Out-of-Distribution Detection using Synthetic Data Generation

2025-02-05

AI Summary

Out-of-distribution (OOD) detection suffers from the scarcity of authentic OOD samples, severely limiting the reliability of existing methods. This paper introduces the first large language model (LLM)-based zero-shot OOD surrogate generation framework: leveraging prompt-driven semantic perturbation modeling, it controllably synthesizes high-fidelity OOD samples without requiring any external OOD data. By innovatively harnessing LLMs’ generative capabilities for OOD detection, the framework enables distribution-agnostic confidence calibration and extends to novel applications—including RLHF reward modeling and alignment detection. Evaluated across nine InD-OOD benchmarks, our method achieves substantial reductions in false positive rates (down to 0% in several cases) while preserving in-distribution classification accuracy, consistently outperforming state-of-the-art baselines.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Distinguishing in- and out-of-distribution (OOD) inputs is crucial for reliable deployment of classification systems. However, OOD data is typically unavailable or difficult to collect, posing a significant challenge for accurate OOD detection. In this work, we present a method that harnesses the generative capabilities of Large Language Models (LLMs) to create high-quality synthetic OOD proxies, eliminating the dependency on any external OOD data source. We study the efficacy of our method on classical text classification tasks such as toxicity detection and sentiment classification as well as classification tasks arising in LLM development and deployment, such as training a reward model for RLHF and detecting misaligned generations. Extensive experiments on nine InD-OOD dataset pairs and various model sizes show that our approach dramatically lowers false positive rates (achieving a perfect zero in some cases) while maintaining high accuracy on in-distribution tasks, outperforming baseline methods by a significant margin.

Problem

• Detecting out-of-distribution inputs for reliable classification systems
• Generating synthetic OOD data using Large Language Models
• Improving OOD detection accuracy and reducing false positive rates

Innovation

• Uses LLMs for synthetic OOD data
• Reduces dependency on external data
• Improves accuracy and lowers false positives