Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate

2026-04-27

AI Summary

This work addresses the high computational cost and verbose outputs of multi-agent debate frameworks for enhancing large language model reasoning. The authors propose a two-stage fine-tuning approach that distills the multi-agent debate structure into a single model, enabling efficient internalized reasoning. Their method integrates debate-structure learning, dynamic reward scheduling, length truncation, and activation steering to carve interpretable agent subspaces within the model’s activation space, facilitating precise suppression of harmful behaviors. Experimental results demonstrate that the internalized model matches or even surpasses explicit multi-agent debate performance across multiple benchmarks, reduces reasoning token usage by up to 93%, and achieves more precise control over harmful content with minimal impact on general capabilities.

Abstract

Multi-agent debate has been shown to improve reasoning in large language models (LLMs). However, it is compute-intensive, requiring generation of long transcripts before answering questions. To address this inefficiency, we develop a framework that distills multi-agent debate into a single LLM through a two-stage fine-tuning pipeline combining debate structure learning with internalization via dynamic reward scheduling and length clipping. Across multiple models and benchmarks, our internalized models match or exceed explicit multi-agent debate performance using up to 93% fewer tokens. We then investigate the mechanistic basis of this capability through activation steering, finding that internalization creates agent-specific subspaces: interpretable directions in activation space corresponding to different agent perspectives. We further demonstrate a practical application: by instilling malicious agents into the LLM through internalized debate, then applying negative steering to suppress them, we show that distillation makes harmful behaviors easier to localize and control with smaller reductions in general performance compared to steering base models. Our findings offer a new perspective for understanding multi-agent capabilities in distilled models and provide practical guidelines for controlling internalized reasoning behaviors. Code available at https://github.com/johnsk95/latent_agents

Problem

• multi-agent debate
• large language models
• reasoning efficiency
• compute-intensive
• internalized reasoning

Innovation

• latent agents
• multi-agent debate distillation
• activation steering
• internalized reasoning
• dynamic reward scheduling