Permit: Permission-Aware Representation Intervention for Controlled Generation in Large Language Models

2026-05-10

AI Summary

This work addresses the critical challenge of sensitive information leakage in enterprise deployments of large language models, where unauthorized generation during inference circumvents conventional access control mechanisms. To tackle this, the authors propose Permit, a permission-aware representation intervention framework that operates with the backbone model frozen. Permit reveals—for the first time—that permission conditions induce separable and directionally concentrated geometric structures in hidden states, enabling the design of a lightweight intervention mechanism. By analyzing activation disparities to extract a permission-sensitive subspace and applying either offset or gating strategies, Permit achieves fine-grained generation control with minimal computational overhead and high precision. Experiments demonstrate that Permit substantially outperforms existing approaches across diverse permission settings, reducing information leakage to near zero, improving F1 scores by over 18%, and requiring more than 98% fewer trainable parameters.

Abstract

Large language models (LLMs) are increasingly deployed in enterprise settings where they handle sensitive documents and user context, raising acute concerns over security and controllability. Conventional access control regulates whether information is accessible to the model, yet leaves how the model uses that information at generation time largely unconstrained: once sensitive content enters the context, outputs may still drift beyond a user's authorized scope. We present Permit, a novel permission-aware representation intervention framework that closes this gap by enforcing fine-grained control directly on the model's hidden states. Through exploratory analysis, we find that permission conditions induce hidden-state shifts that are (i) separable across permissions and (ii) concentrated in a small set of dominant directions. Permit exploits this geometry in two stages: it first identifies a permission-sensitive subspace from activation differences across permission conditions, and then performs lightweight interventions within this subspace to steer generation, with two concrete instantiations (offset-based and gated). Both operate atop a frozen backbone with only a handful of permission-specific parameters, achieving precise control with minimal overhead. Experimental results demonstrate that Permit performs better than the state-of-the-art method across multiple permission settings while driving information leakage to near zero, achieving over 18% F1-score improvement with >98% fewer trainable parameters.

Problem

• access control
• controlled generation
• large language models
• information leakage
• permission-aware

Innovation

• permission-aware intervention
• representation geometry
• controlled generation
• hidden-state subspace
• parameter-efficient control