InsLen Score: Detecting Object Hallucination in Multimodal Large Models Using the Instruction Itself

A study at ICML 2026 proposes InsLen Score (Instruction Lens Score), which can effectively detect object hallucination in multimodal large language models (MLLMs) by analyzing the user instruction itself—without additional training or reference images. This provides a new proactive prevention approach for hallucination management in multimodal AI systems.

Multimodal large language models (MLLMs) have made significant progress in tasks like image understanding, but object hallucination remains a prominent issue—models may mention objects that do not exist in the image. Traditional detection methods require complex post-processing, additional reference images, or expensive fine-tuning, which are costly and difficult to deploy quickly. The industry needs simple and efficient solutions.

Core Insight of InsLen Score: The user instruction itself contains sufficient detection signals—certain wording and structures are correlated with hallucination rates. The technical implementation involves three layers of analysis:

• Lexical level: The number of object nouns is positively correlated with hallucination.
• Syntactic structure: Complex sentence structures increase hallucination probability.
• Semantic intent: Instructions beyond image reasoning easily induce hallucinations. A comprehensive risk score between 0 and 1 is output.

Zero-shot validation was conducted on models like LLaVA and MiniGPT-4. InsLen Score achieved better detection accuracy on the POPE benchmark than some baseline methods that require additional training. Its computational overhead is extremely low (millisecond-level), does not require model internal states, and can be integrated in real time.

Developers can integrate InsLen Score:

• When the score exceeds a threshold, prompt users to optimize the instruction or add constraints.
• Trigger review/secondary verification by combining with model confidence.
• Optimize prompt templates by analyzing high-risk instruction features to reduce hallucinations at the source.

Limitations:

• Only targets object-level hallucinations; limited detection of errors in complex scenarios/relationship reasoning.
• Validated based on English; multilingual performance needs further research.
• No specific instruction improvement suggestions. Future directions: Expand to complex scenarios, support multilingual use, and develop automatic instruction optimization modules.

The project has been open-sourced on GitHub, providing PyTorch implementations, pre-trained models, and example scripts, supporting pip installation and API calls. The community is welcome to submit feedback and contributions via Issues and PRs.