🤖 AI

Coupled Hierarchical Search over Topology and Execution for Agentic Workflow Synthesis

HierFlow is a training-free, test-time architecture that automates agentic workflow synthesis by employing a coupled hierarchical search paradigm, which dynamically integrates topology adjustments with execution-level optimization to achieve superior performance and efficiency across diverse benchmarks.

Original authors: Dong Li, Yanchi Liu, Xujiang Zhao, Wei Cheng, Zhengzhang Chen, Xintao Wu, Zhong Chen, Chen Zhao, Haifeng Chen

Published 2026-07-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dong Li, Yanchi Liu, Xujiang Zhao, Wei Cheng, Zhengzhang Chen, Xintao Wu, Zhong Chen, Chen Zhao, Haifeng Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a super-smart robot how to solve a really hard puzzle, like a complex math problem or a tricky coding challenge. You can't just tell the robot "figure it out" and hope for the best; it needs a plan, a step-by-step recipe of what to do first, second, and third. In the world of artificial intelligence, these plans are called workflows. Think of a workflow like a map for a treasure hunt: it tells the robot where to dig, when to look for clues, and how to use tools to get the treasure.

For a long time, humans had to draw these maps by hand. If the puzzle changed, the map often became useless, and humans had to start over. Recently, scientists figured out how to let the robot draw its own maps. But here's the catch: the number of possible maps is so huge it's like trying to find a single specific grain of sand on every beach on Earth. If the robot tries to check every single map, it takes forever and burns out its brain (or its computer budget). So, the big question for scientists is: How can we teach a robot to quickly find the best map for a specific puzzle without wasting time on bad ones, and without needing to go back to school to learn new things first?

This is where a new idea called HierFlow comes in. The researchers behind this paper, led by Dong Li and colleagues, realized that instead of trying to find the perfect map all at once, you should build it in layers, like a construction crew. They suggest a two-step dance: first, sketch the rough outline of the plan (the "topology"), and then, only if necessary, zoom in to fix the specific details of each step (the "execution").

Here's how their magic trick works. Imagine you are organizing a massive school play. The Top Level is the director. The director doesn't worry about how to sew the costumes or paint the sets; they just decide the big picture: "First, we build the stage. Then, we cast the actors. Finally, we rehearse." This is the Topology Search. The director draws a simple flowchart of these big steps.

But what if the director's plan has a flaw? Maybe the stage can't be built before the actors are cast because the actors need to know their lines to help design the stage. The director might not know this yet. This is where the Lower Level comes in. Think of this as a team of specialized stagehands. When the director says, "Build the stage," the stagehands don't just blindly start hammering. They run a quick, smart test (using a method inspired by a game called Monte Carlo Tree Search, which is like playing out a few possible futures in your head to see which one wins). They try a few different ways to build the stage. If they hit a wall or realize the plan is impossible, they send a signal back to the director: "Hey, this order doesn't work! Let's swap the steps."

The coolest part of HierFlow is its Adaptive Gating. In the past, robots would try to fix every single step of the plan, even if the step was already perfect. It's like a chef tasting every single grain of salt in a soup, even if the soup is already delicious. HierFlow is smarter. It has a "gatekeeper" that asks, "Is this step actually broken or confusing?" If the answer is "No," the gate stays closed, and the robot saves its energy. If the answer is "Yes," the gate opens, and the stagehands zoom in to fix it. This saves a ton of time and computer power.

The paper shows that this method works incredibly well. When they tested HierFlow on tough challenges like answering tricky questions, solving math problems, and writing code, it beat almost every other method they compared it to. It didn't just get better answers; it got them faster and cheaper. The researchers found that by letting the "director" and the "stagehands" talk to each other, the robot could fix its own mistakes on the fly without needing any extra training. It's like giving the robot a self-correcting compass that points it straight to the solution, no matter how twisty the path gets.

In short, HierFlow suggests that we don't need to force robots to memorize every possible solution. Instead, we can teach them to sketch a plan, check if it makes sense, and only dig deep into the details when they actually need to. It's a way to make AI smarter, faster, and more flexible, turning a chaotic search for the perfect plan into a smooth, organized construction project.

Technical Summary: Coupled Hierarchical Search over Topology and Execution for Agentic Workflow Synthesis

Problem Statement

Automating the creation of structured workflows for Large Language Models (LLMs) to solve complex agentic tasks is hindered by a vast, combinatorial search space. Existing approaches face a dichotomy:

Manual Design: Methods like Self-Refine and LLM-Debate rely on handcrafted, static workflows that lack adaptability as task complexity increases.
Automatic Generation: Current automated methods (e.g., AFlow, ADAS) typically operate within a single "proxy space" (e.g., code or natural language programs) using offline search or training-based designers. These approaches suffer from high data/engineering costs, poor generalization to new distributions, and inflexibility due to fixed training distributions.

The core challenge is to perform test-time scaling—adapting workflow generation to specific queries without additional training—while managing the computational cost of searching a complex, discrete workflow space.

Methodology: HierFlow

The authors propose HierFlow, a training-free, test-time hierarchical search framework that treats workflow generation as a coupled topology-and-execution search. Instead of searching a single proxy space, HierFlow operates over two intertwined layers:

1. Dual-Space Search Architecture

Upper-Level (Topology Space): Searches for task decompositions. It constructs a Directed Acyclic Graph (DAG) where nodes represent subtasks. The topology defines the boundaries and dependencies of subtasks, effectively localizing the search space for the lower level.
Lower-Level (Code Execution Space): Performs targeted search within the specific code domain of each subtask defined by the upper level. It uses a Monte Carlo Tree Search (MCTS)-inspired lightweight tree search to explore executable code implementations.
Coupling Mechanism: The search is bidirectional. The topology induces subtask-specific code-search domains, while execution feedback from the lower level can trigger topology refinement (rewinding, adding/removing nodes, or reorganizing dependencies) at the upper level. This creates a closed-loop system where the search object is a coupled state of structure and execution, rather than a fixed pipeline.

2. Adaptive Gating Mechanism

To balance computational cost and quality, HierFlow introduces a gating module that selectively triggers lower-level searches.

Search Necessity Proxy ( $Z_i$ ): For each subtask, the system calculates a proxy based on Search Potential (unexplored space) and Uncertainty (predictive entropy of the evaluation rationale).
Decision Logic: Lower-level search is activated only if $Z_i$ exceeds a threshold $\tau$ . This prevents wasteful computation on subtasks that are either well-defined or unlikely to benefit from further exploration.
Theoretical Basis: The paper provides a coupling-aware analysis (Theorem 2.2 and 2.6), demonstrating that the efficiency of this hierarchical approach depends on the "observable coupling" between subtasks. The method is most efficient when dependencies are sparse and repair rates are low; performance degrades as cross-subtask coupling increases.

3. Algorithmic Components

Initialization: An initiator generates candidate topologies, selecting the best based on a weighted score covering coverage, structural soundness, and searchability.
Refinement: If lower-level execution fails or indicates structural insufficiency (via a refinement signal), a topology refiner (LLM) proposes revised DAGs, preserving completed nodes while modifying dependencies or subtasks.
Lower-Level Search: Uses a soft mixed probability strategy for selection and an LLM-based optimizer for expansion, evaluating candidates via sandboxed execution and LLM-based scoring.

Key Contributions

Coupled Search Formulation: The paper formulates query-level workflow generation as a coupled topology-execution search, where topology is a dynamic search variable that induces code-search domains, and execution feedback revises the topology at test time.
Adaptive Gating: Introduction of a budget-aware gating mechanism that selectively triggers subtask-level code search based on a "search necessity" proxy, optimizing the trade-off between quality and computational cost.
Coupling-Aware Analysis: A theoretical characterization explaining when hierarchical decomposition and gating are beneficial, specifically linking efficiency gains to observable dependency density and execution-triggered repair rates.
Empirical Validation: Comprehensive evaluation across diverse benchmarks (Question Answering, Math Reasoning, Code Generation) demonstrating superior performance without retraining.

Experimental Results

The authors evaluated HierFlow against 11 baselines (including manual methods like CoT and automated methods like AFlow, MaAS, and DyFlow) across five datasets: HotpotQA, GSM8K, MATH, HumanEval, and MBPP.

Performance: HierFlow consistently achieved the highest average performance across all datasets (e.g., 78.74% average score vs. 77.42% for the next best, MaAS). It outperformed both manually designed workflows and other automatic generation methods.
Generalization: Unlike methods relying on offline training or dataset-specific optimization (e.g., MaAS, DyFlow), HierFlow demonstrated robust "plug-and-play" generalization to new data distributions without performance degradation.
Efficiency-Quality Trade-off: By varying the gating threshold $\tau$ , the authors showed a smooth Pareto frontier. A moderate threshold ( $\tau=4$ ) provided a strong balance, achieving high accuracy while significantly reducing API costs and wall-clock time compared to aggressive, un-gated search.
Ablation Studies: Removing the topology space (flat search) or the upper-level refinement significantly degraded performance or increased costs, validating the necessity of the dual-space coupling. Removing the gating mechanism increased costs without proportional accuracy gains, confirming the efficiency of the adaptive trigger.

Significance and Claims

The paper claims that HierFlow addresses the limitations of current agentic workflow synthesis by:

Eliminating Training Overhead: It operates entirely at test time, avoiding the data and engineering costs associated with training workflow designers.
Overcoming Static Limitations: By allowing execution feedback to reshape the topology, it overcomes the rigidity of fixed, handcrafted, or offline-optimized workflows.
Optimizing Search Efficiency: The coupled search and gating mechanisms allow the system to navigate the combinatorial explosion of the workflow space more effectively than single-proxy approaches, providing a scalable solution for complex agentic tasks.

The authors position HierFlow not as a universal asymptotic speedup, but as a method that is particularly effective in regimes where task decomposition can localize search and where cross-subtask coupling is manageable, offering a robust, training-free alternative for high-quality agentic workflow generation.