM evaluation, shifting focus from mere answer correctness (“how many are right?”) to causal understanding of correct responses (“why is it right?”).

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Multiple-choice question (MCQ) benchmarks are widely used for evaluating Large Language Models (LLMs), yet their reliability is undermined by benchmark contamination. In this study, we reframe contamination as an inherent aspect of learning and seek to disentangle genuine capability acquisition from superficial memorization in LLM evaluation. First, by analyzing model performance under different memorization conditions, we uncover a counterintuitive trend: LLMs perform worse on memorized MCQs than on non-memorized ones, indicating the coexistence of two distinct learning phenomena, i.e., rote memorization and genuine capability learning. To disentangle them, we propose TrinEval, a novel evaluation framework that reformulates MCQs into an alternative trinity format, reducing memorization while preserving knowledge assessment. Experiments validate TrinEval's effectiveness in reformulation, and its evaluation reveals that common LLMs may memorize by rote 20.5% of knowledge points (in MMLU on average).

Problem

• Assessing LLM genuine capability vs memorization in MCQ benchmarks
• Reducing benchmark contamination impact on LLM evaluation reliability
• Proposing TrinEval to disentangle rote learning from knowledge acquisition

Innovation

• Analyze model performance under memorization conditions
• Propose TrinEval framework for MCQ reformulation
• Reduce memorization while preserving knowledge assessment