LatentGuard: Efficient and Inspectable Latent Reasoning for LLM Safeguards
LatentGuard is a safeguard framework that uses latent reasoning to reduce the cost of LLM safety moderation. It compresses textual rationales into compact latent states via a staged curriculum, predicting safety verdicts directly from continuous representations. An isolated auxiliary decoder generates compact audit artifacts on demand, preserving inspectability. 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.
LatentGuard introduces an efficient and inspectable approach to LLM safeguards by moving reasoning into continuous latent states, significantly reducing token generation costs while maintaining safety performance. It outperforms GuardReasoner-8B on mean weighted F1 and drastically cuts reasoning tokens, with an on-demand audit decoder providing inspectability.
LatentGuard employs a staged curriculum to compress task-aligned textual rationales into compact latent states, enabling direct safety verdict prediction from continuous representations. The isolated auxiliary decoder generates audit artifacts only when needed, keeping rationale generation off the critical inference path. This design achieves a 168x reduction in reasoning tokens (from 268.56 to 1.60) while improving safety metrics.
The approach addresses a key deployment challenge for reasoning-based guard models: the high cost of generating explicit rationales for every interaction. By reducing token usage, LatentGuard makes advanced safety moderation more economically viable for real-world applications, potentially accelerating adoption in production systems.
LatentGuard offers a cost-effective solution for deploying LLM safeguards at scale, reducing inference costs while maintaining or improving safety performance. This can lower barriers for companies to implement robust content moderation, especially in high-volume applications.
Future work may explore scaling LatentGuard to larger models, integrating it with other safety techniques, and evaluating its robustness against adversarial attacks. The inspectability feature could also be extended to support regulatory compliance and human oversight requirements.