Manifold of Failure: Behavioral Attraction Basins in Language Models

2026-02-25

AI Summary

This work addresses the limited understanding of unsafe regions in large language models by introducing the novel concept of “behavioral attraction basins,” reframing safety vulnerability discovery as a quality-diversity optimization problem. Leveraging the MAP-Elites algorithm combined with an “alignment deviation” metric, the method systematically maps the model’s “failure manifold,” revealing its continuous topological structure. Experiments on Llama-3-8B, GPT-OSS-20B, and GPT-5-Mini demonstrate that the approach achieves up to 63% behavioral coverage and identifies as many as 370 distinct vulnerability niches, substantially outperforming baseline methods such as GCG, PAIR, and TAP. This represents a paradigm shift from isolated failure detection to a comprehensive understanding of the global safety landscape.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Large-scale security measurements

Abstract

While prior work has focused on projecting adversarial examples back onto the manifold of natural data to restore safety, we argue that a comprehensive understanding of AI safety requires characterizing the unsafe regions themselves. This paper introduces a framework for systematically mapping the Manifold of Failure in Large Language Models (LLMs). We reframe the search for vulnerabilities as a quality diversity problem, using MAP-Elites to illuminate the continuous topology of these failure regions, which we term behavioral attraction basins. Our quality metric, Alignment Deviation, guides the search towards areas where the model's behavior diverges most from its intended alignment. Across three LLMs: Llama-3-8B, GPT-OSS-20B, and GPT-5-Mini, we show that MAP-Elites achieves up to 63% behavioral coverage, discovers up to 370 distinct vulnerability niches, and reveals dramatically different model-specific topological signatures: Llama-3-8B exhibits a near-universal vulnerability plateau (mean Alignment Deviation 0.93), GPT-OSS-20B shows a fragmented landscape with spatially concentrated basins (mean 0.73), and GPT-5-Mini demonstrates strong robustness with a ceiling at 0.50. Our approach produces interpretable, global maps of each model's safety landscape that no existing attack method (GCG, PAIR, or TAP) can provide, shifting the paradigm from finding discrete failures to understanding their underlying structure.

Problem

• Manifold of Failure
• Behavioral Attraction Basins
• Alignment Deviation
• AI Safety
• Large Language Models

Innovation

• Manifold of Failure
• Behavioral Attraction Basins
• Quality Diversity
• MAP-Elites
• Alignment Deviation