LatentGuard: Efficient and Inspectable Latent Reasoning for LLM Safeguards

2026-08-04

AI Summary

This work addresses the high computational cost of existing reasoning-based safety alignment methods for large language models (LLMs), which rely on explicit rationale generation, and the lack of auditability in latent-variable approaches. The authors propose LatentGuard, a novel framework that introduces continuous latent-variable reasoning into LLM safety moderation for the first time. It employs staged curriculum learning to compress textual rationales into compact latent states, which directly inform safety decisions. An isolated auxiliary decoder is integrated to generate audit artifacts on demand, ensuring transparency without compromising efficiency. The method drastically reduces inference overhead—shrinking the critical path from 268.56 to 1.60 tokens—while achieving a weighted F1 score of 84.91 and an audit utility of 85.75, thereby balancing performance and interpretability.

Abstract

Reasoning-based guard models improve LLM safeguards, but decoding explicit rationales for every interaction makes them costly to deploy. Although latent-reasoning methods reduce token generation by moving reasoning into continuous states, they remain underexplored for safety moderation and lack an inspection interface for deployment. In this paper, we propose LatentGuard, an efficient and inspectable safeguard framework that brings continuous latent reasoning to guard models. LatentGuard uses a staged curriculum to progressively compress task-aligned textual rationales into compact latent states, enabling safety verdicts to be predicted directly from continuous representations. To preserve inspectability, an isolated auxiliary decoder generates compact audit artifacts on demand, keeping rationale generation off the standard inference path. Experiments show that LatentGuard-8B improves mean weighted F1 from 83.95 to 84.91 over GuardReasoner-8B, while reducing critical-path reasoning cost from 268.56 generated rationale tokens to 1.60 latent reasoning tokens. Its audit decoder achieves an audit utility score of 85.75, demonstrating an efficient and inspectable path toward deployable LLM safeguards.

Problem

• LLM safeguards
• latent reasoning
• inspectability
• safety moderation
• reasoning cost

Innovation

• latent reasoning
• LLM safeguards
• inspectable AI
• efficient inference
• audit artifacts