Skip to content

PDE-Agent: A toolchain-augmented multi-agent framework for PDE solving

πŸ•’ Published (v1): 2025-12-18 06:02 UTC Β· Source: Arxiv Β· link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

πŸ’¬ Ask ChatGPT✦ Ask Claude

TL;DR

PDE-Agent is a toolchain-augmented multi-agent framework that automates PDE solving end-to-end from natural language descriptions by coordinating modular PDE-toolkit tools via LLM-driven agents. It introduces Prog-Act (a dual-loop error correction mechanism with graph memory) and a Resource-Pool for managing implicit runtime artifacts between tools. On the accompanying PDE-Bench (~100 cases), it achieves 90% Pass@1 versus 62% for the next-best baseline.

Problem

Traditional PDE solvers (FEM, FDM) require manual mesh setup and domain expertise. PINN frameworks like DeepXDE partially automate this but still demand expert coding for PDE constraints and network configuration. Existing tool-augmented LLM agents rely on step-by-step validation (causing redundancy) and handle only explicit parameter passing between tools, failing on implicit runtime artifacts (e.g., function objects, class instances) that cannot be serialized β€” a critical gap in multi-tool PDE workflows.

Method

PDE-Agent comprises four agents: Planner (decomposes NL query into ordered subtasks with tool assignments), Parser/Solver (extracts PDE parameters and configures neural network solvers), Executor (invokes DeepXDE-based PDE tools), and Orchestrator (coordinates and corrects).

Two core mechanisms: - Prog-Act (Progressive reasoning and Acting): A dual-loop error-handling scheme. The inner loop runs fixed-checkpoint validation (not step-by-step), with agents proactively requesting help; the outer loop triggers global plan revision when inner-loop correction fails. A directed Graph Memory G=(V,E) tracks tool/subtask dependencies, enabling targeted error localization and retrace along the tool chain. - Resource-Pool: A centralized runtime repository updated after each tool invocation (R ← R βˆͺ {o_{Ο„,f}}), storing non-serializable artifacts (class instances, function objects) accessible to all agents, resolving implicit inter-tool dependencies that explicit parameter passing cannot handle.

Tool-parameter separation keeps tool selection (Planner) distinct from parameter extraction (Parser/Solver) and execution (Executor).

Key Contributions

  • PDE-Agent: first LLM-driven multi-agent framework with modular, encapsulated PDE-toolkits for fully automated NLβ†’solution PDE solving.
  • PDE-Bench (PDE-Data): 100 manually curated test cases across 11 PDE types with full workflow annotations (NL description β†’ tool-collaborative solving).
  • Multi-level evaluation metrics: global (Pass@1), local (STS, BERTScore on tool parameters/outputs), and logical collaboration process (subgraph isomorphism, PageRank-JS similarity, graph edit distance, graph embedding similarity via Node2Vec).
  • Prog-Act dual-loop mechanism and Resource-Pool design, validated via ablation to contribute +14% and essential toolchain completion respectively.

Results

  • PDE-Agent: 90% Pass@1 overall on PDE-Data (100 cases).
  • OctoTools: 62%; OpenAlita: 59%; LLM-DS (pure DeepSeek-v3, no tools): 56%.
  • Ablation: removing Prog-Act drops to 76%; replacing with step-wise validation yields 86% at higher computational cost; Prog-Act achieves 90% with fewer resources.
  • Generalization across LLMs: GPT-4o-mini 81%, Gemini 85%, DeepSeek-v3 90%.
  • On graph-based logical metrics, PDE-Agent outperforms OctoTools across all four dimensions (tool chain completeness, error rate, BERTScore, STS).

Limitations

  • PDE-Data is small (~100 cases), limiting statistical confidence in fine-grained per-category comparisons.
  • Evaluation relies on DeepXDE/PINNs as the underlying solver; accuracy of PDE solutions depends on PINN convergence, which can fail for stiff or high-frequency problems.
  • Resource-Pool design is described but its overhead and memory management strategy under large/long workflows are not analyzed.
  • Generalizability beyond PINNs-solvable PDE classes (e.g., to FEM-requiring problems) is not demonstrated.
  • No wall-clock timing comparison between Prog-Act and step-wise validation is reported, only claims of reduced resource consumption.

Relevance to Harnesses / Meta-Harnesses

PDE-Agent is a concrete domain-specific harness instantiation: it encapsulates a tool suite (PDE-toolkits wrapping DeepXDE), orchestrates multiple specialized sub-agents, and implements a meta-level control loop (Prog-Act's dual-loop with graph memory) that manages tool scheduling, error recovery, and inter-tool dependency resolution. The Resource-Pool pattern β€” a shared runtime state accessible to all agents β€” is directly analogous to the shared context/state management problem central to meta-harness design. The multi-level evaluation framework (global/logical/local) is a reusable harness-evaluation template, separating end-to-end outcome metrics from process-quality metrics, which is a key methodological gap harness researchers are actively addressing.