Source-linked AI summary
TodoEvolve: Learning to Architect Agent Planning Systems
Jiaxi Liu, Yanzuo Jiang, Guibin Zhang, Zihan Zhang, Heng Chang, Zhenfei Yin, Qibing Ren, Junchi Yan
TL;DR
Existing agent planning systems rely on fixed structures that do not flexibly match diverse task requirements. TodoEvolve uses PlanFactory and IGPO to synthesize and revise task-specific planning architectures, and it reports consistent gains across benchmarks and backbones while maintaining economical overhead.
Problem
Existing planning systems rely on fixed, hand-crafted structures despite diverse task-specific planning priors.
Method
TodoEvolve trains Todo-14B with IGPO to synthesize and dynamically revise planning systems structured through PlanFactory’s four planning dimensions.
Results
TodoEvolve reports consistent benchmark gains, including a 16.37% improvement over vanilla Smolagents on GAIA with GPT-5-Mini.
Takeaways & Limitations
Task-specific synthesis offers greater adaptability than static graph-based priors across diverse execution backbones and task domains.
Abstract
from arXiv · showhide
Planning has become a central capability for contemporary agent systems in navigating complex, long-horizon tasks, yet existing approaches predominantly rely on fixed, hand-crafted planning structures that lack the flexibility to adapt to the structural diversity of open-ended problems. To address this limitation, we introduce TodoEvolve, a meta-planning paradigm that autonomously synthesizes and dynamically revises task-specific planning architectures. Specifically, we first construct PlanFactory, a modular design space that standardizes diverse planning paradigms within a unified codebase encompassing topology, initialization, adaptation, and navigation, thereby providing a common interface for heterogeneous planning patterns. Leveraging PlanFactory, we collect high-quality planning trajectories and train Todo-14B via \textit{Impedance-Guided Preference Optimization} (IGPO), a multi-objective reinforcement learning objective that encourages the generation of planning systems that are performant, stable, and token-efficient across arbitrary tasks and agent backbones. Empirical evaluations on five agentic benchmarks demonstrate that TodoEvolve consistently surpasses carefully engineered planning modules while maintaining economical API costs and runtime overhead.
1 Introduction
Existing agent planners use diverse, largely fixed structures, motivating task-specific planning adaptation. TodoEvolve addresses this by synthesizing and revising planning systems across four dimensions, with Todo-14B trained through IGPO and evaluated across agentic benchmarks.
- Motivation: Existing planning systems span single- and multi-agent settings and represent plans with diverse structures.These structures include linear to-do lists and more complex graph-based forms.
- Motivation: Different tasks can favor different planning priors, while no single oracle planning system covers all task domains and settings.The passage contrasts MCTS, linear to-do lists, and DAG-based planning across mathematical reasoning, autonomous driving, document QA, and multimodal search.
- TodoEvolve: TodoEvolve synthesizes task-adaptive planners and dynamically updates planning states as execution unfolds.Todo-14B instantiates task-specific planning topology, revision cadence, and navigation strategy.
- PlanFactory: PlanFactory organizes planning systems through topology, initialization, adaptation, and navigation, providing a unified design space for heterogeneous paradigms.It re-implements representative architectures and supports planning-trajectory synthesis and future research.
- Training: IGPO jointly promotes performance, stability, and token efficiency in generated planning systems.Todo-14B is trained as a meta-planner capable of controlling planning structures across scenarios and agent backbones.
- Evaluation: 16.37% improvement on GAIA is reported for TodoEvolve over vanilla Smolagents with GPT-5-Mini, alongside generalization to diverse backbones.The introduction also reports 75% on xBench-DS with GPT-5-Mini.
2 Related Works
Related work shows a progression from structured reasoning and execution loops toward planning systems with domain-specific structures and reinforcement learning objectives. PlanFactory situates these paradigms by distinguishing their agent compatibility, planning scope, and execution structure.
- Agent Planning Systems: Foundational planners such as CoT, ToT, GoT, ReAct, and Reflexion introduced decomposition, structured reasoning, execution loops, and self-correction.These approaches generally use predetermined topologies.
- PlanFactory Representation: Table 1 distinguishes planning paradigms by single- versus multi-agent compatibility, step-wise versus task-wise scope, and linear versus graph execution structure.These dimensions provide the table’s comparison framework.
- Agent Planning Systems: Recent planning frameworks embed domain priors through architectures including dual hierarchies, modular graphs, orchestration, parallel DAGs, consensus, flow graphs, and cross-check networks.The listed systems vary in dispatch, feedback, routing, pruning, deliberation, traversal, and conflict-resolution mechanisms.
- RL for Agent Planning: Reinforcement-learning work applies verifiable rewards to search trajectories, multi-turn interactions, and tool-use strategies.The related methods include Search-R1, LATS, RAGEN, and ToRL.
3 PlanFactory: Unified Planning Codebase
PlanFactory separates agent execution from planning control and represents planning as a modular configuration. Its shared lifecycle interface supports heterogeneous strategies, parallel benchmarking, and evaluation of success and plan coherence.
- System Abstraction: PlanFactory uses a bi-level abstraction in which the Agent System executes interactions and the Planning System governs high-level control logic.The execution substrate includes agents, states, actions, dynamics, context aggregation, and trajectory reward.
- Planning Configuration: The Planning System imposes structural logic on execution through a configuration composed of four functional modules.These modules correspond to topology, initialization, adaptation, and navigation.
- Planning Configuration: Current systems manually fix the planning configuration, motivating automatic synthesis of a task-tailored configuration P*.PlanFactory is designed to decouple high-level planning logic from low-level execution.
- Implementation: All planning paradigms inherit from a standardized BasePlanning interface encapsulating topology, initialization, adaptation, and navigation.This polymorphism allows heterogeneous strategies to be swapped within a shared runtime.
- Inference Workflow: The inference workflow constructs a customized planning system across four dimensions and deploys it in real time to orchestrate agent execution.The workflow connects planning-system construction with runtime execution.
- Evaluation: PlanFactory evaluates dynamic information-seeking tasks using LLM-as-a-Judge to quantify task success and logical plan coherence.The evaluation suite automates trajectory analysis in open domains.
4 TodoEvolve: Training Meta-Planners
TodoEvolve trains a meta-planner to synthesize task-specific planning systems through a standardized PlanFactory space, verified execution data, and two-stage SFT followed by IGPO alignment. IGPO uses execution outcomes and cognitive impedance to favor planning architectures that are correct, efficient, and operationally stable.
- Meta-planning paradigm: TodoEvolve frames planning generation as conditional code generation for bespoke planning systems tailored to task-specific structural requirements.The paradigm aims to move beyond static protocols and manual engineering by generating customized planning configurations.
- PlanFactory: PlanFactory standardizes existing planning mechanisms into a shared code interface and Plan Space for expressing different topological structures.The toolkit decouples high-level planning logic from low-level execution and supports systematic study of agentic architectures.
- Data construction: Evolutionary sampling combines each user query, Meta-Planner instructions, Meta-Tool documentation, and three randomly sampled structural references to generate query-specific plans.The generated plan can integrate or modify reference patterns rather than merely replicate existing templates.
- Data construction: Execution-based verification retains a generated plan only when its final answer matches ground truth, filtering out hallucinated or unsound architectures.Validated trajectories are then formatted into supervision for correctness learning and preference-based refinement.
- Preference alignment: IGPO constructs preference pairs by prioritizing correctness, discarding pairs where both plans fail, and using cognitive impedance to distinguish equally correct plans.Cognitive impedance incorporates total cost, errors, execution smoothness, and a planning-to-execution cost ratio that penalizes excessive planning effort.
- Training objective: The two-stage curriculum uses SFT to establish structural competence and IGPO to align the Meta-Planner with execution efficiency.The optimization objective balances task success against operational costs, while the resulting approach is intended to minimize cognitive impedance without sacrificing functional correctness.
5 Experiments
Experiments evaluate TodoEvolve across diverse benchmarks, backbones, efficiency measures, ablations, and a qualitative execution case. Results show improved accuracy and adaptability over static planning systems without proportional resource increases.
- Main Results: TodoEvolve improves Smolagents by 16.37% on GAIA with GPT-5-Mini, reaching a 72.12% average score.It also surpasses specialized frameworks on GAIA and performs strongly on WebWalkerQA and xBench-DS.
- Main Results: TodoEvolve retains gains across GPT-5-Mini, DeepSeek V3.2, and Kimi K2 execution backbones.With DeepSeek V3.2, it reaches 70.91% on GAIA, exceeding same-backbone Flash-Searcher by over 10 percentage points.
- Structural Specialization: Fixed planning architectures exhibit task-dependent performance variability because their inductive biases favor different task typologies.Centralized systems trade data-handling capacity for reasoning depth, while DAGs prioritize extraction speed over logical coherence.
- Inference Efficiency: TodoEvolve achieves higher success rates than static baselines while maintaining comparable costs and latency.On Kimi-K2 execution metrics, it exceeds the best static baseline by +10.0% on WebWalker-QA and +14.0% on DeepSearch-QA.
- Ablation Study: On GAIA Level 2, Zero-Shot reaches 55.8% accuracy, while complete TodoEvolve peaks at 72.1%.The complete configuration uses both training stages and reference-augmented inference; its additional steps and cost accompany persistence on complex tasks.
- Case Study: A qualitative case study shows TodoEvolve selecting different planning structures as task states evolve.It uses a high-breadth Fork-Join topology for information deadlocks and strict linear constraints for high-precision targets.
6 Conclusion
TodoEvolve replaces rigid, one-size-fits-all planning with autonomous synthesis of task-specific architectures. Across diverse benchmarks, it outperforms static baselines while balancing success rates and computational efficiency.
- Conclusion: TodoEvolve navigates PlanFactory to dynamically configure task-specific topologies and strategies via IGPO.The paper frames this as transforming planning from manual engineering into autonomous synthesis.
- Conclusion: Across diverse benchmarks, TodoEvolve outperforms static baselines while achieving Pareto optimality between success rates and computational efficiency.The paper presents this as a blueprint for self-evolving agents handling open-ended, long-horizon complexities.
A PlanFactory Details
PlanFactory represents planning systems through modular topologies and execution mechanisms. Its instantiated systems span hierarchical, graph-based, parallel, flow-based, and dual-layer designs.
- PlanFactory Details: Co-Sight uses a cross-check net activated by inconsistency detection and navigated through conflict resolution.Its design targets epistemic discrepancies and mutual verification.
- PlanFactory Details: AgentOrchestra uses an orchestration hierarchy with role definition, global decomposition, and centralized routing.A planning agent dispatches instructions to specialized agents.
- PlanFactory Details: OAgents models objectives as dependent modules, initializes them through SOP configuration, and dispatches dependency-satisfied nodes dynamically.Dynamic programming determines which candidate nodes are executable at each step.
- PlanFactory Details: Flash-Searcher builds a parallel DAG and combines aggressive parallelization with workflow pruning and branch injection.Nodes execute when predecessors are satisfied or partial results provide sufficient validation.
- PlanFactory Details: FlowSearch treats reasoning as an evolving thought graph, incrementally constructing and revising nodes and edges toward the goal.Its refiner can add, delete, or modify graph elements while integrating knowledge contexts with execution states.
- PlanFactory Details: OWL separates strategic management from tactical execution through a dual hierarchy and dynamically dispatches tasks using worker profiles.A high-level planner decomposes tasks against available worker capabilities.
B Datasets
The study evaluates five datasets spanning general assistance, web interaction, deep search, synthetic task planning, and long-horizon research. Their sizes and task compositions provide coverage across diverse agentic demands.
- Datasets: GAIA contains 165 tasks: 53 Level-1, 86 Level-2, and 26 Level-3 problems.It is part of the five-dataset evaluation suite.
- Datasets: WebWalkerQA contains 680 real-world queries across four domains and 1,373 webpages, with 170 queries sampled for evaluation.It evaluates complex, multi-turn web interactions.
- Datasets: xBench-DeepSearch contains 100 tasks assessing agentic planning, tool use, and reasoning.TaskCraft contributes 300 synthetic queries, while DeepSearchQA contributes 50 long-horizon research queries.
C Case Study
TodoEvolve generates distinct planning architectures for different query types rather than relying on static templates. The examples span linear sequential logic, state-aware graph routing, and highly parallelized graph structures tailored to task dependencies.
- Three representative systems show TodoEvolve dynamically tailoring control flow to distinct query types and task dependencies.The examples range from linear sequential logic to complex parallel graph structures.
- For strict multi-stage filtering and calculation, TodoEvolve instantiates a linear execution topology with sequential fetch-and-filter processing.A periodic adaptation trigger validates intermediate retrieval results before final synthesis.
- For structured extraction with sorting and ranking constraints, the meta-planner constructs a Knowledge Flow Graph with acquisition, filtering, and finalization nodes.State-aware routing selects parallel extraction or sequential reasoning according to node status.
- For complex entity resolution across multiple subjects, TodoEvolve evolves a highly parallelized graph architecture with concurrent independent sub-goals.A Parallel Executor is activated to minimize latency while the adaptation layer monitors synchronization.