// Latent Agents // Multi-agent debate makes models reason better. It also burns tokens generating long transcripts before any answer comes …

// Latent Agents // Multi-agent debate makes models reason better. It also burns tokens generating long transcripts before any answer comes out. This new research distills the entire debate into a single LLM. Latent Agents uses a two-stage fine-tuning pipeline: the model first learns debate structure, then internalizes it through dynamic reward scheduling and length clipping. The result: the internalized model matches or beats explicit multi-agent debate using up to 93% fewer tokens. Activation steering reveals something interesting. Internalization carves out agent-specific subspaces, interpretable directions in activation space corresponding to different agent perspectives. The "agents" survive distillation as identifiable circuits, not just behavior. There's a safety angle too. When malicious agents are deliberately embedded via distillation, negative steering suppresses them more cleanly than steering a base model would, with smaller hits to general performance. Why it matters: most production teams can't afford full debate inference at scale. Internalized debate keeps the reasoning gain without the token tax. Paper: https://arxiv.org/abs/2604.24881 Learn to build effective AI agents in our academy: https://academy.dair.ai/

Source: DAIR.AI (X) | 2026-04-29