Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate (arxiv.org)

Researchers have introduced a novel approach called "Latent Agents," aimed at enhancing the efficiency and effectiveness of multi-agent debate frameworks for large language models (LLMs). Traditional multi-agent debate techniques require considerable computational resources, generating extensive transcripts before addressing queries. The Latent Agents framework streamlines this process through a two-stage fine-tuning procedure that combines debate structure learning with internalization using dynamic reward scheduling and length clipping. This results in models that can achieve comparable or superior outcomes to explicit debate methods while utilizing up to 93% fewer tokens. The significance of this work lies in its potential to advance the understanding and capabilities of distilled models in the AI/ML community. By revealing that internalization leads to the formation of distinct agent-specific subspaces within activation space, the research offers insights into how different perspectives can be represented and manipulated. A practical outcome demonstrated includes the ability to control harmful behaviors by introducing malicious agents into the LLM and applying negative steering, ultimately leading to better behavior localization with minimal impact on overall performance. These findings provide a foundational shift in how internalized reasoning can be utilized and controlled within AI systems, presenting valuable guidelines for future developments in AI safety and interpretability.