PRIME: Evaluating Prompt Resolution Under Incompatible Instructions in LLMs

As large language models (LLMs) are increasingly used in complex, real-world applications, they frequently encounter prompts containing mutually exclusive instructions. Current benchmarks primarily assess whether models can follow individual, compatible constraints. This paper investigates how LLMs behave when faced with contradictory meta-instructions—situations where satisfying one directive necessitates violating another—and whether these behaviors are systematic or model-specific.

The authors introduce PRIME (Prompt Resolution under Incompatible Meta-Instructions Evaluation), a framework designed to analyze LLM behavior under controlled conflict. They constructed PRIMEBench, a dataset of 72 source tasks categorized into arithmetic, logic, and conceptual reasoning. For each task, they injected three types of conflicts: reasoning (step-by-step vs. final answer only), length (detailed vs. brief), and format (bullet points vs. no bullets). The researchers evaluated five open-weight LLMs (TinyLlama, Qwen2, Gemma, StableLM, and Mistral) using a deterministic taxonomy to classify responses as following the first instruction, the second instruction, both, or neither.

The study reveals that the nature of the conflict is the primary driver of model behavior. Reasoning conflicts consistently result in the highest rates of instruction neglect (where the model satisfies neither directive). In contrast, format conflicts are typically resolved by the model choosing one of the two instructions. When faced with length conflicts, models show a strong, consistent bias toward providing detailed responses, regardless of the brevity constraint.

Furthermore, the researchers found that directive preference—the tendency to favor the first or second instruction—is a stable, model-specific trait. For instance, Gemma and Mistral consistently favored the first instruction, while TinyLlama favored the second. These preferences remained stable across different task distributions, suggesting that conflict resolution is deeply tied to a model's specific alignment training rather than its parameter scale.