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

BAMAS: Structuring Budget-Aware Multi-Agent Systems

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 1 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 BAMAS, which picks an LLM pool via an ILP under a cost budget, then uses offline REINFORCE to choose a multi-agent collaboration topology, reducing cost up to 86% vs. AutoGen/MetaGPT/ChatDev at comparable accuracy on GSM8K, MBPP, and MATH.

Problem

Existing LLM multi-agent system construction methods focus on performance and treat cost as an afterthought, offering little control over cost-performance trade-offs under an explicit budget.

Approach

Budget-aware construction in three steps: an ILP with tiered decision weights selects a performance-first LLM pool within budget, an offline REINFORCE policy conditioned on task and budget selects one of four collaboration topologies, and the system is instantiated with the chosen LLMs and topology.

Key results

• On MBPP, BAMAS at budget 1,250 reaches 82.6% accuracy vs. 82.2% for MetaGPT/AutoGen-level SOTA, at cost 529.2 vs. 3,735.1 (86% reduction).
• On GSM8K, BAMAS at budget 1,625 reaches 95.3% accuracy vs. 95.4% for AutoGen with DeepSeek-V3, at cost 542.9 vs. 1,425.3 (62% reduction).
• On MATH with budget 2,000, BAMAS reaches 81.2% accuracy vs. 77.6% for the best baseline, at lower cost (646.0 vs. 797.2).

Surprising findings

• The Planner-Driven topology is never selected by the learned policy across datasets and budgets.
• Out-of-budget occurrences are rare, but BAMAS exceeds the budget on up to 30 of 1,000 MATH tasks at budget 2,000 despite the constraint.

Limitations

• Only two LLMs (DeepSeek-V3 and GPT-4.1 nano) are available in experiments, limiting evaluation of LLM pool selection.
• Baselines are restricted to the same two LLMs and fixed configurations, and results rely on a single run with fixed seeds.
• Token-cost estimates are fixed representative values rather than per-task predictions, so budget overruns can occur.

Experimental setup

Datasets: GSM8K MBPP MATH

Benchmarks: GSM8K MBPP MATH

Metrics: accuracy average cost out-of-budget task count

Baselines: AutoGen MetaGPT ChatDev Naive-CostAware

Builds on: AutoGen MetaGPT ChatDev REINFORCE

Closest papers by meaning

• $\texttt{Multi}^2$: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments (ICML 2026 · similarity 0.86)
• Multi-agent Architecture Search via Agentic Supernet (ICML 2025 · similarity 0.86)
• Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments (ICLR 2026 · similarity 0.86)
• Multi-Agent Collaboration via Evolving Orchestration (NeurIPS 2025 · similarity 0.86)
• Conformal Constrained Policy Optimization for Cost-Effective LLM Agents (AAAI 2026 · similarity 0.85)
• MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.85)