OI-Bench: New Benchmark Tests How Easily LLMs Fall for Misleading Directives

A new research project from arXiv proposes OI-Bench, a standardized benchmark designed to measure how vulnerable large language models (LLMs) are to directive interference—the influence of misleading instructions, social cues, and framing effects embedded in their inputs.

The Challenge#

Prior work has shown that LLMs can be swayed by more than just raw information. Their decisions are susceptible to social signals, how choices are framed, and direct instructions. However, there hasn't been a systematic way to evaluate and compare this vulnerability across different models.

The Approach#

The researchers introduce option injection, a technique that embeds a deceptive directive within an extra choice option in multiple-choice question answering (MCQA) tasks. This combines manipulation of the choice interface itself with interference-based attacks, creating a realistic and scalable evaluation setting.

OI-Bench spans 3,000 questions covering knowledge, reasoning, and commonsense domains. The benchmark includes 16 different types of directives targeting:

• Social compliance — leveraging social pressure or authority
• Bonus framing — incentive-based manipulation
• Threat framing — penalty-based coercion
• Instructional interference — direct command injection

Findings#

The team evaluated 12 LLMs and found substantial vulnerabilities, with attack success rates and robustness varying widely across models. They also tested mitigation strategies ranging from inference-time prompting to post-training alignment techniques.

The benchmark is intended to support more rigorous and systematic evaluation of LLM robustness to directive interference in choice-based interfaces—an important consideration as these systems are deployed in real-world applications.