RepV: Safety-Separable Latent Spaces for Scalable Neurosymbolic Plan Verification

2025-10-30

AI Summary

Verifying behavioral compliance of AI systems in safety-critical scenarios remains challenging due to the limited expressiveness of formal methods and the high false-positive/negative rates in natural-language-constraint evaluation. To address this, we propose RepV, a neuro-symbolic verifier that constructs a safety-separable latent space—enabling linear separation of safe versus unsafe plans—and provides position-based probabilistic verification guarantees. RepV integrates model-checking–generated initial labels, a lightweight projector, a frozen linear classifier, and LLM-generated reasoning traces in a synergistic inference framework. It achieves efficient, single-forward-pass verification with <0.2M parameters and zero human annotation. Experiments across multiple tasks demonstrate that RepV improves compliance prediction accuracy by up to 15% over state-of-the-art baselines, significantly outperforming conventional fine-tuning approaches.

Technology Category

• Humans and AI: Human-Aware Planning and Behavior Prediction
• Natural Language Processing: Safety and Robustness
• Computer Vision: Visual Reasoning & Symbolic Representations

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: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

As AI systems migrate to safety-critical domains, verifying that their actions comply with well-defined rules remains a challenge. Formal methods provide provable guarantees but demand hand-crafted temporal-logic specifications, offering limited expressiveness and accessibility. Deep learning approaches enable evaluation of plans against natural-language constraints, yet their opaque decision process invites misclassifications with potentially severe consequences. We introduce RepV, a neurosymbolic verifier that unifies both views by learning a latent space where safe and unsafe plans are linearly separable. Starting from a modest seed set of plans labeled by an off-the-shelf model checker, RepV trains a lightweight projector that embeds each plan, together with a language model-generated rationale, into a low-dimensional space; a frozen linear boundary then verifies compliance for unseen natural-language rules in a single forward pass. Beyond binary classification, RepV provides a probabilistic guarantee on the likelihood of correct verification based on its position in the latent space. This guarantee enables a guarantee-driven refinement of the planner, improving rule compliance without human annotations. Empirical evaluations show that RepV improves compliance prediction accuracy by up to 15% compared to baseline methods while adding fewer than 0.2M parameters. Furthermore, our refinement framework outperforms ordinary fine-tuning baselines across various planning domains. These results show that safety-separable latent spaces offer a scalable, plug-and-play primitive for reliable neurosymbolic plan verification. Code and data are available at: https://repv-project.github.io/.

Problem

• Verifying AI plan compliance with safety rules in critical domains
• Overcoming limitations of formal methods and opaque deep learning
• Creating linearly separable latent spaces for neurosymbolic verification

Innovation

• Learns latent space for linear safety separation
• Uses lightweight projector with language model rationales
• Provides probabilistic guarantees for verification correctness