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

ResMAS: Resilience Optimization in LLM-based Multi-agent Systems

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 ResMAS, a two-stage framework that uses a GNN reward model plus GRPO-trained Qwen2.5-7B topology generator, followed by topology-aware prompt optimization, to improve resilience of LLM multi-agent systems under random agent failures.

Problem

LLM-based multi-agent systems are vulnerable to non-malicious perturbations such as random agent failures, and prior work focuses on reactive attack defense rather than proactively designing resilient topologies and prompts.

Approach

Define MAS resilience as area under the accuracy-vs-error-rate curve, train a GNN reward model to predict per-problem correctness under perturbation to avoid costly evaluation, use it as GRPO reward to fine-tune an LLM topology generator, then refine each agent's prompt using its neighbors and interaction history (positive/negative flip examples).

Key results

• ResMAS achieves the best resilience in all MATH and MMLU-Pro settings, e.g., 0.849 on MATH with 20 agents and 40 edges versus 0.820 for the best baseline (OPRO/TextGrad-level) at similar settings.
• On Chess, ResMAS reaches 0.823 with 20 agents and 40 or 60 edges, outperforming the best baseline GPTSwarm/TextGrad at about 0.81.
• The trained reward model predicts per-problem correctness with 0.86 accuracy on its test split.

Surprising findings

• MAS with random topology and no optimization are far more resilient than a single agent, and resilience grows with agent and edge count.
• Prompt refinement alone, using one GPT-4o call, markedly increases MAS resilience under the same topology.

Limitations

• Resilience is evaluated only under a single perturbation type (random agent failure with random responses).
• Topology and prompt are optimized in separate stages rather than jointly.
• Prompt optimization relies on a stronger LLM (gpt-4o-mini) as optimizer and a small training set of 32 questions per dataset.

Experimental setup

Datasets: MATH MMLU-Pro Chess (BIG-Bench subset) HumanEval

Benchmarks: MATH MMLU-Pro Chess HumanEval

Metrics: resilience (normalized area under accuracy vs error-rate cur accuracy

Baselines: G-Designer OPRO TextGrad GPTSwarm

Builds on: GRPO G-Designer GPTSwarm TextGrad OPRO

Closest papers by meaning

• On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents (ICML 2025 · similarity 0.91)
• SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication (AAAI 2026 · similarity 0.88)
• Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems (NeurIPS 2026 · similarity 0.88)
• AdvEvo-MARL: Shaping Internalized Safety through Adversarial Co-Evolution in Multi-Agent Reinforcement Learning (ICML 2026 · similarity 0.88)
• HieraMAS: Optimizing Intra-Node LLM Mixtures and Inter-Node Topology for Multi-Agent Systems (ICML 2026 · similarity 0.88)
• Empirical Study on Robustness and Resilience in Cooperative Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.87)