Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Verification-Aware Planning for Multi-Agent Systems

Tianyang Xu, Dan Zhang, Kushan Mitra, Estevam Hruschka

TL;DR

The paper addresses the fragility of multi-agent LLM systems arising from misinterpretations and handoffs during planning and execution. It introduces VeriMAP, a verification-aware planning framework that decomposes tasks into a directed acyclic graph (DAG), assigns per-subtask verification functions (VFs) in Python or natural language, and coordinates execution and verification through dedicated modules. Empirical results across five diverse datasets show VeriMAP outperforms both single-agent baselines and MAP-based multi-agent systems, with notable gains on harder tasks and clear benefits from replanning. The work advances robustness, interpretability, and scalability in multi-agent LLM systems without relying on external labels.

Abstract

Large language model (LLM) agents are increasingly deployed to tackle complex tasks, often necessitating collaboration among multiple specialized agents. However, multi-agent collaboration introduces new challenges in planning, coordination, and verification. Execution failures frequently arise not from flawed reasoning alone, but from subtle misalignments in task interpretation, output format, or inter-agent handoffs. To address these challenges, we present VeriMAP, a framework for multi-agent collaboration with verification-aware planning. The VeriMAP planner decomposes tasks, models subtask dependencies, and encodes planner-defined passing criteria as subtask verification functions (VFs) in Python and natural language. We evaluate VeriMAP on diverse datasets, demonstrating that it outperforms both single- and multi-agent baselines while enhancing system robustness and interpretability. Our analysis highlights how verification-aware planning enables reliable coordination and iterative refinement in multi-agent systems, without relying on external labels or annotations.

Verification-Aware Planning for Multi-Agent Systems

TL;DR

The paper addresses the fragility of multi-agent LLM systems arising from misinterpretations and handoffs during planning and execution. It introduces VeriMAP, a verification-aware planning framework that decomposes tasks into a directed acyclic graph (DAG), assigns per-subtask verification functions (VFs) in Python or natural language, and coordinates execution and verification through dedicated modules. Empirical results across five diverse datasets show VeriMAP outperforms both single-agent baselines and MAP-based multi-agent systems, with notable gains on harder tasks and clear benefits from replanning. The work advances robustness, interpretability, and scalability in multi-agent LLM systems without relying on external labels.

Abstract

Large language model (LLM) agents are increasingly deployed to tackle complex tasks, often necessitating collaboration among multiple specialized agents. However, multi-agent collaboration introduces new challenges in planning, coordination, and verification. Execution failures frequently arise not from flawed reasoning alone, but from subtle misalignments in task interpretation, output format, or inter-agent handoffs. To address these challenges, we present VeriMAP, a framework for multi-agent collaboration with verification-aware planning. The VeriMAP planner decomposes tasks, models subtask dependencies, and encodes planner-defined passing criteria as subtask verification functions (VFs) in Python and natural language. We evaluate VeriMAP on diverse datasets, demonstrating that it outperforms both single- and multi-agent baselines while enhancing system robustness and interpretability. Our analysis highlights how verification-aware planning enables reliable coordination and iterative refinement in multi-agent systems, without relying on external labels or annotations.
Paper Structure (73 sections, 10 figures, 9 tables)

This paper contains 73 sections, 10 figures, 9 tables.

Table of Contents

  1. Introduction
  2. VeriMAP
  3. Verification-Aware Planner
  4. Planning for Information Flow
  5. Planner-Defined Subtask Verification
  6. Executor
  7. Verifier
  8. Coordinator
  9. Task Sequencing and Context Management
  10. Execution and Verification Management
  11. Error Handling and Replanning
  12. Evaluation
  13. Setup
  14. Evaluation Datasets
  15. Tools
  16. ...and 58 more sections

Figures (10)

  • Figure 1: Overview of VeriMAP. VeriMAP integrates planning and verification by decomposing tasks into a DAG of subtasks, where the planner specifies Structured and Named I/O and Verification Functions (VFs) at subtask-level. Executors produce JSON outputs verified by paired VFs, while the coordinator manages contexts, retries, and replanning to ensure reliable final results.
  • Figure 2: Total API cost (planner, executor, verifier) across five benchmarks. The y-axis indicates average per-task cost in USD, using gpt-4.1 as the planner and gpt-4o-mini as the executor and verifier (when applicable).
  • Figure 3: Comparison of planning and execution–verification traces between MAP-V and VeriMAP. At step 3, the MAP-V verifier failed to detect the incorrect answer $-2$ (marked with a thumbs-down), whereas VeriMAP, guided by the planner-generated VF, successfully identified the error and used the feedback to guide the executor to the correct answer in the next attempt.
  • Figure 4: Prompt for VeriMAP Planner
  • Figure 5: Replanning Prompt for VeriMAP Planner
  • ...and 5 more figures