CEDAR: Agent-Orchestrated Tree Search for Goal-Directed Optimization of Complex Systems
Yingtao Tian
cs.AI
2026-08-07
CEDAR (Sakana) wraps an LLM Editor that rewrites system code and an LLM Judge that scores it inside MCTS; with no formulae it beats Optuna even when Optuna is handed the full ground-truth equations.
Complex systems model the world with nonlinear, feedback-driven equations: how population, resources, and pollution push back on each other, how an epidemic spreads, what a policy will provoke. The hard part is that the link between a system's structure (who influences whom, how the feedback loops) and the behavior it ultimately produces is notoriously hard to predict, so designing a system toward a goal is an open problem in artificial life. On top of that, the traditional tooling (specialized languages like DYNAMO and STELLA) is hand-written by experts, with heavy workflows that keep complex-systems methods out of domains where they could help.
CEDAR recasts this as a Monte Carlo Tree Search driven by two LLM agents. The system is represented as a runnable restricted subset of Python with domain primitives, so the model can edit its dynamics directly. Each round, MCTS picks a candidate node to expand; an LLM Editor acts as a variation operator, proposing a rewritten system under one of eight strategies (breakthrough, aggressive, amplify, conservative, and the like); the new system is actually executed; and an LLM Judge acts as a fitness function, scoring the emergent behavior against a natural-language goal and writing a short analysis. This grafts the generate-and-evaluate loop of evolutionary computation onto tree search, with the tree playing a structured population. Because it is a tree rather than convergence to a single point, it keeps multiple solutions alive.
Two comparisons stand out. The first fits a concrete target trajectory (a stochastic population-growth model) against the black-box optimizer Optuna, and the setup deliberately favors the baseline: Optuna gets 100 trials and, in its strongest setting, is handed the full ground-truth formulae.
| Method | Formulae given | L1 lower is better | DTW lower is better |
| Optuna | none | 29.06 | 5503.68 |
| Optuna | simple | 26.01 | 4927.94 |
| Optuna | full truth | 3.71 | 477.52 |
| CEDAR (Claude) | none | 3.29 | 757.21 |
| CEDAR (GPT-5.1) | none | 2.22 | 433.13 |
CEDAR starts from a bare skeleton with no formulae yet lands lower error than Optuna with the full equations. The second task optimizes the classic World Dynamics model toward a natural-language goal (more population, less resource use, less pollution, all at once); CEDAR finds several structurally distinct systems that all meet the goal, with different backends favoring different tradeoffs (Claude leans toward population, GPT-5.1 toward resources). An ablation shows MCTS beats linear search on score and trajectory smoothness and avoids the overfitting-like collapse. A run does about 100 expansions, takes 30 minutes, and costs $10 to $50.
Complex-systems modeling has been mostly expert craft. CEDAR turns "search for a system that meets a fuzzy natural-language goal" into a repeatable pipeline, and because tree search keeps multiple solutions it supports sensitivity analysis. For anyone in system dynamics, artificial life, or policy simulation, this lowers the barrier to building and tuning models. It reads more as lowering the barrier and opening new capability than as squeezing fractions of a point off an existing leaderboard.
The authors flag the main risk themselves: the LLM Judge both scores candidates and shares a model class with the Editor, so in the core abstract-goal setting the evaluator is not anchored to anything external, leaving room for self-congratulation. The paper reports trends rather than tight statistical tests, and the in-depth experiments cover only two systems. The missing comparison is against classical evolutionary and multi-objective optimization; it is still unclear when LLM-based variation and evaluation are necessary rather than merely convenient.