💬 NLP

TodoEvolve: Learning to Architect Agent Planning Systems

TodoEvolve introduces a meta-planning paradigm that uses a modular design space called PlanFactory and a novel reinforcement learning objective (IGPO) to autonomously synthesize and dynamically refine task-specific planning architectures that outperform hand-crafted systems.

Original authors: Jiaxi Liu, Yanzuo Jiang, Guibin Zhang, Zihan Zhang, Heng Chang, Zhenfei Yin, Qibing Ren, Junchi Yan

Published 2026-02-10

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiaxi Liu, Yanzuo Jiang, Guibin Zhang, Zihan Zhang, Heng Chang, Zhenfei Yin, Qibing Ren, Junchi Yan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "One-Size-Fits-All" Suit Dilemma

Imagine you are hiring a personal assistant to help you with your life.

Most current AI "agents" are like assistants who only know how to follow one specific type of instruction manual. If you ask them to organize a simple grocery list, they use a linear manual (Step 1, Step 2, Step 3). If you ask them to plan a massive, multi-city wedding, they try to use that same grocery list manual. They get overwhelmed, lose track of details, or try to do everything in a straight line when they should be multitasking.

In the AI world, we call these "fixed planning structures." They are rigid. They don't change whether the task is as simple as "check the weather" or as complex as "write a software program."

The Solution: TodoEvolve (The "Master Architect" Assistant)

The researchers created TodoEvolve. Instead of just giving the AI a manual, they gave the AI the ability to design its own manual every time you give it a new task.

Think of TodoEvolve not as the assistant, but as the Master Architect who sits behind the assistant. Before the assistant starts working, the Architect looks at your request and says: "Wait, this isn't a simple task. For this, we don't need a checklist; we need a complex web of interconnected teams working in parallel."

How It Works: The Four Pillars of Design

To build these custom "manuals," TodoEvolve uses four specific design tools (which they call the PlanFactory):

Topology (The Blueprint): Should the plan be a straight line (like a recipe), a tree (like a family tree), or a complex web (like a spiderweb)?
Initialization (The Setup): How do we write the first draft of this plan?
Adaptation (The Pivot): If something goes wrong (e.g., a website is down), how does the plan change on the fly?
Navigation (The Steering): How does the assistant actually move from one step to the next?

The Secret Sauce: IGPO (The "Efficiency Coach")

How did they teach the AI to be such a good architect? They used a training method called IGPO (Impedance-Guided Preference Optimization).

Imagine you are training a chef. You don't just tell them, "Make a good meal." You tell them, "Make a meal that tastes amazing, but don't waste any ingredients, don't use ten pans if one will do, and don't take three hours if you can do it in thirty minutes."

The researchers rewarded the AI not just for getting the right answer, but for getting it efficiently. They penalized "Cognitive Impedance"—which is a fancy way of saying "wasted brainpower and wasted time." This taught the AI to avoid being "too bureaucratic" (making a plan so complex it becomes a headache) or "too lazy" (making a plan so simple it fails).

Why This Matters (The Results)

When they tested this "Master Architect" approach, the results were huge:

It’s Smarter: It significantly outperformed existing AI assistants on difficult real-world tests.
It’s Versatile: It works whether the "worker" is a small, cheap AI or a massive, expensive one (like GPT-5). It’s like a manager who can lead a group of interns just as effectively as a group of CEOs.
It’s Efficient: It doesn't just get the job done; it finds the "sweet spot" between being thorough and being fast.

Summary in a Nutshell

Old AI: An assistant with one rigid instruction manual for every job.
TodoEvolve: An architect who builds a custom, perfect workflow for every specific problem you throw at it.

Technical Summary: TodoEvolve: Learning to Architect Agent Planning Systems

TodoEvolve is a meta-planning paradigm designed to move beyond the limitations of current LLM-based agent systems, which typically rely on fixed, hand-crafted planning structures (e.g., linear lists, DAGs, or trees). The authors argue that no single planning architecture is optimal for all tasks; instead, the "best" architecture depends on the specific structural characteristics of the problem at hand.

1. The Problem: The "One-Size-Fits-All" Limitation

Current agentic planning systems suffer from structural rigidity. While some tasks benefit from linear execution (to minimize token cost) and others require complex Directed Acyclic Graphs (DAGs) for parallel reasoning, existing frameworks force a specific topology onto every query. This leads to:

Suboptimality: Using a complex graph for a simple task wastes tokens and latency.
Failure in Complexity: Using a simple linear plan for a high-conflict or multi-dependency task leads to reasoning drift or execution failure.
Manual Engineering Overhead: Developers must manually design and tune scaffolds for different domains (web search, coding, embodied AI).

2. Methodology: The Meta-Planning Paradigm

The authors propose a two-pronged approach: building a standardized design space and training a meta-planner to navigate it.

A. PlanFactory (The Modular Design Space)
To enable autonomous synthesis, the authors first created PlanFactory, a unified codebase that decomposes all planning paradigms into four functional dimensions:

♣ Topology: The structural organization (e.g., Linear, DAG, Tree, Hierarchy).
♦ Initialization: How the task is first decomposed into the chosen topology.
♥ Adaptation: The logic for when and how to revise the plan based on environment feedback.
♠ Navigation: The mechanism for issuing directives to the acting agent.

B. Todo-14B (The Meta-Planner)
The core innovation is Todo-14B, a model trained to take a task description as input and output a customized planning configuration (code and YAML). The training follows a two-stage curriculum:

Stage 1: Structural Competence (SFT): The model is fine-tuned on a dataset of verified, successful planning trajectories. This teaches the model the "syntax" of the PlanFactory API and how to generate executable architectures.
Stage 2: Impedance-Guided Preference Optimization (IGPO): This is a novel reinforcement learning objective. Instead of just optimizing for "correctness," IGPO optimizes for efficiency and stability. It uses a metric called Cognitive Impedance ( $I$ ), which penalizes high token costs, high latency, and execution errors. The model learns to prefer architectures that achieve the correct answer with the least "bureaucratic" overhead (the ratio of planning cost to execution cost).

3. Key Contributions

PlanFactory: A standardized, modular toolkit that provides a common interface for diverse planning patterns, serving as both a research benchmark and a data synthesis engine.
TodoEvolve Paradigm: A shift from designing plans to designing the systems that create plans.
IGPO Objective: A multi-objective optimization strategy that treats planning efficiency as a first-class citizen, preventing the model from over-engineering simple tasks.
Todo-14B Model: A specialized meta-planner capable of synthesizing bespoke architectures for arbitrary tasks and agent backbones.

4. Experimental Results

The authors evaluated TodoEvolve across five agentic benchmarks (including GAIA and xBench-DS) using various LLM backbones (GPT-5-mini, DeepSeek V3.2, Kimi K2).

Performance Gains: Integrating TodoEvolve with the Smolagents framework improved GAIA scores by up to 16.37%.
Generalization: The meta-planner is backbone-agnostic; for example, it boosted GPT-5-mini to 75% on xBench-DS.
Efficiency (Pareto Optimality): Unlike many methods that increase performance at the cost of massive compute, TodoEvolve maintains competitive latency and cost by selecting the most "economical" topology for the task.
Complexity Handling: On the most difficult tasks (GAIA Level 3), TodoEvolve outperformed much larger proprietary models by utilizing optimal dynamic topologies.

5. Significance

TodoEvolve represents a significant step toward self-evolving autonomous agents. By decoupling the reasoning engine (the LLM) from the architectural scaffold (the planner), the research demonstrates that agentic intelligence can be significantly enhanced by "meta-reasoning"—the ability to decide not just what to do, but how the system should be structured to do it most effectively. This paves the way for agents that can autonomously adapt their own cognitive workflows to match the complexity of the real world.