Research map / Large language models / Agents, code and math
FlowBot: Inducing LLM Workflows with Bilevel Optimization and Textual Gradients
Hongyeon Yu, Young-Bum Kim, Yoon Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add bilevel workflow induction on top of prompt optimization (ablation) | FlowBot · HotpotQA1 | score: 60.07 -> 72.80 | Improved |
| add bilevel optimization formulation (ablation) | FlowBot · HotpotQA1 | score: 61.73 -> 72.80 | Improved |
| add bilevel optimization formulation (ablation) | FlowBot · HotpotQA2 | score: 79.26 -> 80.19 | Improved |
| add layer-by-layer textual gradient backpropagation (ablation) | FlowBot · HotpotQA1 | score: 71.13 -> 72.80 | Improved |
| add layer-by-layer textual gradient backpropagation (ablation) | FlowBot · HotpotQA2 | score: 76.51 -> 80.19 | Improved |
| add workflow induction vs prompt-only optimization (Qwen3.5-4B executor) (ablation) | Qwen3.5-4B · HotPotQA1 | score: 0.513 -> 0.681 | Improved |
| add workflow induction vs prompt-only optimization (Qwen3.5-397B executor) (ablation) | Qwen3.5-397B-A17B · HotPotQA2 | score: 0.816 -> 0.806 | Hurt |
| use FlowBot instead of GEPA | GPT-4.1 mini · HoVer | score: 51.67 -> 63.20 | Improved |
| use FlowBot instead of GEPA+Merge | GPT-4.1 mini · PUPA | score: 96.46 -> 90.67 | Hurt |
| use FlowBot instead of reimplemented AFlow | GPT-4.1 mini · DROP | score: 79.69 -> 92.28 | Improved |
| use FlowBot instead of reimplemented AFlow | GPT-4.1 mini · MBPP | score: 74.78 -> 73.84 | Hurt |
| induce one unified workflow instead of separate ones | FlowBot Qwen3.5-397B · 6 AFlow tasks average | score: 87.85 -> 80.48 | Hurt |
Closest papers by meaning
- Optimize Once, Execute Fast: Latency-Aware Multi-Agent Workflow Learning for Recurrent Queries (NeurIPS 2026 · similarity 0.88)
- WorkflowLLM: Enhancing Workflow Orchestration Capability of Large Language Models (ICLR 2025 · similarity 0.88)
- AFlow: Automating Agentic Workflow Generation (ICLR 2025 · similarity 0.87)
- Benchmarking Agentic Workflow Generation (ICLR 2025 · similarity 0.87)
- FlowSteer: Towards Agents Designing Agentic Workflows via Reinforced Progressive Canvas Editing (NeurIPS 2026 · similarity 0.86)
- Flow-of-Options: Diversified and Improved LLM Reasoning by Thinking Through Options (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
