AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

EvoCF: Multi-Agent Collaboration via Agentic Memory-Driven Evolutionary Counterfactual Planning

Haotian Chi, Zeyu Feng, Xingrui Yu, Linbo Luo, Yew-Soon Ong, Ivor Tsang, Hechang Chen, Yi Chang, Haiyan Yin

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EvoCF, an agentic memory-driven evolutionary counterfactual planning framework for discovering multi-agent collaboration strategies for embodied systems via counterfactual plan generation.

Problem

LLM-based planners for multi-agent embodied systems often fail to capture the physical and coordination constraints of real-world environments when planning collaboration strategies.

Approach

Combines agentic memory with evolutionary search over counterfactual plans so that the planner can generate and refine alternative collaboration strategies, as described in the abstract.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add memory-guided counterfactual plan generator (vs random mutations) (ablation)EvoCF · MAP-THOR (2 agents)success rate: 0.68 -> 0.84Improved
add agentic memory-driven evaluator (vs heuristic evaluator) (ablation)EvoCF · MAP-THOR (2 agents)success rate: 0.72 -> 0.84Improved
add symbolic constraint inductor (ablation)EvoCF · MAP-THOR (2 agents)success rate: 0.75 -> 0.84Improved
add memory-guided counterfactual generator, steps (ablation)EvoCF · MAP-THOR (2 agents)average steps: 21.37 -> 18.69Improved
increase retrieved trajectories K from 1 to 5 (ablation)EvoCF evaluator · MAP-THOR (2 agents)success rate: 0.77 -> 0.84Improved
increase retrieved trajectories K from 5 to 10 (ablation)EvoCF evaluator · MAP-THOR (2 agents)success rate: 0.84 -> 0.82Hurt
scale number of agents from 2 to 3EvoCF · MAP-THORsuccess rate: 0.84 -> 0.87Improved
scale number of agents from 3 to 4EvoCF · MAP-THORbalance: 0.81 -> 0.74Hurt
inject transferred constraints into planner (put pots and pans task)Planner · MAP-THOR put pots and pans on stovesuccess rate: 0.60 -> 0.80Improved
inject transferred constraints into planner (open all drawers task)Planner · MAP-THOR open all drawerssuccess rate: 0.40 -> 0.60Improved
use EvoCF instead of LLaMAR with Claude 3.5 SonnetLLaMAR · MAP-THOR (2 agents)success rate: 0.70 -> 0.81Improved
use EvoCF instead of LLaMAR with GPT-4o miniLLaMAR · MAP-THOR (2 agents)success rate: 0.61 -> 0.74Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy