agent/multi - agent

MACNET - Scaling Large-Language-Model-based Multi-Agent Collaboration 논문리뷰

jinuklee 2024. 8. 5. 16:44

https://arxiv.org/pdf/2406.07155

LLM의 power law와 같이 agent의 수를 늘린다고 해서 emergent capability가 생기는가?

Inspired by the neural scaling law, a natural question arises: does increasing agents in multi-agent collaboration exhibit emergent capabilities?

 

3. Multi-Agent Collaboration Network

 

This arrangement ensures that each node-occupied agent ai precedes its corresponding edge-occupied agent aij , and aij precedes aj , thereby guaranteeing ensuring orderly information transmission

Specifically, ai requests feedback, aij offers optimization suggestions and requests further refinement, and aj provides the refined solutio

 

3.3 Memory Control

 

Context  문제 해결을 위해

we adopt a heuristic mechanism(chatdev)

Short-term memory captures the intra-interaction working memory during each dual-agent interaction, ensuring contextaware decision-making. Long-term memory maintains inter-interaction context continuity by transmitting only the final solutions derived from dialogues, not the entire conversational history

 

마지막의 convergent agent가 solution을 종합해 강점을 살리고 약점을 버리는(assess solution's weakness and strength) 합성과정 진행(synthesizing