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

GraphFlow: A Graph-Based Workflow Management for Efficient LLM-Agent Serving

ICML 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 GraphFlow, which represents workflows as a shared operation DAG (wGraph), generates task-specific workflow subgraphs with a GCN+MLP edge scorer, and cuts KV-cache memory via base-plus-sparse-residual caching with path pruning.

Problem

Existing workflow-assisted LLM agent serving relies on predefined templates and shallow retrieval, which generalizes poorly to unseen tasks, and per-workflow KV caching duplicates memory for operations shared across workflows.

Approach

Unify atomic operations from many workflows into one global operation DAG (wGraph); learn a GNN+MLP that scores candidate dependencies conditioned on a virtual task node to synthesize workflows; and store each operation's KV cache as a context-free base plus sparse prefix-specific residuals, pruning rarely used transitions.

Key results

• On Qwen-2.5-7B, GraphFlow reaches 92.1 GSM8K accuracy and 86.2 HumanEval pass@1, versus 89.2 and 78.1 for the best baseline AFlow.
• Reported average gain of about 4.95 points across five benchmarks and about 4x KV-cache memory reduction versus stateful caching.
• Stateless KV caching drops MATH accuracy to about 39.4% while the differential design retains about 52.6% vs 53.8% for stateful on Llama-3.1-8B.

Limitations

• Workflow generation relies on supervision from GPT-4o-synthesized traces and offline mined workflows, which may limit generalization.
• Gains reported only on five benchmarks and three backbones; latency gains are not uniform across settings.
• Path pruning falls back to on-the-fly KV computation for rare paths, adding latency in those cases.

Experimental setup

Datasets: GSM8K MATH HotpotQA HumanEval MBPP

Benchmarks: GSM8K MATH HotpotQA HumanEval MBPP

Metrics: accuracy F1 pass@1 P90 latency KV memory footprint

Baselines: Vanilla MetaGPT LLMCompiler TaskWeaver AgentKB AutoFlow AFlow

Builds on: AFlow LLMCompiler TaskWeaver MetaGPT PagedAttention

Closest papers by meaning

• Benchmarking Agentic Workflow Generation (ICLR 2025 · similarity 0.94)
• Flow: Modularized Agentic Workflow Automation (ICLR 2025 · similarity 0.91)
• Optimize Once, Execute Fast: Latency-Aware Multi-Agent Workflow Learning for Recurrent Queries (NeurIPS 2026 · similarity 0.90)
• DyFlow: Dynamic Workflow Framework for Agentic Reasoning (NeurIPS 2025 · similarity 0.90)
• KVFlow: Efficient Prefix Caching for Accelerating LLM-Based Multi-Agent Workflows (NeurIPS 2025 · similarity 0.89)
• FlowSteer: Towards Agents Designing Agentic Workflows via Reinforced Progressive Canvas Editing (NeurIPS 2026 · similarity 0.88)