Skip to content

CARD: Towards Conditional Design of Multi-agent Topological Structures

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Conditional topology design for multi-agent LLM collaboration structures

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CARD (Conditional Agentic Graph Designer) is a framework that generates adaptive multi-agent communication topologies conditioned on dynamic environmental signals (model versions, tool availability, data-source quality), rather than relying on fixed or statically-learned graphs. It formalizes AMACP (Adaptive Multi-Agent Communication Protocol) and instantiates it via a conditional variational graph encoder-decoder. On HumanEval, MATH, and MMLU, CARD outperforms static and prompt-based topology baselines while adapting topology at runtime without retraining.

Problem

Existing multi-agent LLM topologies are either manually designed (brittle) or automatically learned but statically fixed—they cannot adapt when runtime conditions change (e.g., model upgrades from GPT-4o to GPT-5, tool unavailability, degraded data sources). Static topologies produce redundant interactions or disrupted information flows under such shifts, and naive prompt-level condition injection can actively hurt performance (up to −12.50% on MATH).

Method

CARD represents each agent \(v_i\) via two channels: a profile vector \(P_i\) (role, model base, tools) and a condition vector \(C_i\) (runtime environment state). The environment condition is a Cartesian product of semantic feature spaces \(C = F_1 \times F_2 \times \cdots \times F_k\), encoded via a pretrained language model into a shared embedding space.

A conditional graph encoder-decoder then operates as follows: - Two separate graph encoders \(\phi_p, \phi_c\) produce latent profile states \(H^p\) and condition states \(H^c\) over an anchor topology \(A\) (chain, star, or fully connected). - A decoder \(\psi_\theta\) estimates pairwise edge probabilities conditioned on both channels and a query embedding \(h_Q\): $\(\psi(S \mid H^p, H^c) = \prod_{i,j} \psi\!\left(S_{ij} \mid h_i^p, h_i^c, h_j^p, h_j^c, h_Q; \Theta_d\right)\)$ - Edges are activated by thresholding: \(E_{\text{com}} = \{(v_i, v_j) \mid S_{ij} > \tau\}\).

Training minimizes the AMACP objective \(\mathcal{L}_{\text{CARD}} = -u(\alpha^{(K)}) + \beta \cdot w(G_{\text{com}}; C)\), where \(u(\cdot)\) is task utility (accuracy) and \(w(\cdot)\) is a condition-aware communication cost regularizer \(\sum_{(i,j)} \text{Cost}_{ij} \cdot p_{ij}\) over soft edge probabilities. At runtime, topology updates require only one forward pass through refreshed condition embeddings—no retraining.

Key Contributions

  • Formalization of AMACP, a protocol specifying effectiveness, cost-efficiency, and adaptiveness requirements for multi-agent communication topologies.
  • CARD, a conditional graph-generation framework with separate profile and condition encoder channels, trained end-to-end on the AMACP objective.
  • Runtime adaptation via one-pass re-decoding from updated condition embeddings, enabling zero-retraining topology updates.
  • Empirical demonstration that structured condition embedding outperforms prompt-level condition injection, which can degrade performance.
  • Topology visualization showing interpretable behavioral patterns (e.g., weaker models generate denser graphs; lower-quality retrieval increases external-knowledge reliance).

Results

  • HumanEval: CARD 90.50% avg vs. best static baseline Aflow 89.83%; wins or ties 13/15 model–benchmark combinations.
  • MATH: CARD 74.50% avg vs. Aflow 80.83% (note: CARD leads on out-domain generalization; G-Designer drops from 91.66% → 79.16% switching deepseek-v3 → qwen-72B, CARD drops less: 91.66% → 82.50%).
  • MMLU: CARD 86.67% avg vs. G-Designer 84.44% and Aflow 82.87%.
  • Ablation: prompt conditioning (w/ Cond.\(^p\)) causes up to −12.50% on MATH; CARD yields +0.83% to +3.34% on MATH, +0.66% to +2.62% on MMLU, +2.50% to +23.33% on HumanEval vs. unconditioned baseline.
  • Topology correlation analysis: search-engine swap preserves global structure (\(r = 0.9797\), \(p = 0.0006\)) while adjusting local edge weights; lowest-capacity + lowest-quality search produces densest graph.

Limitations

  • Evaluated only under simulated environmental changes (model swaps, tool perturbations); real-world deployment validation is absent.
  • Scalability to large agent ensembles is deferred to future work.
  • The threshold \(\tau\) and trade-off \(\beta\) require tuning; sensitivity analysis details are in appendices but not ablated in the main results.
  • Online/continual adaptation (reinforcement learning) is not yet integrated; runtime adaptation is limited to one-pass re-decoding.
  • MATH average (74.50%) does not clearly exceed all baselines (Aflow 80.83% avg), suggesting the benefit is concentrated in out-domain generalization rather than overall peak performance.

Relevance to Agentic AI / LLM Agents

CARD directly addresses a core bottleneck in deploying multi-agent LLM systems at scale: communication topology brittleness under real-world churn in model versions, tools, and data sources. By framing topology as a conditional generative problem rather than a static optimization, it provides a principled mechanism for agents to reconfigure their coordination structure without human intervention or retraining—a key capability for autonomous, long-running agent pipelines. This connects to the broader trend of treating multi-agent architecture itself as a learnable, adaptive component (alongside GPT-Swarm, G-Designer, Aflow), extending that line toward environment-conditioned, zero-retraining reconfiguration.