Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Autonomous Quantum Simulation through Large Language Model Agents

Weitang Li, Jiajun Ren, Lixue Cheng, Cunxi Gong

TL;DR

This work shows that large language model agents can autonomously perform tensor‑network simulations of quantum many‑body systems by coupling in‑context learning with carefully curated documentation and a multi‑agent workflow. A central Conductor coordinates seven specialized agents, enabling domain‑specific reasoning and rigorous validation throughout the physics–code–data pipeline. Across three benchmark tasks—the 2D transverse‑field Ising model, sub‑Ohmic spin‑boson dynamics, and retinal photoisomerization—the multi‑agent, in‑context approach yields approximately 90% success with publication‑quality outputs, while baseline configurations suffer from errors and hallucinations. The results suggest that documenting domain knowledge inside the agent prompt and decomposing tasks among specialized agents can democratize access to tensor‑network methods, with ongoing work aimed at reducing error rates and extending capabilities to more complex problems and modalities.

Abstract

We demonstrate that large language model (LLM) agents can autonomously perform tensor network simulations of quantum many-body systems, achieving approximately 90% success rate across representative benchmark tasks. Tensor network methods are powerful tools for quantum simulation, but their effective use requires expertise typically acquired through years of graduate training. By combining in-context learning with curated documentation and multi-agent decomposition, we create autonomous AI agents that can be trained in specialized computational domains within minutes. We benchmark three configurations (baseline, single-agent with in-context learning, and multi-agent with in-context learning) on problems spanning quantum phase transitions, open quantum system dynamics, and photochemical reactions. Systematic evaluation using DeepSeek-V3.2, Gemini 2.5 Pro, and Claude Opus 4.5 demonstrates that both in-context learning and multi-agent architecture are essential. Analysis of failure modes reveals characteristic patterns across models, with the multi-agent configuration substantially reducing implementation errors and hallucinations compared to simpler architectures.

Autonomous Quantum Simulation through Large Language Model Agents

TL;DR

This work shows that large language model agents can autonomously perform tensor‑network simulations of quantum many‑body systems by coupling in‑context learning with carefully curated documentation and a multi‑agent workflow. A central Conductor coordinates seven specialized agents, enabling domain‑specific reasoning and rigorous validation throughout the physics–code–data pipeline. Across three benchmark tasks—the 2D transverse‑field Ising model, sub‑Ohmic spin‑boson dynamics, and retinal photoisomerization—the multi‑agent, in‑context approach yields approximately 90% success with publication‑quality outputs, while baseline configurations suffer from errors and hallucinations. The results suggest that documenting domain knowledge inside the agent prompt and decomposing tasks among specialized agents can democratize access to tensor‑network methods, with ongoing work aimed at reducing error rates and extending capabilities to more complex problems and modalities.

Abstract

We demonstrate that large language model (LLM) agents can autonomously perform tensor network simulations of quantum many-body systems, achieving approximately 90% success rate across representative benchmark tasks. Tensor network methods are powerful tools for quantum simulation, but their effective use requires expertise typically acquired through years of graduate training. By combining in-context learning with curated documentation and multi-agent decomposition, we create autonomous AI agents that can be trained in specialized computational domains within minutes. We benchmark three configurations (baseline, single-agent with in-context learning, and multi-agent with in-context learning) on problems spanning quantum phase transitions, open quantum system dynamics, and photochemical reactions. Systematic evaluation using DeepSeek-V3.2, Gemini 2.5 Pro, and Claude Opus 4.5 demonstrates that both in-context learning and multi-agent architecture are essential. Analysis of failure modes reveals characteristic patterns across models, with the multi-agent configuration substantially reducing implementation errors and hallucinations compared to simpler architectures.
Paper Structure (16 sections, 9 equations, 5 figures)

This paper contains 16 sections, 9 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Results
  3. Agent Architecture
  4. Benchmark Tasks
  5. Typical Workflow
  6. Performance Analysis
  7. Discussion
  8. Methods
  9. Matrix Product States and Renormalizer
  10. Agent Implementation and LLM Configuration
  11. Benchmark Tasks
  12. Two-Dimensional Transverse-Field Ising Model
  13. Sub-Ohmic Spin-Boson Model
  14. Retinal Photoisomerization Model
  15. Prompts of the Benchmark Tasks
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Multi-agent architecture for autonomous tensor network simulation. The Conductor agent orchestrates seven specialized agents: Strategist, Guide, Programmer, Executor, Aggregator, Validator, and Visualizer. Agents communicate through structured messages and share access to tools such as file operations and code execution. The user interacts through natural language problem specifications and receives publication-ready outputs.
  • Figure 2: Model schematics and representative simulation results.(a--c) Schematic diagrams of the three benchmark models: (a) 2D transverse-field Ising model with ferromagnetic coupling $J$ and transverse field $h$; (b) spin-boson model showing the two-level system coupled to bath modes and the diabatic potential energy surfaces; (c) retinal photoisomerization model with torsional coordinate connecting cis and trans configurations. (d--f) Successful runs from the multi-agent architecture showing correct phase transition behavior, accurate phase classification, and converged population dynamics. (g--i) Failed runs from the baseline configuration illustrating characteristic failure modes, which are missing data and overlapping curves, incomplete phase diagrams dominated by errors, and unphysical oscillatory dynamics.
  • Figure 3: Complete workflow for spin-boson model phase diagram construction. The workflow shows coordination between specialized agents across four tracks: planning (Strategist, Guide), programming (Programmer), execution (Executor), and analysis (Aggregator, Validator, Visualizer). The right column provides a step-by-step narrative of the 20 actions taken during the simulation. The Validator provides quality scores at each stage, ensuring numerical reliability before production runs and flagging borderline cases in the final phase classification.
  • Figure 4: Performance comparison across agent architectures and LLM backends.(a--c) Score distributions for each model across the three benchmark tasks (Ising, Spin-boson, Retinal). Small markers show individual runs (5 per task), while large markers indicate mean scores. (d) Average score deduction per run decomposed by error type for each model and architecture. Implementation errors and hallucination dominate failures for DeepSeek-V3.2 and Gemini 2.5 Pro, while response errors are the primary failure mode for Claude Opus 4.5.
  • Figure 5: Token consumption and agent behavior analysis.(a) Input and (b) output token counts for each model-task combination. (c) Total tool call counts. (d--f) Distribution of high-level agent calls across the seven specialized agents for the Ising model, spin-boson model, and retinal model, respectively.