Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MIRROR: Modular Internal Processing for Personalized Safety in LLM Dialogue

Nicole Hsing

TL;DR

MIRROR addresses persistent safety failures in personalized multi-turn dialogue by separating immediate response (Talker) from asynchronous deliberation (Thinker) and by maintaining a bounded, regenerating internal state that conserves user-specific constraints across turns. The Thinker consists of an Inner Monologue Manager and a Cognitive Controller that generate parallel reasoning threads and synthesize them into a compact internal state used by the Talker, enabling safe, responsive interactions with minimal latency. On CuRaTe, MIRROR improves safety by an average of $21%$ relative (from $69%$ to $84%$) across seven models, with open-source Llama 4 Scout reaching $91%$, often outperforming frontier baselines like GPT-4o and Claude 3.7 Sonnet and even their MIRROR-augmented variants. This production-oriented architectural augmentation democratizes AI safety by enabling safer open-source systems at low cost and demonstrates that architectural innovations can equalize safety benefits across model scales.

Abstract

Large language models frequently generate harmful recommendations in personal multi-turn dialogue by ignoring user-specific safety context, exhibiting sycophantic agreement, and compromising user safety for larger group preferences. We introduce MIRROR, a modular production-focused architecture that prevents these failures through a persistent, bounded internal state that preserves personal conversational information across conversational turns. Our dual-component design inspired by Dual Process Theory separates immediate response generation (Talker) from asynchronous deliberative processing (Thinker), which synthesizes parallel reasoning threads between turns with marginal latency. On the CuRaTe personalized safety benchmark, MIRROR-augmented models achieve a 21% relative improvement (69% to 84%) across seven diverse frontier models, with open-source Llama 4 and Mistral 3 variants surpassing both GPT-4o and Claude 3.7 Sonnet at only \$0.0028 to \$0.0172 additional cost per turn, narrowing the gap between affordable open-source models to frontier systems in the safety space. The modular architecture enables flexible deployment: full internal processing for affordable models or single-component configurations for expensive systems, democratizing access to safer, personalized AI.

MIRROR: Modular Internal Processing for Personalized Safety in LLM Dialogue

TL;DR

MIRROR addresses persistent safety failures in personalized multi-turn dialogue by separating immediate response (Talker) from asynchronous deliberation (Thinker) and by maintaining a bounded, regenerating internal state that conserves user-specific constraints across turns. The Thinker consists of an Inner Monologue Manager and a Cognitive Controller that generate parallel reasoning threads and synthesize them into a compact internal state used by the Talker, enabling safe, responsive interactions with minimal latency. On CuRaTe, MIRROR improves safety by an average of relative (from to ) across seven models, with open-source Llama 4 Scout reaching , often outperforming frontier baselines like GPT-4o and Claude 3.7 Sonnet and even their MIRROR-augmented variants. This production-oriented architectural augmentation democratizes AI safety by enabling safer open-source systems at low cost and demonstrates that architectural innovations can equalize safety benefits across model scales.

Abstract

Large language models frequently generate harmful recommendations in personal multi-turn dialogue by ignoring user-specific safety context, exhibiting sycophantic agreement, and compromising user safety for larger group preferences. We introduce MIRROR, a modular production-focused architecture that prevents these failures through a persistent, bounded internal state that preserves personal conversational information across conversational turns. Our dual-component design inspired by Dual Process Theory separates immediate response generation (Talker) from asynchronous deliberative processing (Thinker), which synthesizes parallel reasoning threads between turns with marginal latency. On the CuRaTe personalized safety benchmark, MIRROR-augmented models achieve a 21% relative improvement (69% to 84%) across seven diverse frontier models, with open-source Llama 4 and Mistral 3 variants surpassing both GPT-4o and Claude 3.7 Sonnet at only \0.0172 additional cost per turn, narrowing the gap between affordable open-source models to frontier systems in the safety space. The modular architecture enables flexible deployment: full internal processing for affordable models or single-component configurations for expensive systems, democratizing access to safer, personalized AI.

Paper Structure

This paper contains 90 sections, 8 figures, 16 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Advances in Conversational Reasoning and Memory Systems
  4. Safety and Alignment in Conversational Systems
  5. Cognitive Science Foundations
  6. The MIRROR Architecture
  7. Overview
  8. The Thinker
  9. Inner Monologue Manager
  10. Cognitive Controller
  11. The Talker
  12. Processing Flow
  13. Results
  14. Experimental Setup and Benchmark Selection
  15. Safety Performance Results
  16. ...and 75 more sections

Figures (8)

  • Figure 1: An overview of the MIRROR architecture.
  • Figure 2: MIRROR component overview showing the information consolidation cycle.
  • Figure 3: Visualization of the Inner Monologue Manager's reasoning process.
  • Figure 4: Visualization of the Cognitive Controller's internal state process.
  • Figure 5: Mean success rate comparison across models showing absolute performance.
  • ...and 3 more figures