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

CARD: Towards Conditional Design of Multi-agent Topological Structures

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 CARD, a conditional variational graph generator that designs LLM multi-agent communication topologies conditioned on environment signals (model, tool, data-source), evaluated on HumanEval, MATH, and MMLU.

Problem

Multi-agent LLM systems use communication topologies that are fixed or statically learned, so they fail to adapt when models, tools, or knowledge sources change.

Approach

Encode agent profiles and runtime condition text with separate graph encoders, decode edge probabilities conditioned on a query embedding, and train with a utility-plus-condition-aware-cost objective so topology can be regenerated at runtime without retraining when conditions change.

Key results

• CARD reaches 90.50% average on HumanEval, 74.50% on MATH, and 86.67% on MMLU, the best averages in Table 1 (G-Designer: 86.50, 72.66, 84.44).
• On MATH with DeepSeek-V3, G-Designer drops from 91.66% to 79.16% on qwen-72B while CARD goes from 91.66% to 82.50%.
• Under node attack on HumanEval, CARD reaches 87.50% vs G-Designer 86.66%.

Surprising findings

• Naive prompt-level condition injection can hurt performance, dropping up to 12.50% on MATH with one base model.
• On simple tasks with a capable model, a single vanilla agent slightly outperforms multi-agent setups due to redundant communication.

Limitations

• Does not update agent-level prompts, profiles, or tool selection in response to environment shifts.
• Graph abstraction may miss domain-specific constraints such as procedural or tool dependencies in software engineering.
• Environmental shifts are simulated (tool/source swaps, node attacks) rather than measured in real deployments.

Experimental setup

Datasets: HumanEval MATH MMLU

Benchmarks: HumanEval MATH MMLU

Metrics: accuracy inference cost (USD per instance)

Baselines: Vanilla CoT Random-graph LLM-Debate GPTSwarm AFlow G-Designer

Builds on: G-Designer GPTSwarm AFlow LLM-Debate chain-of-thought

Closest papers by meaning

• Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation (AAAI 2026 · similarity 0.90)
• AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation (ICML 2026 · similarity 0.90)
• G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks (ICML 2025 · similarity 0.88)
• RADAR: Redundancy-Aware Diffusion for Multi-Agent Communication Structure Generation (ICML 2026 · similarity 0.88)
• Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · similarity 0.87)
• HieraMAS: Optimizing Intra-Node LLM Mixtures and Inter-Node Topology for Multi-Agent Systems (ICML 2026 · similarity 0.87)