Hidden in the Request: Explaining Unethical LLM Compliance through Token Relevance
2026-08-24 • Artificial Intelligence
Artificial IntelligenceComputation and Language
AI summaryⓘ
The authors studied why large language models sometimes fail to act ethically, especially when asked directly for help with unethical requests. They tested models using different kinds of prompts and found that the models struggled more when the request was phrased as a direct assistance request. Using a method called Layer-wise Relevance Propagation, they discovered the models focus more on polite phrasing than on the unethical parts of the prompt. The authors then developed techniques to make models pay more attention to these unethical cues, which helped the models respond more safely.
Large Language ModelsalignmentethicsLayer-wise Relevance Propagationprompt engineeringmodel attributionsafetycompliance failurestoken relevancedecoding methods
Authors
Or Biton, Tomer Krichli, Itai Allouche, Joseph Keshet
Abstract
Although Large Language Models (LLMs) are aligned to optimize for both helpfulness and harmlessness, these dual objectives may conflict, inevitably leading to alignment failures. This work systematically investigates instances where LLMs fail to exhibit ethical behavior. To understand the underlying mechanics of these vulnerabilities, we introduce a probing methodology that presents unethical scenarios to LLMs in three distinct structural modalities: objective classification tasks, subjective first-person statements, and direct requests for assistance. We find that model performance degrades in the request-for-assistance-based form. Using Layer-wise Relevance Propagation (LRP), we trace this discrepancy to an attribution bias: the model places greater emphasis on benign task-framing tokens (e.g., "Can you help me...") than on tokens signaling the underlying unethical behavior (e.g., "without getting caught"), which we term cue-tokens. We hypothesize that this under-attribution contributes to harmful compliance. To test this, we introduce two LRP-guided decoding methods that steer generation toward trajectories more relevant to cue tokens. Empirical evaluations show that these interventions promote safer responses, supporting cue-token attribution's role in compliance failures.