MemLens: Uncovering Memorization in LLMs with Activation Trajectories

2025-09-25

AI Summary

Large language models (LLMs) exhibit susceptibility to training data contamination on benchmarks such as AIME and Math500; existing detection methods—relying on lexical overlap or perplexity—suffer from poor generalization and fail to identify implicit contamination. To address this, we propose a novel memorization detection paradigm based on **digit-token probability activation trajectories**: contaminated samples exhibit premature answer locking (“shortcut reasoning”) in early Transformer layers, whereas clean samples accumulate evidence incrementally across layers, yielding discriminative dynamic trajectories. We further validate the causal validity and authenticity of these trajectories via LoRA-based fine-tuning to inject controlled contamination samples. Our method significantly improves detection accuracy for implicit contamination, achieves superior cross-task generalization across multiple benchmarks, and provides an interpretable, empirically verifiable tool for trustworthy LLM evaluation.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Large language models (LLMs) are commonly evaluated on challenging benchmarks such as AIME and Math500, which are susceptible to contamination and risk of being memorized. Existing detection methods, which primarily rely on surface-level lexical overlap and perplexity, demonstrate low generalization and degrade significantly when encountering implicitly contaminated data. In this paper, we propose MemLens (An Activation Lens for Memorization Detection) to detect memorization by analyzing the probability trajectories of numeric tokens during generation. Our method reveals that contaminated samples exhibit ``shortcut'' behaviors, locking onto an answer with high confidence in the model's early layers, whereas clean samples show more gradual evidence accumulation across the model's full depth. We observe that contaminated and clean samples exhibit distinct and well-separated reasoning trajectories. To further validate this, we inject carefully designed samples into the model through LoRA fine-tuning and observe the same trajectory patterns as in naturally contaminated data. These results provide strong evidence that MemLens captures genuine signals of memorization rather than spurious correlations.

Problem

• Detecting memorization in LLMs contaminated benchmark data
• Overcoming limitations of surface-level lexical overlap methods
• Identifying shortcut behaviors in contaminated samples trajectories

Innovation

• Analyzing numeric token probability trajectories
• Detecting shortcut behaviors in early layers
• Validating with LoRA fine-tuning injection