COMP 590H Presentation

Multi-Agent Evolve

instantiates a closed-loop co-evolution framework containing three interacting roles (proposer, solver, and judge) from a shared LLM backbone and jointly optimized via reinforcement learning

CoMAS (Co-evolving MultiAgent Systems)

emphasizes intrinsic interaction rewards, extracting learning signals directly from multi-agent discussion dynamics through an LLM-based judge

MARFT (Multi-Agent Reinforcement Finetuning)

Provides a framework for training LLM-based teams, accounting for differences between agents' roles and their extended conversations

Stronger-MAS (Stronger MultiAgent Systems)

Develops training strategies for systems where agents have different roles and communicate over multiple turns.

MAPoRL (MultiAgent Post Co-Training for Collaborative Large Language Models with Reinforcement Learning)

Uses a verifier that evaluates not only the team's final answer but also the quality of the intermediate discussions.

MALT

Trains agents with different responsibilities in a sequence, helping them improve their individual roles while also improving the team's final performance

MARS

Combines two complementary kinds of agents:

• System 1: Fast, intuitive agents that make quick decisions.

• System 2: Slower, more deliberate agents that use tools and perform deeper reasoning.

Alignment Waltz

where a generation agent and a feedback agent improve together to produce safer and more helpful behavior.

Planning Agent

ProtAgents: can help organize protein-related research workflows.

Tool Use and Execution Agent

ChemCrow: which integrates language-model reasoning with chemistry-specific tools to support complex chemical tasks.

Search and Retrieval Agent

PaperQA2: retrieves information from scientific papers and uses citations to support its responses