Learning Safety Constraints for Large Language Models

2025-05-30

AI Summary

Large language models (LLMs) face a fundamental trade-off between safety—e.g., mitigating harmful outputs and adversarial attacks—and task performance degradation. Method: This paper introduces Safety Polytope (SaP), a geometric approach that explicitly encodes multidimensional safety constraints within the LLM’s latent space. SaP represents safe and unsafe regions as interpretable polyhedral boundaries, enabling weight-invariant, post-hoc safety calibration. Leveraging geometric deep learning and adversarial robustness optimization, SaP induces spontaneous specialization of polyhedral facets across distinct safety dimensions (e.g., ethics, factual consistency). Contribution/Results: SaP significantly improves detection of ethically violating inputs (+23.6%), reduces adversarial attack success rates (−41.2%), and preserves original task performance without degradation. Crucially, it provides human-interpretable, geometry-grounded safety decisions—offering the first framework to unify rigorous safety enforcement with transparent, constraint-aware latent-space reasoning in LLMs.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Security and Privacy: Security and privacy of machine learning and AI applications

Abstract

Large language models (LLMs) have emerged as powerful tools but pose significant safety risks through harmful outputs and vulnerability to adversarial attacks. We propose SaP, short for Safety Polytope, a geometric approach to LLM safety that learns and enforces multiple safety constraints directly in the model's representation space. We develop a framework that identifies safe and unsafe regions via the polytope's facets, enabling both detection and correction of unsafe outputs through geometric steering. Unlike existing approaches that modify model weights, SaP operates post-hoc in the representation space, preserving model capabilities while enforcing safety constraints. Experiments across multiple LLMs demonstrate that our method can effectively detect unethical inputs, reduce adversarial attack success rates while maintaining performance on standard tasks, thus highlighting the importance of having an explicit geometric model for safety. Analysis of the learned polytope facets reveals emergence of specialization in detecting different semantic notions of safety, providing interpretable insights into how safety is captured in LLMs' representation space.

Problem

• Preventing harmful outputs from large language models
• Reducing vulnerability to adversarial attacks in LLMs
• Enforcing safety constraints without modifying model weights

Innovation

• Geometric approach enforces safety constraints
• Post-hoc representation space operation
• Interpretable polytope facets specialization