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

MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems

ICLR 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 MAS² (recursive self-generation), where a tri-agent meta-MAS (generator, implementer, rectifier) builds and repairs task-specific multi-agent systems, trained with Collaborative Tree Optimization and value-scaled preference loss.

Problem

Automatically generated multi-agent systems follow a rigid generate-once-and-deploy paradigm and are brittle to runtime failures and dynamic environments, while also being constrained by external search modules or single meta-agents.

Approach

A multi-agent system that generates other multi-agent systems: a generator drafts a workflow template, an implementer assigns LLM backbones, and a rectifier monitors execution and repairs the workflow online. The three meta-agents are trained from a collaborative decision tree with path-level credit propagation and a value-weighted DPO-style loss.

Key results

• On the main benchmarks MAS² reaches 89.3% on HotpotQA, 97.0% on HumanEval and 71.3% on MATH, outperforming the best baseline in each category (e.g., ScoreFlow at 86.0% on HotpotQA).
• With unseen stronger LLMs added to the pool, MAS² reaches 90.6% on MATH and 84.0% on Bamboogle, reported as gains of up to 15.1%.
• Cost-performance analysis places MAS² on the Pareto frontier; on Bamboogle it is 12.8 points above SC (GPT-4o) at over 25x lower cost.

Surprising findings

• Several handcrafted and automated MAS baselines underperform single LLMs on some tasks, e.g., MultiPersona drops 4.3 points on MATH and ADAS scores 7.0% on BrowseComp+ versus 13.2% for GPT-4o.

Limitations

• Paper does not report training data size, compute, or training hardware.
• Benchmark evaluation relies on API-hosted LLMs, so results depend on provider model versions.
• Inconsistent benchmark counts (six, seven, eight) across the text; the rectifier's contribution is shown only in ablations on three tasks.

Experimental setup

Datasets: HotpotQA Bamboogle Natural Questions BrowseComp+ HumanEval MBPP MATH

Benchmarks: HotpotQA Bamboogle NQ BrowseComp+ HumanEval MBPP MATH

Metrics: accuracy / pass rate API cost

Baselines: CoT Self-Consistency MedPrompt MultiPersona LLM-Debate DyLAN ADAS MaAS AFlow ScoreFlow

Builds on: MaAS ScoreFlow AFlow FlowReasoner MAS-GPT DPO

Closest papers by meaning

• Agent Primitives: Reuseable Latent Building Blocks for Multi-Agent Systems (ICML 2026 · similarity 0.90)
• EvoMAS: Evolutionary Generation of Multi-Agent Systems (ICML 2026 · similarity 0.90)
• PACE: Two-Timescale Self-Evolution for Small Language Model Agents (NeurIPS 2026 · similarity 0.90)
• Large Language Models Miss the Multi-agent Mark (NeurIPS 2025 · similarity 0.90)
• Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies (ICLR 2026 · similarity 0.89)
• MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks (ICML 2026 · similarity 0.89)