Multi-Agent Debate 내재화를 통한 토큰 사용량 93% 절감 및 추론 성능 유지
Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate
Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate
추론 시점 Activation Steering 통한 모델 거부 제거 및 57.07 t/s 성능 달성
Shared expert pool reduces parameters while maintaining performance
You Don't Have to Fine-Tune Your LLM to change it's Behavior. You Can Just… Steer It.