Research map / Large language models / Agents, code and math
DyFlow: Dynamic Workflow Framework for Agentic Reasoning
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 DyFlow, a designer-executor framework where a trained planner (DyPlanner, Phi-4 based) dynamically generates subgoal operator graphs conditioned on intermediate feedback, trained by SFT distillation then KTO self-play preference optimization.
Problem
Existing LLM agent workflows are static or only optimized offline per dataset/query, and do not adapt subgoal plans using real-time intermediate feedback, limiting robustness and cross-task generalization.
Approach
A hierarchical designer generates stage subgraphs of context-parameterized operators after each execution step, based on summarized state and feedback; the designer is trained with SFT distillation from GPT-4.1 trajectories followed by KTO on self-play success/failure labels, while the executor (Phi-4) is frozen.
Key results
- DyFlow achieves average 61.45 across five domains, versus 57.74 for the strongest baseline MaAS and 53.44 for Vanilla with Phi-4 executor.
- Zero-shot on held-out SocialMaze, DyFlow scores 17.18 vs 13.36 for MaAS and 6.49 for Vanilla.
- DyPlanner (14B Phi-4 designer) matches GPT-4.1 and Claude-3.7-Sonnet designers (avg 61.45 vs 59.83 and 61.16) at roughly 0.42 USD total inference cost versus 21.26 and 34.61.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DyFlow instead of MaAS | Phi-4 executor · SocialMaze | accuracy: 13.36 -> 17.18 | Helped |
| Use DyFlow instead of AFlow | Phi-4 executor · PubMedQA | accuracy: 69.72 -> 72.91 | Helped |
| Use DyFlow instead of AFlow | Phi-4 executor · MATH | accuracy: 74.00 -> 76.40 | Helped |
| Use DyFlow instead of MaAS | Phi-4 executor · LiveBench | accuracy: 44.00 -> 48.67 | Helped |
| Use DyFlow instead of AFlow | Phi-4 executor · HumanEval | pass@1: 89.02 -> 92.07 | Helped |
| Add KTO preference optimization (ablation) | DyFlow designer · SocialMaze | accuracy: 13.36 -> 17.18 | Helped |
| Add SFT initialization (ablation) | DyFlow designer · SocialMaze | accuracy: 15.65 -> 17.18 | Helped |
| Add dynamic operator selection (ablation) | DyFlow · SocialMaze | accuracy: 12.21 -> 17.18 | Helped |
| Add dynamic planning with intermediate feedback (ablation) | DyFlow · SocialMaze | accuracy: 11.45 -> 17.18 | Helped |
Surprising findings
- Removing SFT (keeping only KTO) slightly improves average over the full system's SFT+KTO variant in some domains, e.g. SocialMaze 15.65 vs 17.18 is lower but HumanEval 91.46 vs 92.07 is close; overall the w/o SFT variant scores 59.00 versus 61.45 for full.
- Removing dynamic planning causes the largest drop (56.48 avg), larger than removing KTO or SFT.
Limitations
- No integration with external tools or APIs such as search or databases.
- Designer trained on only a few hundred to thousands of trajectories and evaluated on small test sets (e.g., 109 HumanEval problems).
- Relies on GPT-4.1 for initial distillation data, adding training cost.
Experimental setup
Datasets: MATH PubMedQA LiveBench SocialMaze HumanEval
Benchmarks: SocialMaze PubMedQA MATH LiveBench HumanEval
Metrics: accuracy pass@1 pass@k cost (USD) and tokens
Baselines: Vanilla CoT Self-Consistency LLM-Debate Self-Refine ADAS AFlow MaAS
Builds on: KTO AFlow MaAS ADAS DyLAN
Closest papers by meaning
- Benchmarking Agentic Workflow Generation (ICLR 2025 · similarity 0.90)
- GraphFlow: A Graph-Based Workflow Management for Efficient LLM-Agent Serving (ICML 2026 · similarity 0.90)
- Tools as Continuous Flow for Evolving Agentic Reasoning (NeurIPS 2026 · similarity 0.90)
- Flow: Modularized Agentic Workflow Automation (ICLR 2025 · similarity 0.88)
- AFlow: Automating Agentic Workflow Generation (ICLR 2025 · similarity 0.88)
- ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.88)
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.