PRIME: Evaluating Prompt Resolution Under Incompatible Instructions in LLMs

2026-06-21

AI Summary

Existing benchmarks evaluate large language models (LLMs) solely on isolated instructions, which fails to reveal their behavior under conflicting directives. This work proposes PRIME, a framework that systematically assesses LLMs’ ability to handle incompatible instructions by constructing controlled conflicts in response length, output format, and reasoning style. Using a deterministic behavioral taxonomy, the study conducts experiments under both balanced and natural distribution settings. Findings indicate that conflict type exerts a significantly stronger influence on model behavior than model scale, leading to the identification of multiple failure modes. Evaluations across five open-source LLMs demonstrate that assessing instruction following on isolated prompts is insufficient for a comprehensive understanding of model capabilities.

Abstract

Large language models (LLMs) often encounter conflicting prompts, although current instruction following benchmarks assess those meta-instructions in isolation, limiting the insights about how models process conflicting instructions. We introduce a framework \textit{PRIME}(\textit{Prompt Resolution under Incompatible Meta-Instructions Evaluation}) to analyze behavior of LLMs when provided with conflicting instructions. \textit{PRIME} purposefully produces calibrated conflicts across response length, output format, and reasoning; classifying model responses with a deterministic behavioral taxonomy. We are evaluating five instruction tuned open weight LLMs in two distinct settings, balanced and naturally distributed. The conclusion we reach upon analysis is that conflict type is more significant in affecting behavior than model scale, and various failure modes across different categories of conflict. Our findings emphasize the value of developing conflict awareness and suggest ability of LLM to follow instructions cannot be assessed through isolated constraints alone.

Problem

• conflicting instructions
• instruction following
• large language models
• prompt resolution
• evaluation benchmark

Innovation

• conflicting instructions
• instruction following
• behavioral taxonomy
• prompt resolution
• large language models