"Dynamic Ensemble Reasoning for LLM Experts"

Dynamic Ensemble Reasoning turns a group of AI models into a well-coordinated team, each contributing their best skills.

→ Achieves 98% of ChatGPT performance using only 10% of parameters

Dynamic Ensemble Reasoning for LLM (DER) introduces a smart agent that picks the right LLM expert at each step, sharing knowledge between models while using minimal computational resources.

-----

https://arxiv.org/abs/2412.07448

🤔 Original Problem:

Using multiple LLMs together is computationally expensive and inefficient, as traditional methods require running all models simultaneously and combining their outputs.

-----

🔧 Solution in this Paper:

→ DER frames LLM ensemble reasoning as a Markov Decision Process where an agent selects which LLM to use next

→ A Knowledge Transfer Prompt enables efficient information sharing between different LLMs

→ The system uses a Terminator mechanism to determine when answers are good enough

→ PPO training optimizes the agent for both answer quality and computational efficiency

-----

💡 Key Insights:

→ Dynamic model selection is more efficient than running all LLMs simultaneously

→ Knowledge transfer between models improves overall performance

→ Role-playing in prompts helps bridge knowledge gaps between different LLMs

-----

📊 Results:

→ 7x parameter reduction compared to full ensemble methods (17B vs 117B parameters)

→ 9% improvement in BARTScore with Knowledge Transfer Prompt

→ Achieves 98% of ChatGPT performance using only 10% of parameters