CEDAR: Agent‑Orchestrated Tree Search for Goal‑Directed Optimization of Complex Systems

Yingtao TianSubmission to ICLR 2026 · Rejected

TL;DR

CEDAR is a new autonomous framework that combines Large Language Models (LLMs) with Monte Carlo Tree Search (MCTS) to design and optimize complex systems, such as economic or biological models. By representing these systems as Python code and using "Editor" and "Judge" agents to iteratively refine them, CEDAR can achieve sophisticated goals—like balancing global population growth with environmental health—far more effectively than traditional manual modeling or black-box optimization.

What Problem Does This Paper Address?

Modeling complex systems (think: pandemic spread, global economics, or ecological niches) is notoriously difficult. These systems are defined by non-linear feedback loops where a small change in one variable can lead to a catastrophic "tipping point" or an unexpected "boom" elsewhere.

Historically, experts used specialized languages like DYNAMO or STELLA to build these models. However, this process is a manual "fitting" nightmare. The relationship between the model’s structure (the equations) and its emergent behavior (the simulation results) is often opaque. If a model doesn't meet a goal, a human has to guess which equation to tweak, run it, and try again. This labor-intensive workflow limits our ability to use "systems thinking" for rapid decision-making in critical areas like climate policy or resource management.

Methods

The authors introduce CEDAR (Complex-systems Exploration and Design via Agent-Orchestrated Refinement). It moves away from niche languages and represents systems using a restricted subset of Python. This allows LLMs to leverage their existing coding prowess to modify system dynamics directly.

The "brain" of CEDAR is an LLM-driven Monte Carlo Tree Search (MCTS). Here is how the loop works:

  1. Selection: MCTS chooses a promising "node" (a version of the system code) to expand.
  2. Expansion (The LLM Editor): An LLM agent acts as an "Editor," proposing specific code modifications based on a strategy (e.g., "be aggressive" or "be conservative") and the current search history.
  3. Simulation: The new Python code is executed to produce a simulation record.
  4. Evaluation (The LLM Judge): Another LLM agent, the "Judge," looks at the simulation data and the original goal. It provides a numerical score and a textual analysis of why the system performed the way it did.
  5. Backpropagation: The score is used to update the search tree, guiding future iterations toward better designs.

Datasets

The researchers evaluated CEDAR on a diverse set of 20 complex systems:

  • World Dynamics: A famous, high-dimensional model from Jay Forrester (1971) that tracks the co-evolution of human population, capital investment, natural resources, and pollution.
  • Biological Systems: 19 distinct models from the book Modeling Dynamic Biological Systems, covering everything from predator-prey dynamics to nutrient cycling.
  • Baselines: They compared CEDAR against Optuna, a state-of-the-art black-box optimizer, across three levels of "formula support" (giving Optuna no math, simple math, or the full correct equations).

Key Insights and Ideas

  • Optimizing for the "Vague": One of the most impressive feats is CEDAR's ability to handle natural language goals like "Make the population grow sustainably without creating excessive pollution." Traditional optimizers need a rigid mathematical objective function; CEDAR uses the LLM Judge to translate human intent into a guiding signal.
  • Structure over Parameters: While traditional optimizers just tune the "knobs" (coefficients), CEDAR can change the "wiring" (the equations themselves). In experiments, CEDAR (with no prior knowledge of the math) outperformed Optuna even when Optuna was given the full, correct mathematical skeleton.
  • Preserving Diversity: Unlike many optimizers that collapse onto a single "best" answer, the MCTS structure allows CEDAR to discover multiple, structurally distinct ways to solve a problem. This is vital for policy-makers who need to see different trade-offs (e.g., prioritizing resource preservation vs. prioritizing economic growth).
  • Interpretable Optimization: Because the LLM Judge provides a textual rationale for every score, the entire optimization process becomes a "glass box." A human can read the Judge's feedback to understand exactly why a specific model variant failed or succeeded.

Key Results

  • World Dynamics Mastery: CEDAR successfully optimized the complex World Dynamics model, finding a configuration that achieved higher population growth with significantly lower pollution and resource depletion than the original human-designed baseline from 1971.
  • Quantitative Superiority: In "fitting a record" tasks (trying to match a specific data curve), CEDAR achieved an L1 distance of 2.22, compared to Optuna’s 3.71—even though Optuna was "cheating" by having access to the full ground-truth formulas.
  • Efficiency: The system is practical; a typical run takes about 30 minutes and costs between $10 and $50 in API fees, making it a viable tool for researchers.

Limitations and Future Work

The authors acknowledge that while CEDAR works empirically, the theoretical convergence of MCTS when using LLMs as transition kernels is still an open research question.

Future work points toward a "Human-in-the-loop" paradigm, where CEDAR acts as a co-pilot. A human could inspect the diverse solutions CEDAR finds, refine the goals in natural language, and send the agent back in to explore further. There is also interest in extending CEDAR to even more specialized domains like epidemiology or robotics by adding domain-specific "primitives" to its Python library.

Summary updated

Abstract

Complex systems modeling analyzes nonlinear, feedback-driven phenomena from population dynamics to economic policy, supporting decisions with significant societal impact. In established practice, models are often authored in specialized system-dynamics languages (e.g., DYNAMO, STELLA) that specify the models' structure. However, building and refining such models requires extensive manual effort due to (1) the opaque relationship between the structure and emergent behavior and (2) the labor-intensive workflows imposed by these languages. These barriers limit adoption and hinder effective decision-making. To address these challenges, we introduce \method (\methodlong), an autonomous method that uses LLM (Large Language Model) agents to discover and improve complex systems that satisfy user-specified goals. Our key innovation is an LLM-driven MCTS (Monte Carlo Tree Search) process deeply coupled with complex system} at each iteration, an LLM Judge evaluates performance against goals and an LLM Editor proposes improved system variants. We represent systems using a restricted, runnable subset of Python with domain-specific primitives, enabling LLMs to meaningfully and modify system dynamics directly. CEDAR is theoretically designed with formalization in mind, and empirically enables automatic optimization of vague goals, thereby reducing human effort while achieving capabilities beyond existing approaches. The unified design handles diverse systems across domains, constructing complex systems that would otherwise require extensive manual fitting. Moreover, by using LLMs to interpret systems, \method makes system design transparent and accessible, facilitating broader adoption of complex systems modeling.
BibTeX
@misc{tian2026cedar,
  author = {Yingtao Tian},
  title = {{CEDAR: Agent‑Orchestrated Tree Search for Goal‑Directed Optimization of Complex Systems}},
  year = {2026},
  eprint = {2608.06871},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.06871},
}
Figure