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

FlowSteer: Towards Agents Designing Agentic Workflows via Reinforced Progressive Canvas Editing

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Software engineering agent benchmarks · 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 FlowSteer, where a lightweight RL-trained policy (Flow-Director) iteratively builds agentic workflows via atomic edits on an executable Workflow Canvas that returns syntax-checked feedback, then a fixed executor LLM runs the designed workflow.

Problem

Existing workflow orchestration relies on human design, lacks graph-level execution feedback, and cannot repair errors during long-horizon workflow construction.

Approach

Treat workflow design as a multi-turn MDP: a small policy issues one atomic graph edit per turn conditioned on canvas feedback, is trained end-to-end with masked GRPO and a structural-then-answer reward, and designer and executor backends are decoupled so the policy transfers across backends and operator libraries.

Key results

• FlowSteer with GPT-4o-mini backend reaches 78.12 EM and 84.98 F1 on HotPotQA, versus 67.19 EM for Qwen3-8B direct and 68.75 EM for AFlow (GPT-4o-mini).
• Average IID accuracy/pass rate of 88.67 versus 84.67 for AFlow and 82.37 for the best agent-RL baseline Orchestrator.
• Average OOD EM of 67.19 versus 63.68 for Orchestrator, with a reported gain of +11.72 over GPT-4o-mini.

Surprising findings

• SFT on Qwen3-8B underperforms the untuned Qwen3-8B baseline on several code benchmarks (e.g., HumanEval 61.72 vs 81.25).

Limitations

• Main results use API-based GPT-4o-mini as executor, so cost and dependence on closed backends remain.
• Reported Table scores lack clear separation of standard deviations and unclear statistical protocol; several ablation numbers appear inconsistent with the main table.

Experimental setup

Datasets: GSM8K MATH HotPotQA SQuAD v2 MBPP HumanEval TriviaQA NaturalQuestions MathQA AIME 2025 APPS DS-1000

Benchmarks: GSM8K MATH HotPotQA SQuAD v2 MBPP HumanEval TriviaQA NaturalQuestions MathQA AIME 2025 APPS DS-1000

Metrics: EM F1 Accuracy Pass@1

Baselines: Qwen3-8B GPT-4o-mini SFT GRPO AFlow AgentFlow Router-R1 Orchestrator

Builds on: GRPO ReAct AFlow DAPO

Closest papers by meaning

• Flow: Modularized Agentic Workflow Automation (ICLR 2025 · similarity 0.89)
• GraphFlow: A Graph-Based Workflow Management for Efficient LLM-Agent Serving (ICML 2026 · similarity 0.88)
• Optimize Once, Execute Fast: Latency-Aware Multi-Agent Workflow Learning for Recurrent Queries (NeurIPS 2026 · similarity 0.88)
• Benchmarking Agentic Workflow Generation (ICLR 2025 · similarity 0.88)
• AFlow: Automating Agentic Workflow Generation (ICLR 2025 · similarity 0.86)
• DyFlow: Dynamic Workflow Framework for Agentic Reasoning (NeurIPS 2025 · similarity 0.86)