Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs

Researchers demonstrate that large language models exhibit brittle instruction-following when faced with competing behavioral patterns, with compliance rates ranging from 1% to 99% across 13 models. The study reveals that output diversity and format—rather than reasoning ability—are the primary determinants of robustness against induction pressure, highlighting fundamental vulnerabilities in current LLM training.

This research exposes a critical tension between two core capabilities in large language models: their ability to follow explicit instructions versus their tendency to complete patterns based on training data. The study constructs adversarial scenarios where user instructions directly conflict with demonstrated behavioral patterns, measuring how models navigate these contradictions across multiple scales and domains. The findings are striking: instruction-following rates vary dramatically across models independent of their overall capability scores, suggesting that robustness against induction pressure represents a separate, underexplored dimension of model reliability.

The research builds on growing concerns about alignment and control in advanced language models. While previous work has examined instruction-following and jailbreaking vulnerabilities separately, this study uniquely quantifies the head-to-head competition between these opposing forces. The observation that models can reason correctly through chain-of-thought prompting while still producing incorrect outputs indicates a disconnect between internal deliberation and behavioral output—a finding with significant implications for interpretability and safety.

For the AI industry, these results suggest that current benchmarks measuring instruction-following may be incomplete. Organizations deploying LLMs in sensitive applications—financial advisory, medical guidance, security systems—cannot rely solely on standard capability metrics to assess reliability. The finding that output diversity provides robustness introduces a potential design principle for more stable models, though it remains unclear how to balance this with the efficiency demands of production systems.

Future research should investigate whether architectural changes or training methodologies can fundamentally resolve this tension rather than merely mitigating it. The systematic underestimation models show regarding their own susceptibility also raises questions about model self-awareness and calibration reliability.

• →Instruction-following rates across 13 models range from 1% to 99% when competing with hardcoded patterns, largely independent of standard capability benchmarks.
• →Output format diversity provides significantly more robustness against induction pressure than single-token outputs, suggesting format design matters more than semantic reasoning.
• →Chain-of-thought reasoning can improve robustness but creates dissociation between correct internal reasoning and incorrect final outputs.
• →Models systematically underestimate their own susceptibility to induction pressure by an average of 16.5%, raising concerns about model self-calibration.
• →Instruction robustness correlates more with alignment to trained value priors than with overall model capability, indicating a distinct vulnerability dimension.