@beyondwork/docx-react-component 1.0.0 → 1.0.1

package/docs/context7/planner-agent/papers/todoevolve-learning-to-architect-agent-planning-systems.md

@@ -1,1675 +0,0 @@
----
-summary: 'Converted paper text and source links for TodoEvolve: Learning to Architect Agent Planning Systems.'
-read_when:
-  - Reviewing harness and coordination research source material in the docs tree
-  - You want the extracted paper text with source links preserved
-topics:
-  - planning-and-orchestration
-  - harnesses-and-practice
-kind: 'paper'
-title: 'TodoEvolve: Learning to Architect Agent Planning Systems'
----
-# TodoEvolve: Learning to Architect Agent Planning Systems
-
-<Note>
-Converted from the source document on 2026-03-22. The repo does not retain downloaded source files; they were fetched transiently, converted to Markdown, and deleted after extraction.
-</Note>
-
-## Metadata
-
-| Field | Value |
-| --- | --- |
-| Content type | Paper / report |
-| Authors | Jiaxi Liu, Yanzuo Jiang, Guibin Zhang, Zihan Zhang, Heng Chang, Zhenfei Yin, Qibing Ren, Junchi Yan |
-| Year | 2026 |
-| Venue | arXiv 2602.07839 |
-| Research bucket | P0 direct hits |
-| Maps to | Meta-planning, task-specific planning topology, and dynamic planning revision. |
-| Harness fit | Useful when the planning loop itself should adapt instead of staying hand-designed. |
-| Source page | [Open source](https://arxiv.org/abs/2602.07839) |
-| Source PDF | [Open PDF](https://arxiv.org/pdf/2602.07839.pdf) |
-
-## Extracted text
-### Page 1
-
-TodoEvolve: Learning to Architect Agent Planning Systems
-
-TodoRL Team
-
-Abstract
-
-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 plan-
-
-ning patterns. Leveraging PlanFactory, we collect high-quality planning trajectories and train Todo-14B
-
-via 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.
-
-Date: February 10, 2026
-
-Code: https://github.com/EcthelionLiu/TodoEvolve
-
-1 Introduction
-
-With the rapid advancement of foundation models (Team et al., 2025b,a,c), large language model (LLM)-powered
-
-agents have begun to demonstrate strong capabilities across domains such as deep research (Hu et al., 2025a; Shi
-
-et al., 2025b), complex software engineering (iQuest, 2025; Yang et al., 2024), and real-world transactions Andon
-
-(2025); Backlund and Petersson (2025). Beyond improvements in base model capacity, increasingly sophisticated
-
-agent scaffolds are equally critical (Wang et al., 2025a), equipping LLMs with essential agentic support including
-
-planning (Parmar et al., 2025; Wu et al., 2025b; Erdogan et al., 2025a), memory (Hu et al., 2026a), reflection, etc. Among
-
-these, planning stands out as a central capability, enabling agents to navigate complex environments by maintaining a
-
-coherent global state, preserving behavioral consistency, and coordinating actions across tasks (Cao et al., 2025).
-
-Existing planning systems developed for LLM-based agents exhibit substantial diversity. From the perspective of
-
-planning target, some are designed to support single agent, primarily addressing long-horizon execution and mitigating
-
-the risk of “lost in the middle” (Erdogan et al., 2025b), while others are tailored for multi-agent systems, focusing on
-
-subtask allocation and contextual coordination across agents with distinct roles (Parmar et al., 2025; Hu et al., 2025b).
-
-In terms of representational form, plans have been instantiated using a wide range of structures, including linear to-do
-
-lists (LangChain, 2025), directed acyclic graphs (DAG) (Qin et al., 2025), tree-structured plans (Hu et al., 2026b), and
-
-hierarchical notes. Moreover, planning systems differ markedly across task domains, with domain-specific designs
-
-emerging for embodied action (Wang et al., 2024b), web search (Kim et al., 2024), and programming. Faced with this
-
-diversity, practitioners may naturally ask: is there a single planning structure that can serve as a one-size-fits-all solution
-
-that generalizes well across settings?
-
-1
-
-arXiv:2602.07839v1 [cs.CL] 8 Feb 2026
-
-### Page 2
-
-We posit that such an oracle planning system does not exist. Beyond distinct task domains require different planning
-
-priors (for instance, MCTS-based planning may be effective for mathematical reasoning yet is rarely adopted for
-
-autonomous driving agents due to the vastness of its action space (Wang et al., 2024a)), even within a single task
-
-class, alternative planning priors exhibit performance disparities. For example, in web search, AOP (Li et al., 2025a)
-
-employs a simple linear to-do list coupled with a reward model to solve document QA in a token-efficient manner, but
-
-it is substantially outperformed in more complex multimodal settings by DAG-based planning structures (Qin et al.,
-
-2025). Similarly, while linear tasks require minimal revision (Hu et al., 2025b), high-conflict environments demand
-
-continuous topological restructuring (Zhang et al., 2025), rendering a single, universal planning system unrealistic.
-
-Accordingly, we contend that the central challenge is not to design a one-size-fits-all planner, but to customize planning
-
-systems to the structural characteristics of each task. To this end, we propose TodoEvolve, a meta-planning paradigm
-
-that synthesizes task-adaptive agentic planners and dynamically updates their planning states as execution unfolds.
-
-Concretely, we train Todo-14B using Impedance-Guided Preference Optimization (IGPO), a multi-objective preference
-
-learning objective that jointly promotes high performance, stability, and token efficiency in the generated planning
-
-systems. The resulting meta-planner Todo-14B takes a task instance as input and instantiates a tailored planning
-
-topology, revision cadence, and navigation strategy, operationalized as a task-specific to-do structure. Todo-14B
-
-integrates seamlessly with single/multi-agent execution frameworks, remains compatible with diverse LLM backbones,
-
-and generalizes across heterogeneous task domains.
-
-To ground TodoEvolve within the diverse landscape of existing planning systems, we introduce a modular planning
-
-design space comprising four dimensions: ♣ Topology (the structural organization of task decomposition), ♦ Initializa-
-
-tion (how the task topology is instantiated), ♥ Adaptation (when and how the topology is revised), and ♠ Navigation
-
-(the mechanism that issues executable directives to the acting agent). This design space provides a unified abstraction
-
-capable of accommodating and localizing a wide spectrum of existing planning paradigms. Building on this formula-
-
-tion, we decompose and re-implement ten representative planning architectures, including Plan-and-Act (Erdogan
-
-et al., 2025b), linear planning (Hu et al., 2025b), DAG-based planning (Qin et al., 2025), and parallel and dynamic
-
-planning (Zhu et al., 2025). The resulting framework, denoted as PlanFactory, serves both as (i) a data synthesis engine
-
-for generating high-quality planning trajectories to train TodoEvolve and (ii) a standardized codebase to facilitate
-
-future research on agentic planning capabilities. Our contributions are as follows:
-
-❶ Unified Codebase: We introduce PlanFactory, a modular design space for agentic planning systems encompassing
-
-four key components (topology, initialization, adaptation, and navigation), providing unified implementations and
-
-benchmark support for a wide range of prevailing planning structues.
-
-❷ Meta Planners: We introduce TodoEvolve, a meta-planning paradigm that synthesizes task-adaptive planning
-
-systems and dynamically revises planning states. Through impedance-guided preference optimization (IGPO), we
-
-train Todo-14B, a meta-planner capable of instantiating and controlling planning structures across diverse scenarios
-
-and agent backbones.
-
-❸ Experimental Evaluation: Extensive experiments on four challenging agentic benchmarks demonstrate that TodoE-
-
-volve delivers (I) substantial performance gains, improving frameworks such as Smolagents by up to 16.37% on
-
-GAIA; and (II) robust generalization, generalizing across diverse LLM backbones, for example boosting GPT-5-Mini
-
-to 75% on xBench-DS.
-
-2 Related Works
-
-Agent Planning Systems. Agentic planning has evolved from static prompting to structured reasoning. Foundational
-
-works like CoT (Wei et al., 2022), ToT (Yao et al., 2023a), and GoT (Besta et al., 2023) enabled cognitive decomposition,
-
-while ReAct (Yao et al., 2023b) and Reflexion (Shinn et al., 2023) introduced execution loops with self-correction.
-
-However, these approaches typically rely on rigid, predetermined topologies, limiting adaptability in open-ended
-
-environments where optimal structures vary dynamically. Recent frameworks address this by embedding domain
-
-priors: Flash-Searcher (Qin et al., 2025) and OAgents (Zhu et al., 2025) leverage DAG-based parallelism; OWL (Hu et al.,
-
-2025b) and AgentOrchestra (Li et al., 2025a) utilize hierarchical coordination; and systems like FlowSearch (Hu et al.,
-
-2026b), JoyAgent (Han et al., 2025), and Co-Sight (Zhang et al., 2025) optimize workflows via structured verification.
-
-Crucially, these systems remain bound by pre-designed architectures. This necessitates a meta-planning approach
-
-capable of autonomously synthesizing and customizing planning structures tailored to each task’s unique complexity.
-
-2
-
-### Page 3
-
-Table 1 An overview of agentic planning paradigms decomposed in PlanFactory. The “Mul” column distinguishes between
-
-single-agent (S) and multi-agent (M) compatibility. “Scope” specifies the granularity at which planning is performed (α for
-
-step-wise vs. Ω for task-wise), and “Struct” indicates whether the execution flow is linear (ℓ) or organized as a complex graph
-
-structure (G).
-
-Mul. Scope Struct. ♣ Topology ♦ Initialization ♥ Adaptation ♠ Navigation
-
-Method Date
-
-(M/S) (Ω/α) (G/ℓ) Structural Organization Instantiation Mechanism Revision Logic Execution Directives
-
-OWL 2025.6 M Ω G Dual Hierarchy Planner Decompose Manager Intervention Dynamic Dispatch
-
-OAgents 2025.6 M α ℓ Modular Graph SOP Configuration Critic-Loop Feedback Loop Execution
-
-AgentOrchestra 2025.9 M Ω G Orch. Hierarchy Role Definition Env Feedback Centralized Routing
-
-Flash-Searcher 2025.9 S Ω G Parallel DAG Dependency Parsing Workflow Pruning Concurrent Paths
-
-JoyAgent 2025.10 M Ω G Collective Hierarchy Hybrid Planning Consensus Voting Joint Deliberation
-
-FlowSearch 2025.10 M Ω G Thought Graph Flow Construction Dynamic Expansion Graph Traversal
-
-Co-Sight 2025.10 M α ℓ Cross-Check Net Inconsistency Trigger Meta-Verification Conflict Resolution
-
-RL for Agent Planning. Training paradigms have shifted from preference alignment (Rafailov et al., 2023; Schulman
-
-et al., 2017) toward reinforcement learning with verifiable rewards (RLVR) (Guo et al., 2025), optimizing against
-
-objective ground truths fosters emergent self-verification. Recent works apply this to diverse dimensions: Search-
-
-R1 (Jin et al., 2025) and LATS (Zhou et al., 2023) optimize search trajectories; RAGEN (Wang et al., 2025b) targets
-
-multi-turn interactions; and ToRL (Li et al., 2025b) refines tool-use strategies. More related works include (Li et al., 2025c;
-
-Xi et al., 2025; Feng et al., 2024; Paglieri et al., 2025). However, a critical limitation persists: these approaches primarily
-
-optimize the agent’s action policy or tool selection within fixed topological loops. In contrast, our work leverages
-
-verifiable trajectories to train a meta-planner, moving beyond policy optimization to autonomously synthesize the
-
-underlying planning structure itself.
-
-3 PlanFactory: Unified Planning Codebase
-
-3.1 Preliminary
-
-We adopt a bi-level agentic inference abstraction where the Agent System executes environment interactions, while
-
-the Planning System governs high-level control logic.
-
-Agent Systems. We formalize the execution substrate as a tuple M = ⟨I, S, A, Ψ, Ω⟩, comprising an agent roster I, a
-
-global state space S, and a joint action space A = ⋃i∈I Ai. The state dynamics follow Ψ(st+1 ∣ st, at, μ(t)), where
-
-μ(t) ∈ I identifies the active agent at time t. To support action generation, a context mechanism Ω aggregates the
-
-execution history Ht, such that at = πμ(t)(st, Ht, Q ∣ Ω). Finally, the resulting trajectory τ is evaluated by a reward
-
-R(τ), positioning M as a flexible execution engine orchestrated by higher-level logic.
-
-Planning Systems. The Planning System imposes structural logic on execution. We formalize it as a configuration P
-
-comprising four key functional modules:
-
-P = ⟨G, Iinit, Fadapt, Nnav⟩ (1)
-
-defining the mechanisms respectively. As shown in Table 1, existing paradigms represent static instances of P,
-
-augmenting the policy as at = π(⋅ ∣ P). Crucially, current systems rely on manual engineering to fix P, limiting
-
-adaptability. This motivates our meta-level framework, which automatically synthesizes an optimal P ∗ tailored to
-
-each task.
-
-3.2 PlanFactory Codebase
-
-We present PlanFactory, a modular toolkit designed to decouple high-level planning logic from low-level execution,
-
-facilitating the systematic study of agentic architectures.
-
-Implementation. The core of PlanFactory is a standardized lifecycle interface. All planning paradigms (Table 1)
-
-inherit from the BasePlanning abstract class, which encapsulates the four essential components: ♣ Topology,
-
-3
-
-### Page 4
-
-Topology
-
-Structural Organization
-
-Initialization
-
-Instantiation Mechanism
-
-Adaptation
-
-Task Revision Logic
-
-Navigation
-
-Execution Directives
-
-Query
-
-Task Description
-
-PlanFactory
-
-Tools
-
-Context
-
-Concrete
-
-Prompts
-
-Topology
-
-Architecture
-
-DAG
-
-Tree
-
-Others
-
-Linear
-
-Feedback
-
-Alert/Error Dynamic Update
-
-Plan State
-
-Action
-
-Answer
-
-Issue
-
-TodoEvolve Agent Execution Loop
-
-Question: Identify the sequence of key locations
-
-traversed on the route from the Shire to Mordor.
-
-System Prompts: You are an expert AI Architect
-
-for the our Agent Framework. Your goal is to
-
-create a NEW Agent Planning Module in Python
-
-and its corresponding Prompt Configuration
-
-(YAML) based on a specific task description,....
-
-Exam-
-
-ples
-
-Instantiated
-
-Agent
-
-Follow Code & Config
-
-Init Topology
-
-Execute Tool
-
-Update
-
-Sync State
-
-Performance Metrics
-
-Solution
-
-Planning
-
-Execution
-
-Adaptation
-
-Summary
-
-Modular
-
-design
-
-Tools
-
-Info
-
-Meta-Planner
-
-Todo-14B
-
-Meta-designing thinking...
-
-class LoTRPlanner(BasePlanning):
-
-def topology_initialize(self):
-
-# Topology: Linear Chain
-
-# 1. Identify Start & End
-
-# 2. Decompose into Segments
-
-return PlanningStep(plan)
-
-def adaptation(self, step):
-
-# 1. Check current location
-
-# 2. Verify next connection
-
-return SummaryStep(status)
-
-Planning Class
-
-which topo? when to adapt?...
-
-system_prompt:
-
-role: "Middle-earth Cartographer"
-
-goal: "Trace route sequentially"
-
-output: "JSON file"
-
-planning:
-
-strategy: "Linear_Chain_Topology"
-
-instruction: "List key stops from Shire
-
-to Mordor"
-
-step:
-
-action: "Find next location"
-
-Planing System Config
-
-Delivering task-customized system...
-
-Optimize
-
-Input
-
-Customized
-
-Planning System
-
-StepLatencyCost Metrics
-
-...... (more iterations)
-
-Final
-
-Answer
-
-Tool Output
-
-Reward
-
-Spearhead
-
-Feedback
-
-Figure 1 The overall inference workflow of TodoEvolve first constructs a customized planning system along four dimen-
-
-sions—topology, initialization, adaptation, and navigation, and then deploys it in real time to orchestrate agent execution.
-
-♦ Initialization, ♥ Adaptation, and ♠ Navigation. For more details, please refer to Appendix A.. This polymorphism
-
-allows heterogeneous strategies to be swapped seamlessly within a shared runtime. Crucially, this design supports
-
-highly parallelized inference, enabling users to benchmark disparate configurations concurrently on a unified backend
-
-without refactoring the agent loop.
-
-Evaluation. PlanFactory provides a comprehensive evaluation suite tailored for dynamic information-seeking tasks.
-
-To ensure reliable assessment in open domains, we employ an LLM-as-a-Judge mechanism. This automates trajectory
-
-analysis, rigorously quantifying both task success rates and the logical coherence of the generated plans.
-
-4 TodoEvolve: Training Meta-Planners
-
-Current agentic systems predominantly rely on static protocols, which inherently lack the flexibility to address the
-
-diverse distribution of real-world queries. To break the shackles of manual engineering, we propose a Generative
-
-Planning Paradigm. The core of this paradigm is Impedance-Guided Preference Optimization (IGPO), a novel training
-
-strategy designed to endue Todo-14B with the ability to dynamically synthesize bespoke planning systems Pcustom
-
-tailored to unique structural requirements. Unlike standard alignment which focuses on stylistic imitation, IGPO
-
-explicitly optimizes the meta-planner to maximize execution stability while minimizing computational overhead. This
-
-section elaborates on our dual-track methodology: (I) constructing a high-quality verifiable planning dataset, and (II)
-
-employing IGPO to establish robust architectural reasoning.
-
-4.1 Data Construction
-
-To enable generative planning, we formulate the system design as a conditional code generation task. To bridge the
-
-lack of architectural priors in standard LLMs, we propose a Bootstrap-and-Filter pipeline within PlanFactory that
-
-transforms the search for optimal plans into a high-quality supervised dataset. This process involves four stages:
-
-Phase 1: Standardization via Unified Tool Interface. First, we utilize the modular nature of PlanFactory to deconstruct
-
-the functional primitives of existing representative planning systems, specifically the 7 paradigms listed in Table 1.
-
-4
-
-### Page 5
-
-We decompose their discrete mechanisms into standardized tools. These tools are encapsulated within our unified
-
-framework, creating a shared Plan Space where different topological structures can be expressed using a consistent
-
-code interface.
-
-Phase 2: Evolutionary Sampling. With the standardized tools ready, we employ an evolutionary strategy to generate
-
-diverse planning candidates. For each query Qi, we construct a specialized input context Ci consisting of:
-
-• The specific user query Qi.
-
-• The system prompt defining the Meta-Planner’s role.
-
-• Detailed documentation of the available Meta-Tools.
-
-• A randomly sampled subset of 3 static planning samples {P 1
-
-ref, P 2
-
-ref, P 3
-
-ref} from our standardized pool, serving as
-
-structural references to guide the architectural design.
-
-The model is tasked with synthesizing a unique, query-specific plan Pgen by integrating or modifying these patterns
-
-to best suit Qi. This process encourages the model to adapt the structural logic to the specific task requirements,
-
-rather than simply replicating existing templates.
-
-Phase 3: Execution-Based Verification. We validate each synthesized plan Pgen by executing it within the PlanFactory
-
-runtime to generate a trajectory τ and final answer Af inal. We apply a strict Execution-as-Judge filter: Pgen is retained
-
-into the dataset if and only if Af inal matches the ground truth. This mechanism effectively purges hallucinated or
-
-unsound architectures, ensuring the Meta-Planner learns exclusively from successful design patterns.
-
-Phase 4: Preference Construction for SFT and IGPO. Finally, we format the validated execution trajectories into training
-
-supervision. To instill both correctness and efficiency into the Meta-Planner, we employ a dual-track alignment
-
-strategy, that separates fundamental capability learning from preference-based refinement:
-
-SFT Data Construction: During SFT, we adopt a strict outcome-supervised filtering protocol. We iterate through the
-
-generated plan candidates and retain only those pairs (Ci, Pgen) that successfully execute. By grounding the target
-
-plan Pgen on the reference-augmented context Ci, we ensure that the base model learns to synthesize valid, executable
-
-architectures from the provided structural inspirations.
-
-IGPO Data Construction: To further align the model with high-quality planning logic via process supervision, we
-
-construct preference pairs (Pwin, Plose) for IGPO. We process the sampling results in pairs and determine the winner
-
-using a hierarchical criterion:
-
-• Correctness First: Correctness is the prerequisite. If one plan succeeds and the other fails, the successful plan is
-
-strictly preferred (Pwin ≻ Plose).
-
-• Noise Filtering: Pairs where both failed are discarded.
-
-• Efficiency as Tie-Breaker: In “expert scenarios” where both candidates yield correct answers, we introduce a novel
-
-metric, Cognitive Impedance (I), to resolve the tie. We define I as a compound cost function:
-
-I(τ) = Ctot ⋅ exp (λ1Nf ail + λ2(1 − Sstab) + λ3
-
-Cplan
-
-Cexec
-
-) (2)
-
-where Ctot is the total cost, Nf ail counts errors, and Sstab quantifies execution smoothness. Crucially, the ratio of
-
-planning cost (Cplan) to execution cost (Cexec) acts as a bureaucracy penalty, ensuring planning effort does not
-
-outweigh execution.
-
-Formally, this pipeline yields two corpora: DSF T = {(Ci, Pgen) ∣ Correct(Pgen)} for structural competence, and
-
-DIGP O = {(Ci, Pwin, Plose) ∣ Pwin ≻ Plose} for efficiency alignment.
-
-4.2 Todo-14B: Training Meta-Planner
-
-This section details the training methodology for Todo-14B. We optimize the Meta-Planner πθ to synthesize planning
-
-configurations that maximize downstream agent performance. We adopt a two-stage curriculum: SFT establishes
-
-structural competence, followed by IGPO to align the planner with execution efficiency.
-
-5
-
-### Page 6
-
-Table 2 Detailed statistics of the constructed datasets. We operate in a long-context regime, where the input LContext (∼13k
-
-tokens) is a composite sequence comprising the system prompt, tool definitions, retrieved structural examples, and the specific user
-
-query.
-
-Dataset Stage Samples Input (LContext) Reasoning (LCoT) Code (LCode)
-
-Stage 1: SFT 3360 ∼ 13,199 ∼ 423 ∼ 1,642
-
-Stage 2: IGPO 2000 ∼ 13,168 ∼ 497 ∼ 1,636
-
-4.2.1 Stage 1: Structural Competence via SFT
-
-We first instill the fundamental capabilities of code generation and architectural reasoning into the Meta-Planner.
-
-Leveraging DSF T, we treat the verified pairs (C, P gen) as expert demonstrations. We optimize πθ using the standard
-
-next-token prediction objective by minimizing the negative log-likelihood of the target sequence. This supervised
-
-training serves as a crucial warm-start phase, ensuring that the model acquires the necessary syntactic rules and API
-
-constraints. Consequently, it learns to synthesize valid instances of P that are structurally grounded in the context C,
-
-providing a stable initialization for subsequent alignment.
-
-4.2.2 Stage 2: Impedance-Guided Preference Alignment
-
-While SFT ensures syntactic viability, it does not guarantee execution efficiency. The subspace of functionally correct
-
-plans is vast, yet the subset of optimal configurations—those that minimize resource consumption while maximizing
-
-success—is sparse. To transition from static correctness to dynamic optimality, we formulate planning generation as a
-
-meta-level optimization problem.
-
-Let P ∈ P denote an executable plan configuration. The Meta-Planner searches the plan space for an optimal
-
-configuration P ∗ that maximizes the expected return, balancing task success against operational costs:
-
-P ∗
-
-= arg max
-
-P ∈P
-
-Eτ ∼M(P)[R(τ) − λI(τ)] (3)
-
-where R(τ) is the binary success reward and I(τ) represents the cognitive impedance. To solve this, we employ our
-
-IGPO method.
-
-Impedance-Contrastive Rejection Sampling. Unlike standard preference collection which often relies on subjective
-
-human ranking, our framework constructs preference pairs based on objective execution metrics. The data curation
-
-process functions as a rejection sampling mechanism designed to distill efficiency signals from stochastic exploration:
-
-• Exploratory Synthesis: Given a context C, the current policy πθ samples K candidate plans {ϕ1,..., ϕK}, instantiat-
-
-ing varied transition dynamics for the Agent System.
-
-• Execution & Evaluation: The Agent System executes these plans to generate trajectories τi. Each trajectory is
-
-evaluated using the composite impedance metric I(τi), aggregating token consumption, temporal latency, and
-
-runtime errors.
-
-• Contrastive Pair Construction: We construct the preference dataset DIGP O by selecting pairs (ϕwin, ϕlose). To
-
-ensure functional validity, we enforce R(τwin) = 1. A pair is selected only if there exists a significant impedance
-
-gap I(τlose) − I(τwin) > δ, ensuring the optimization is driven by high-confidence efficiency signals.
-
-Implicit Reward Alignment. We posit that the optimal policy π∗ should assign probability mass to a configuration
-
-ϕ inversely proportional to its impedance, subject to a KL-divergence constraint that prevents deviation from the
-
-reference distribution. Defining the implicit reward as r(ϕ) = −E[I(τ)] for successful trajectories, the optimal policy
-
-follows the Boltzmann distribution:
-
-π∗
-
-(ϕ ∣ C) ∝ πref (ϕ ∣ C) ⋅ exp (
-
-1
-
-β
-
-r(ϕ)) (4)
-
-This formulation allows us to bypass training an explicit reward model. Following the DPO derivation, the implicit
-
-reward rθ(ϕ) can be re-parameterized by the log-ratio of the policy likelihoods:
-
-rθ(ϕ) = β log
-
-πθ(ϕ ∣ C)
-
-πref (ϕ ∣ C)
-
-(5)
-
-6
-
-### Page 7
-
-Table 3 Performance of various agent frameworks on the WebWalerQA, xBench-Ds, TaskCraft, and GAIA benchmarks. For each
-
-column, the best and second-best pass@1 scores are highlighted in bold and underlined respectively.
-
-Framework Model Family
-
-WebWalker
-
-QA
-
-xBench
-
--DS
-
-Task
-
-Craft
-
-GAIA
-
-Avg. Level 1 Level 2 Level 3
-
-OWL Workforce pass@3 GPT-4o+o3-mini 57.64 55.0 58.33 60.61 81.14 58.14 26.92
-
-OWL RP pass@3 GPT-4o+o3-mini ---58.18 81.14 54.65 23.08
-
-TapeAgents Claude 3.7 etc. ---55.76 71.70 53.49 30.77
-
-AutoAgent Claude 3.5 etc. ---55.15 71.70 53.40 26.92
-
-Smolagents GPT-4.1 ---55.15 67.92 53.49 34.62
-
-Smolagents GPT-5-mini 58.82 51.0 64.00 55.75 69.81 54.65 30.77
-
-Magnetic-1 OpenAI o1 etc. ---46.06 56.60 46.51 23.08
-
-Cognitive Kernel-Pro Claude-3.7 etc. 60.64 56.0 66.00 60.00 79.25 56.98 30.77
-
-Cognitive Kernel-Pro pass@3 Claude-3.7 etc. ---75.15 84.91 73.26 61.54
-
-OAgents Claude-3.7 etc. 58.23 47.0 -66.67 77.36 66.28 46.15
-
-Agent KB GPT-4.1 60.59 48.0 61.67 61.21 79.25 58.14 34.62
-
-Agent KB pass@2 GPT-4.1 68.82 58.0 72.67 67.27 83.02 67.44 34.62
-
-Agent KB pass@3 GPT-4.1 73.53 68.0 75.33 73.94 84.91 73.26 53.85
-
-Flash-Searcher GPT-5-mini 71.18 69.0 69.67 69.09 79.25 69.77 46.15
-
-Flash-Searcher Kimi K2 52.35 66.0 58.00 52.12 58.49 52.33 34.62
-
-Flash-Searcher DeepSeek V3.2 69.41 68.0 69.33 60.61 79.25 53.49 46.15
-
-TodoEvolve + Smolagents GPT-5-Mini 73.53 75.0 72.67 72.12 81.14 72.09 46.15
-
-TodoEvolve + Smolagents Kimi K2 64.71 71.0 69.33 60.00 73.58 55.81 46.15
-
-TodoEvolve + Smolagents DeepSeek V3.2 70.59 74.0 71.33 70.91 84.91 67.44 53.85
-
-The final IGPO loss function maximizes the margin between efficient and inefficient architectures by minimizing:
-
-LIGP O(θ) = −E(ϕw,ϕl)∼DIGP O [log σ(rθ(ϕw) − rθ(ϕl))] (6)
-
-This approach directly aligns the Meta-Planner with the execution environment, teaching it to architect systems that
-
-minimize cognitive impedance while maintaining functional correctness.
-
-5 Experiments
-
-5.1 Experiment Setup
-
-Training. To equip our model with robust planning capabilities, we construct a high-quality composite dataset sourced
-
-from diverse domains. Our training corpus aggregates samples from TaskCraft (Shi et al., 2025a), MoNaCo (Wolfson
-
-et al., 2026), WebWalkerQA (Wu et al., 2025a), and DeepSearchQA (Google, 2025).The data construction pipeline
-
-leverages a teacher-student paradigm, utilizing Gemini-3-Flash as the expert planner to generate high-level reasoning
-
-traces, and DeepSeek V3.2 as the executor to verify actionable outcomes.The final curated dataset detail is shown in
-
-Table 2. We employ Qwen3-14B (Yang et al., 2025) as our backbone model.
-
-Testing & Baselines. To rigorously evaluate the model’s ability to handle diverse and multimodal queries, we
-
-employ a comprehensive evaluation suite. Our benchmarks include the complete GAIA (Mialon et al., 2023) and
-
-XBench-DS (Chen et al., 2025). Additionally, we construct specific test splits from TaskCraft (Shi et al., 2025a) and
-
-WebWalkerQA (Wu et al., 2025a). Crucially, the test samples from these datasets are distinct and non-overlapping
-
-with the training splits to prevent data leakage. For fair comparison during inference, the underlying LLMs driving
-
-the agents include DeepSeek V3.2 (DeepSeek-AI et al., 2025), Kimi-K2 (Team et al., 2025b), and GPT-5-mini (OpenAI,
-
-2025). We utilize Gemini-3-Flash (Comanici et al., 2025) as the judge model to provide unbiased evaluation of agent
-
-trajectories. To validate efficacy, we benchmark Todo-14B against a wide spectrum of state-of-the-art systems. Please
-
-refer to Table 3 for the detailed list of all baselines compared.
-
-5.2 Main Results
-
-Substantial Performance Enhancement over Baselines. As presented in Table 3, integrating TodoEvolve with the
-
-Smolagents framework yields significant performance gains across all evaluated benchmarks. On the comprehensive
-
-7
-
-### Page 8
-
-Table 4 Comprehensive comparison of execution performance across different agent frameworks. The framework achieving the
-
-highest accuracy on each benchmark is highlighted in bold.
-
-Benchmark Metric Co-Sight FlowSearch Flash-Searcher AgentOrchestra OAgents JoyAgent OWL TodoEvolve
-
-WebWalker-QA
-
-Accuracy (%) 16.67 30.00 60.00 46.67 33.33 63.33 53.33 70.00
-
-Avg Cost ($) 0.0013 0.0053 0.0134 0.0112 0.0236 0.0028 0.0062 0.0167
-
-Avg Time (s) 190.52 94.79 164.78 137.69 150.74 212.83 127.63 216.59
-
-Avg Step 2.1 4.0 5.3 6.5 7.2 4.0 3.8 7.7
-
-DeepSearch-QA
-
-Accuracy (%) 4.00 16.00 22.00 20.00 28.00 28.00 30.00 42.00
-
-Avg Cost ($) 0.0025 0.0109 0.0408 0.0263 0.0454 0.0034 0.0191 0.0495
-
-Avg Time (s) 895.88 351.76 522.36 437.06 519.91 548.70 428.63 875.26
-
-Avg Step 2.8 5.5 10.0 9.9 10.8 4.0 6.9 11.7
-
-GAIA-level2 Text-only
-
-Accuracy (%) 17.14 25.71 25.71 14.29 15.71 30.00 24.29 57.14
-
-Avg Cost ($) 0.0018 0.0069 0.0255 0.0149 0.0317 0.0027 0.0130 0.0282
-
-Avg Time (s) 250.23 159.14 305.67 222.75 292.12 304.38 299.78 323.65
-
-Avg Step 2.6 4.6 8.0 7.7 8.7 4.1 6.2 9.1
-
-GAIA benchmark, our approach using GPT-5-Mini achieves an average score of 72.12%, marking a remarkable absolute
-
-improvement of 16.37% over the vanilla Smolagents baseline. Furthermore, our method outperforms specialized
-
-frameworks operating with the same backbone; for instance, it surpasses Flash-Searcher on GAIA Avg and demonstrates
-
-superior versatility on domain-specific benchmarks like WebWalkerQA and xBench-DS. These results empirically
-
-validate that the autonomous synthesis of task-specific planning architectures offers greater adaptability than static
-
-graph-based priors.
-
-Consistent Gains across Diverse Backbones. The scalability of TodoEvolve is evidenced by its consistent improvements
-
-across diverse execution backbones, including GPT-5-Mini, DeepSeek V3.2 and Kimi K2. Notably, when equipped with
-
-the DeepSeek V3.2, our framework achieves a GAIA average of 70.91%, significantly outperforming the Flash-Searcher
-
-implementation using the same model by over 10 percentage points. This consistency suggests that the meta-planner
-
-acquires transferable architectural reasoning capabilities that function independently of the execution model’s internal
-
-knowledge, effectively acting as a general-purpose performance booster for agentic systems.
-
-Complex Reasoning with Open-Source Frameworks. The advantages of TodoEvolve are particularly pronounced in
-
-high-complexity scenarios requiring long-horizon reasoning. On GAIA Level 3, the most challenging subset, our
-
-framework driven by DeepSeek V3.2 attains a success rate of 53.85%. This performance not only surpasses the standard
-
-Agent KB using the more powerful GPT-4.1 but also matches the performance of Agent KB with pass@3 voting. This
-
-finding highlights a critical insight: with optimal dynamic planning topology, cost-effective open-weights models can
-
-rival or exceed the capabilities of resource-intensive proprietary models in complex problem-solving.
-
-5.3 Structural Specialization
-
-We first investigate the performance variability of fixed planning architectures across diverse task typologies, leveraging
-
-the GPT-5-mini (OpenAI, 2025) to evaluate a multi-category benchmark extracted from TaskCraft (Shi et al., 2025a).
-
-As visualized in Figure 2, distinct planning priors exhibit strong inductive biases suitable for specific domains but
-
-lack universality. For instance, centralized systems trade data-handling capacity for reasoning depth, whereas DAG
-
-topologies prioritize extraction speed over logical coherence. This heterogeneity highlights a critical limitation that
-
-rigid topologies cannot optimally address the structural diversity of open-ended queries. This empirical evidence
-
-validates the core premise of TodoEvolve: by dynamically synthesizing architectures that integrate the complementary
-
-strengths of diverse planning paradigms, our meta-planner achieves cross-domain robustness that no single static
-
-framework can match.
-
-5.4 Inference Efficiency
-
-Beyond task adaptability, we evaluate whether the performance gains of TodoEvolve come at the expense of excessive
-
-computational overhead. Table 4 details the execution metrics on three benchmarks using the Kimi-K2 (Team
-
-et al., 2025b) backbone. TodoEvolve consistently achieves dominant accuracy, surpassing the best static baseline by
-
-substantial margins (e.g., +10.0% on WebWalker-QA, +14.0% on DeepSearch-QA). Crucially, this performance does
-
-not incur a proportional spike in resource consumption, TodoEvolve demonstrates superior Pareto optimality: it
-
-8
-
-### Page 9
-
-Figure 2 Task-Dependent Performance Variability.
-
-Figure 3 Ablation Analysis on GAIA Level 2. We compare the following variants, BS (Base Model), SFT (SFT-Only), ZS (Zero-Shot)
-
-and TodoEvolve.
-
-maintains comparable costs and latency to sophisticated baselines while delivering significantly higher success rates.
-
-This indicates that the meta-planner effectively minimizes cognitive impedance, avoiding the redundant loops of
-
-inefficient planners and the premature failures of overly simple ones.
-
-5.5 Ablation Study
-
-To dissect the efficacy of our training components, we conduct an ablation study on the GAIA Level 2 validation set,
-
-comparing four configurations: (1) Base Model, utilizing the unaligned Qwen3-14B to generate planning systems;
-
-(2) SFT-Only, fine-tuned exclusively on verified planning trajectories; (3) Zero-Shot, which incorporates our IGPO
-
-training but performs inference without few-shot examples; and (4) TodoEvolve, the complete framework employing
-
-both training stages and reference-augmented inference. As illustrated in Figure 3, the Base Model fails to synthesize
-
-executable plans due to a lack of syntactic grounding, a capability established by SFT-Only. Notably, the Zero-Shot
-
-setting not only improves accuracy to 55.8% but also reduces API costs relative to SFT-Only, confirming that IGPO
-
-effectively optimizes execution efficiency. Finally, TodoEvolve achieves a peak accuracy of 72.1%; the concomitant
-
-increase in steps and cost reflects the planner’s enhanced capability to persist through and resolve complex, long-
-
-horizon tasks that simpler variants abandon.
-
-5.6 Case Study
-
-To intuitively illustrate how TodoEvolve facilitates complex reasoning, we present a qualitative analysis of a planning
-
-system synthesized during a real execution. As shown in Figure 4, unlike static, "one-size-fits-all" scaffolds, TodoEvolve
-
-9
-
-### Page 10
-
-Figure 4 Evolved planning architectures in real-world instantiation. The system provides adaptive, state-aware structural
-
-scaffolding that spans from macro-topology initialization to granular adaptation and navigation during the execution stage,
-
-effectively steering the agent toward robust and resilient inference.
-
-delivers a dynamic planning architecture that is adaptively tailored to the evolving task state.
-
-We present a qualitative analysis of the planning system synthesized during real execution, as shown in Figure 4.
-
-The results illustrate that TodoEvolve delivers a dynamic planning architecture that is adaptively tailored to the
-
-evolving task state. Specifically, the planner identifies the optimal computational shape for impedance reduction: it
-
-instantiates a high-breadth Fork-Join topology to break information deadlocks (Task A), while conversely enforcing
-
-strict linear constraints to prune search-space noise for high-precision targets (Task B). Notably, the system exhibits
-
-predictive resilience by anticipating access barriers—such as paywalled reports—and proactively staging fallback paths
-
-to secondary sources. Together, these mechanisms ensure the plan acts as a state-aware anchor, preventing reasoning
-
-drift and transforming passive generation into active, strategic solving.
-
-We present more concrete visualizations of the planning systems designed by Todo-14B in Section C.
-
-6 Conclusion
-
-Traditional agentic planning relies on "one-size-fits-all" workflows, often proving rigid and suboptimal for diverse task
-
-demands. This paper aims to transform planning from manual engineering into an autonomous synthesis process,
-
-making architectural design as adaptive as the underlying model’s reasoning. To this end, we introduce TodoEvolve, a
-
-meta-planning paradigm that navigates a unified design space, PlanFactory, to dynamically configure task-specific
-
-topologies and strategies via IGPO. Our extensive evaluations across diverse benchmarks demonstrate that TodoEvolve
-
-outperforms static baselines, achieving Pareto optimality between success rates and computational efficiency. By
-
-bridging the gap between internal reasoning and external architectural scaffolding, TodoEvolve provides a blueprint
-
-for self-evolving agents capable of mastering open-ended, long-horizon complexities.
-
-10
-
-### Page 11
-
-Contributions
-
-Core Contributors
-
-• Jiaxi Liu
-
-• Yanzuo Jiang
-
-Project Lead
-
-• Guibin Zhang
-
-Contributors
-
-• Zihan Zhang
-
-• Heng Chang
-
-Corresponding Authors
-
-• Zhenfei Yin
-
-• Qibing Ren
-
-• Junchi Yan
-
-11
-
-### Page 12
-
-References
-
-Andon (2025). Vending-Bench 2 | Andon Labs — andonlabs.com. https://andonlabs.com/evals/
-
-vending-bench-2. [Accessed 15-01-2026].
-
-Backlund, A. and Petersson, L. (2025). Vending-bench: A benchmark for long-term coherence of autonomous agents.
-
-Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Gianinazzi, L., Gajda, J., Lehmann, T., Podstawski, M., Niewiadomski,
-
-H., Nyczyk, P., and Hoefler, T. (2023). Graph of thoughts: Solving elaborate problems with large language models.
-
-Cao, P., Men, T., Liu, W., Zhang, J., Li, X., Lin, X., Sui, D., Cao, Y., Liu, K., and Zhao, J. (2025). Large language models
-
-for planning: A comprehensive and systematic survey.
-
-Chen, K., Ren, Y., Liu, Y., Hu, X., Tian, H., Xie, T., Liu, F., Zhang, H., Liu, H., Gong, Y., Sun, C., Hou, H., Yang, H., Pan,
-
-J., Lou, J., Mao, J., Liu, J., Li, J., Liu, K., Liu, K., Wang, R., Li, R., Niu, T., Zhang, W., Yan, W., Wang, X., Zhang, Y.,
-
-Hung, Y.-H., Jiang, Y., Liu, Z., Yin, Z., Ma, Z., and Mo, Z. (2025). xbench: Tracking agents productivity scaling with
-
-profession-aligned real-world evaluations.
-
-Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D.,
-
-Rosen, E., et al. (2025). Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and
-
-next generation agentic capabilities. arXiv preprint arXiv:2507.06261.
-
-DeepSeek-AI, Liu, A., Mei, A., Lin, B., Xue, B., Wang, B., Xu, B., Wu, B., Zhang, B., Lin, C., Dong, C., Lu, C., Zhao, C.,
-
-Deng, C., Xu, C., Ruan, C., Dai, D., Guo, D., Yang, D., Chen, D., Li, E., Zhou, F., Lin, F., Dai, F., Hao, G., Chen, G., Li,
-
-G., Zhang, H., Xu, H., Li, H., Liang, H., Wei, H., Zhang, H., Luo, H., Ji, H., Ding, H., Tang, H., Cao, H., Gao, H., Qu,
-
-H., Zeng, H., Huang, J., Li, J., Xu, J., Hu, J., Chen, J., Xiang, J., Yuan, J., Cheng, J., Zhu, J., Ran, J., Jiang, J., Qiu, J., Li, J.,
-
-Song, J., Dong, K., Gao, K., Guan, K., Huang, K., Zhou, K., Huang, K., Yu, K., Wang, L., Zhang, L., Wang, L., Zhao, L.,
-
-Yin, L., Guo, L., Luo, L., Ma, L., Wang, L., Zhang, L., Di, M. S., Xu, M. Y., Zhang, M., Zhang, M., Tang, M., Zhou, M.,
-
-Huang, P., Cong, P., Wang, P., Wang, Q., Zhu, Q., Li, Q., Chen, Q., Du, Q., Xu, R., Ge, R., Zhang, R., Pan, R., Wang, R.,
-
-Yin, R., Xu, R., Shen, R., Zhang, R., Liu, S. H., Lu, S., Zhou, S., Chen, S., Cai, S., Chen, S., Hu, S., Liu, S., Hu, S., Ma, S.,
-
-Wang, S., Yu, S., Zhou, S., Pan, S., Zhou, S., Ni, T., Yun, T., Pei, T., Ye, T., Yue, T., Zeng, W., Liu, W., Liang, W., Pang,
-
-W., Luo, W., Gao, W., Zhang, W., Gao, X., Wang, X., Bi, X., Liu, X., Wang, X., Chen, X., Zhang, X., Nie, X., Cheng, X.,
-
-Liu, X., Xie, X., Liu, X., Yu, X., Li, X., Yang, X., Li, X., Chen, X., Su, X., Pan, X., Lin, X., Fu, X., Wang, Y. Q., Zhang, Y.,
-
-Xu, Y., Ma, Y., Li, Y., Li, Y., Zhao, Y., Sun, Y., Wang, Y., Qian, Y., Yu, Y., Zhang, Y., Ding, Y., Shi, Y., Xiong, Y., He, Y.,
-
-Zhou, Y., Zhong, Y., Piao, Y., Wang, Y., Chen, Y., Tan, Y., Wei, Y., Ma, Y., Liu, Y., Yang, Y., Guo, Y., Wu, Y., Wu, Y.,
-
-Cheng, Y., Ou, Y., Xu, Y., Wang, Y., Gong, Y., Wu, Y., Zou, Y., Li, Y., Xiong, Y., Luo, Y., You, Y., Liu, Y., Zhou, Y., Wu,
-
-Z. F., Ren, Z. Z., Zhao, Z., Ren, Z., Sha, Z., Fu, Z., Xu, Z., Xie, Z., Zhang, Z., Hao, Z., Gou, Z., Ma, Z., Yan, Z., Shao, Z.,
-
-Huang, Z., Wu, Z., Li, Z., Zhang, Z., Xu, Z., Wang, Z., Gu, Z., Zhu, Z., Li, Z., Zhang, Z., Xie, Z., Gao, Z., Pan, Z., Yao,
-
-Z., Feng, B., Li, H., Cai, J. L., Ni, J., Xu, L., Li, M., Tian, N., Chen, R. J., Jin, R. L., Li, S. S., Zhou, S., Sun, T., Li, X. Q.,
-
-Jin, X., Shen, X., Chen, X., Song, X., Zhou, X., Zhu, Y. X., Huang, Y., Li, Y., Zheng, Y., Zhu, Y., Ma, Y., Huang, Z., Xu,
-
-Z., Zhang, Z., Ji, D., Liang, J., Guo, J., Chen, J., Xia, L., Wang, M., Li, M., Zhang, P., Chen, R., Sun, S., Wu, S., Ye, S.,
-
-Wang, T., Xiao, W. L., An, W., Wang, X., Sun, X., Wang, X., Tang, Y., Zha, Y., Zhang, Z., Ju, Z., Zhang, Z., and Qu, Z.
-
-(2025). Deepseek-v3.2: Pushing the frontier of open large language models.
-
-Erdogan, L. E., Lee, N., Kim, S., Moon, S., Furuta, H., Anumanchipalli, G., Keutzer, K., and Gholami, A. (2025a).
-
-Plan-and-act: Improving planning of agents for long-horizon tasks.
-
-Erdogan, L. E., Lee, N., Kim, S., Moon, S., Furuta, H., Anumanchipalli, G., Keutzer, K., and Gholami, A. (2025b).
-
-Plan-and-act: Improving planning of agents for long-horizon tasks. arXiv preprint arXiv:2503.09572.
-
-Feng, P., He, Y., Huang, G., Lin, Y., Zhang, H., Zhang, Y., and Li, H. (2024). Agile: A novel reinforcement learning
-
-framework of llm agents.
-
-Google (2025). DeepSearchQA — kaggle.com. https://www.kaggle.com/datasets/deepmind/
-
-deepsearchqa. [Accessed 05-01-2026].
-
-Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. (2025). Deepseek-r1:
-
-Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
-
-12
-
-### Page 13
-
-Han, A., Hu, J., Wei, P., Zhang, Z., Guo, Y., Lu, J., and Zhang, Z. (2025). Joyagents-r1: Joint evolution dynamics for
-
-versatile multi-llm agents with reinforcement learning. arXiv preprint arXiv:2506.19846.
-
-Hu, C., Du, H., Wang, H., Lin, L., Chen, M., Liu, P., Miao, R., Yue, T., You, W., Ji, W., Yuan, W., Deng, W., Yuan, X.,
-
-Zhang, X., Liu, X., Liu, X., Xu, Y., Cao, Y., Zhang, Y., Wang, Y., Shu, Y., Zhang, Y., Zhang, Y., Gong, Z., Chang, Z., Li,
-
-B., Ma, D., Jia, F., Wang, H., Liu, J., Bai, J., Liu, J., Liu, M., Wang, N., Wu, Q., Du, Q., Li, S., Sun, W., Gong, Y., Chen, Y.,
-
-Zhao, Y., Lin, Y., Ren, Z., Wang, Z., Zhang, A., Li, B., Ma, B., An, K., Xie, L., Li, M., Li, P., Yang, S., Chen, X., Liu, X.,
-
-Luo, Y., Song, Y., Ding, Y., Liang, Y., Li, Z., Zhang, Z., Zhang, Z., Jiao, B., Jiang, D., Chen, J., Li, J., Zhang, X., and Zhu,
-
-Y. (2025a). Step-deepresearch technical report.
-
-Hu, M., Zhou, Y., Fan, W., Nie, Y., Xia, B., Sun, T., Ye, Z., Jin, Z., Li, Y., Chen, Q., Zhang, Z., Wang, Y., Ye, Q., Ghanem, B.,
-
-Luo, P., and Li, G. (2025b). Owl: Optimized workforce learning for general multi-agent assistance in real-world task
-
-automation.
-
-Hu, Y., Liu, S., Yue, Y., Zhang, G., Liu, B., Zhu, F., Lin, J., Guo, H., Dou, S., Xi, Z., Jin, S., Tan, J., Yin, Y., Liu, J., Zhang,
-
-Z., Sun, Z., Zhu, Y., Sun, H., Peng, B., Cheng, Z., Fan, X., Guo, J., Yu, X., Zhou, Z., Hu, Z., Huo, J., Wang, J., Niu, Y.,
-
-Wang, Y., Yin, Z., Hu, X., Liao, Y., Li, Q., Wang, K., Zhou, W., Liu, Y., Cheng, D., Zhang, Q., Gui, T., Pan, S., Zhang, Y.,
-
-Torr, P., Dou, Z., Wen, J.-R., Huang, X., Jiang, Y.-G., and Yan, S. (2026a). Memory in the age of ai agents.
-
-Hu, Y., Ma, R., Fan, Y., Shi, J., Cao, Z., Zhou, Y., Yuan, J., Zhang, S., Feng, S., Yan, X., Zhang, S., Zhang, W., Bai, L., and
-
-Zhang, B. (2026b). Flowsearch: Advancing deep research with dynamic structured knowledge flow.
-
-iQuest (2025). IQuest Coder — iquestlab.github.io. https://iquestlab.github.io/. [Accessed 15-01-2026].
-
-Jin, B., Zeng, H., Yue, Z., Yoon, J., Arik, S., Wang, D., Zamani, H., and Han, J. (2025). Search-r1: Training llms to reason
-
-and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516.
-
-Kim, M., Bursztyn, V., Koh, E., Guo, S., and Hwang, S.-w. (2024). RaDA: Retrieval-augmented web agent planning
-
-with LLMs. In Ku, L.-W., Martins, A., and Srikumar, V., editors, Findings of the Association for Computational
-
-Linguistics: ACL 2024, pages 13511–13525, Bangkok, Thailand. Association for Computational Linguistics.
-
-LangChain (2025). GitHub - langchain-ai/deepagents: Deep Agents is an agent harness built on langchain and
-
-langgraph. Deep Agents are equipped with a planning tool, a filesystem backend, and the ability to spawn sub-
-
-agents - making them well-equipped to handle complex agentic tasks. — github.com. https://github.com/
-
-langchain-ai/deepagents. [Accessed 15-01-2026].
-
-Li, A., Xie, Y., Li, S., Tsung, F., Ding, B., and Li, Y. (2025a). Agent-oriented planning in multi-agent systems.
-
-Li, X., Zou, H., and Liu, P. (2025b). Torl: Scaling tool-integrated rl. arXiv preprint arXiv:2503.23383.
-
-Li, Z., Hu, Y., and Wang, W. (2025c). Encouraging good processes without the need for good answers: Reinforcement
-
-learning for llm agent planning.
-
-Mialon, G., Fourrier, C., Wolf, T., LeCun, Y., and Scialom, T. (2023). Gaia: a benchmark for general ai assistants. In The
-
-Twelfth International Conference on Learning Representations.
-
-OpenAI (2025). Introducing GPT-5.2 — openai.com. https://openai.com/index/
-
-introducing-gpt-5-2/. [Accessed 08-01-2026].
-
-Paglieri, D., Cupiał, B., Cook, J., Piterbarg, U., Tuyls, J., Grefenstette, E., Foerster, J. N., Parker-Holder, J., and
-
-Rocktäschel, T. (2025). Learning when to plan: Efficiently allocating test-time compute for llm agents. arXiv
-
-preprint arXiv:2509.03581.
-
-Parmar, M., Liu, X., Goyal, P., Chen, Y., Le, L., Mishra, S., Mobahi, H., Gu, J., Wang, Z., Nakhost, H., Baral, C., Lee,
-
-C.-Y., Pfister, T., and Palangi, H. (2025). Plangen: A multi-agent framework for generating planning and reasoning
-
-trajectories for complex problem solving.
-
-Qin, T., Chen, Q., Wang, S., Xing, H., Zhu, K., Zhu, H., Shi, D., Liu, X., Zhang, G., Liu, J., Jiang, Y. E., Gao, X., and Zhou,
-
-W. (2025). Flash-searcher: Fast and effective web agents via dag-based parallel execution.
-
-Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. (2023). Direct preference optimization: Your
-
-language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728–53741.
-
-13
-
-### Page 14
-
-Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. (2017). Proximal policy optimization algorithms.
-
-arXiv preprint arXiv:1707.06347.
-
-Shi, D., Cao, J., Chen, Q., Sun, W., Li, W., Lu, H., Dong, F., Qin, T., Zhu, K., Liu, M., Yang, J., Zhang, G., Liu, J., Zhang, C.,
-
-Wang, J., Jiang, Y. E., and Zhou, W. (2025a). Taskcraft: Automated generation of agentic tasks.
-
-Shi, Z., Chen, Y., Li, H., Sun, W., Ni, S., Lyu, Y., Fan, R.-Z., Jin, B., Weng, Y., Zhu, M., Xie, Q., Guo, X., Yang, Q., Wu, J.,
-
-Zhao, J., Tang, X., Ma, X., Wang, C., Mao, J., Ai, Q., Huang, J.-T., Wang, W., Zhang, Y., Yang, Y., Tu, Z., and Ren, Z.
-
-(2025b). Deep research: A systematic survey.
-
-Shinn, N., Labash, B., and Gopinath, A. (2023). Reflexion: an autonomous agent with dynamic memory and self-
-
-reflection. arXiv preprint, abs/2303.11366.
-
-Team,., Zeng, A., Lv, X., Zheng, Q., Hou, Z., Chen, B., Xie, C., Wang, C., Yin, D., Zeng, H., Zhang, J., Wang, K., Zhong,
-
-L., Liu, M., Lu, R., Cao, S., Zhang, X., Huang, X., Wei, Y., Cheng, Y., An, Y., Niu, Y., Wen, Y., Bai, Y., Du, Z., Wang, Z.,
-
-Zhu, Z., Zhang, B., Wen, B., Wu, B., Xu, B., Huang, C., Zhao, C., Cai, C., Yu, C., Li, C., Ge, C., Huang, C., Zhang, C.,
-
-Xu, C., Zhu, C., Li, C., Yin, C., Lin, D., Yang, D., Jiang, D., Ai, D., Zhu, E., Wang, F., Pan, G., Wang, G., Sun, H., Li, H.,
-
-Li, H., Hu, H., Zhang, H., Peng, H., Tai, H., Zhang, H., Wang, H., Yang, H., Liu, H., Zhao, H., Liu, H., Yan, H., Liu, H.,
-
-Chen, H., Li, J., Zhao, J., Ren, J., Jiao, J., Zhao, J., Yan, J., Wang, J., Gui, J., Zhao, J., Liu, J., Li, J., Li, J., Lu, J., Wang,
-
-J., Yuan, J., Li, J., Du, J., Du, J., Liu, J., Zhi, J., Gao, J., Wang, K., Yang, L., Xu, L., Fan, L., Wu, L., Ding, L., Wang, L.,
-
-Zhang, M., Li, M., Xu, M., Zhao, M., Zhai, M., Du, P., Dong, Q., Lei, S., Tu, S., Yang, S., Lu, S., Li, S., Li, S., Shuang-Li,
-
-Yang, S., Yi, S., Yu, T., Tian, W., Wang, W., Yu, W., Tam, W. L., Liang, W., Liu, W., Wang, X., Jia, X., Gu, X., Ling, X.,
-
-Wang, X., Fan, X., Pan, X., Zhang, X., Zhang, X., Fu, X., Zhang, X., Xu, Y., Wu, Y., Lu, Y., Wang, Y., Zhou, Y., Pan, Y.,
-
-Zhang, Y., Wang, Y., Li, Y., Su, Y., Geng, Y., Zhu, Y., Yang, Y., Li, Y., Wu, Y., Li, Y., Liu, Y., Wang, Y., Li, Y., Zhang, Y.,
-
-Liu, Z., Yang, Z., Zhou, Z., Qiao, Z., Feng, Z., Liu, Z., Zhang, Z., Wang, Z., Yao, Z., Wang, Z., Liu, Z., Chai, Z., Li, Z.,
-
-Zhao, Z., Chen, W., Zhai, J., Xu, B., Huang, M., Wang, H., Li, J., Dong, Y., and Tang, J. (2025a). Glm-4.5: Agentic,
-
-reasoning, and coding (arc) foundation models.
-
-Team, K., Bai, Y., Bao, Y., Chen, G., Chen, J., Chen, N., Chen, R., Chen, Y., Chen, Y., Chen, Y., Chen, Z., Cui, J., Ding, H.,
-
-Dong, M., Du, A., Du, C., Du, D., Du, Y., Fan, Y., Feng, Y., Fu, K., Gao, B., Gao, H., Gao, P., Gao, T., Gu, X., Guan, L.,
-
-Guo, H., Guo, J., Hu, H., Hao, X., He, T., He, W., He, W., Hong, C., Hu, Y., Hu, Z., Huang, W., Huang, Z., Huang, Z.,
-
-Jiang, T., Jiang, Z., Jin, X., Kang, Y., Lai, G., Li, C., Li, F., Li, H., Li, M., Li, W., Li, Y., Li, Y., Li, Z., Li, Z., Lin, H., Lin, X.,
-
-Lin, Z., Liu, C., Liu, C., Liu, H., Liu, J., Liu, J., Liu, L., Liu, S., Liu, T. Y., Liu, T., Liu, W., Liu, Y., Liu, Y., Liu, Y., Liu, Y.,
-
-Liu, Z., Lu, E., Lu, L., Ma, S., Ma, X., Ma, Y., Mao, S., Mei, J., Men, X., Miao, Y., Pan, S., Peng, Y., Qin, R., Qu, B., Shang,
-
-Z., Shi, L., Shi, S., Song, F., Su, J., Su, Z., Sun, X., Sung, F., Tang, H., Tao, J., Teng, Q., Wang, C., Wang, D., Wang, F.,
-
-Wang, H., Wang, J., Wang, J., Wang, J., Wang, S., Wang, S., Wang, Y., Wang, Y., Wang, Y., Wang, Y., Wang, Y., Wang,
-
-Z., Wang, Z., Wang, Z., Wei, C., Wei, Q., Wu, W., Wu, X., Wu, Y., Xiao, C., Xie, X., Xiong, W., Xu, B., Xu, J., Xu, J., Xu,
-
-L. H., Xu, L., Xu, S., Xu, W., Xu, X., Xu, Y., Xu, Z., Yan, J., Yan, Y., Yang, X., Yang, Y., Yang, Z., Yang, Z., Yang, Z., Yao,
-
-H., Yao, X., Ye, W., Ye, Z., Yin, B., Yu, L., Yuan, E., Yuan, H., Yuan, M., Zhan, H., Zhang, D., Zhang, H., Zhang, W.,
-
-Zhang, X., Zhang, Y., Zhang, Y., Zhang, Y., Zhang, Y., Zhang, Y., Zhang, Y., Zhang, Z., Zhao, H., Zhao, Y., Zheng, H.,
-
-Zheng, S., Zhou, J., Zhou, X., Zhou, Z., Zhu, Z., Zhuang, W., and Zu, X. (2025b). Kimi k2: Open agentic intelligence.
-
-Team, T. D., Li, B., Zhang, B., Zhang, D., Huang, F., Li, G., Chen, G., Yin, H., Wu, J., Zhou, J., Li, K., Su, L., Ou, L., Zhang,
-
-L., Xie, P., Ye, R., Yin, W., Yu, X., Wang, X., Wu, X., Chen, X., Zhao, Y., Zhang, Z., Tao, Z., Zhang, Z., Qiao, Z., Wang,
-
-C., Yu, D., Fu, G., Shen, H., Yang, J., Lin, J., Zhang, J., Zeng, K., Yang, L., Yin, H., Song, M., Yan, M., Liao, M., Xia, P.,
-
-Xiao, Q., Min, R., Ding, R., Fang, R., Chen, S., Huang, S., Wang, S., Cai, S., Shen, W., Wang, X., Guan, X., Geng, X.,
-
-Shi, Y., Wu, Y., Chen, Z., Li, Z., and Jiang, Y. (2025c). Tongyi deepresearch technical report.
-
-Wang, C., Deng, Y., Lyu, Z., Zeng, L., He, J., Yan, S., and An, B. (2024a). Q*: Improving multi-step reasoning for llms
-
-with deliberative planning.
-
-Wang, X., Li, B., Song, Y., Xu, F. F., Tang, X., Zhuge, M., Pan, J., Song, Y., Li, B., Singh, J., Tran, H. H., Li, F., Ma, R.,
-
-Zheng, M., Qian, B., Shao, Y., Muennighoff, N., Zhang, Y., Hui, B., Lin, J., Brennan, R., Peng, H., Ji, H., and Neubig, G.
-
-(2025a). Openhands: An open platform for ai software developers as generalist agents.
-
-Wang, Z., Cai, S., Chen, G., Liu, A., Ma, X., and Liang, Y. (2024b). Describe, explain, plan and select: Interactive
-
-planning with large language models enables open-world multi-task agents.
-
-14
-
-### Page 15
-
-Wang, Z., Wang, K., Wang, Q., Zhang, P., Li, L., Yang, Z., Jin, X., Yu, K., Nguyen, M. N., Liu, L., et al. (2025b). Ragen:
-
-Understanding self-evolution in llm agents via multi-turn reinforcement learning. arXiv preprint arXiv:2504.20073.
-
-Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., and Zhou, D. (2022). Chain-of-thought
-
-prompting elicits reasoning in large language models.
-
-Wolfson, T., Trivedi, H., Geva, M., Goldberg, Y., Roth, D., Khot, T., Sabharwal, A., and Tsarfaty, R. (2026). Monaco:
-
-More natural and complex questions for reasoning across dozens of documents. Transactions of the Association
-
-for Computational Linguistics, 14:23–46.
-
-Wu, J., Yin, W., Jiang, Y., Wang, Z., Xi, Z., Fang, R., Zhang, L., He, Y., Zhou, D., Xie, P., and Huang, F. (2025a). Webwalker:
-
-Benchmarking llms in web traversal.
-
-Wu, J., Zhao, Q., Chen, Z., Qin, K., Zhao, Y., Wang, X., and Yao, Y. (2025b). Gap: Graph-based agent planning with
-
-parallel tool use and reinforcement learning.
-
-Xi, Z., Huang, J., Liao, C., Huang, B., Guo, H., Liu, J., Zheng, R., Ye, J., Zhang, J., Chen, W., He, W., Ding, Y., Li, G.,
-
-Chen, Z., Du, Z., Yao, X., Xu, Y., Chen, J., Gui, T., Wu, Z., Zhang, Q., Huang, X., and Jiang, Y.-G. (2025). Agentgym-rl:
-
-Training llm agents for long-horizon decision making through multi-turn reinforcement learning.
-
-Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., Zheng, C., Liu, D., Zhou, F.,
-
-Huang, F., Hu, F., Ge, H., Wei, H., Lin, H., Tang, J., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Zhou, J., Lin,
-
-J., Dang, K., Bao, K., Yang, K., Yu, L., Deng, L., Li, M., Xue, M., Li, M., Zhang, P., Wang, P., Zhu, Q., Men, R., Gao, R.,
-
-Liu, S., Luo, S., Li, T., Tang, T., Yin, W., Ren, X., Wang, X., Zhang, X., Ren, X., Fan, Y., Su, Y., Zhang, Y., Zhang, Y.,
-
-Wan, Y., Liu, Y., Wang, Z., Cui, Z., Zhang, Z., Zhou, Z., and Qiu, Z. (2025). Qwen3 technical report.
-
-Yang, J., Jimenez, C. E., Wettig, A., Lieret, K., Yao, S., Narasimhan, K., and Press, O. (2024). Swe-agent: Agent-computer
-
-interfaces enable automated software engineering.
-
-Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T. L., Cao, Y., and Narasimhan, K. (2023a). Tree of thoughts: Deliberate
-
-problem solving with large language models.
-
-Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K. R., and Cao, Y. (2023b). React: Synergizing reasoning and
-
-acting in language models. In The Eleventh International Conference on Learning Representations.
-
-Zhang, H., Lu, J., Jiang, S., Zhu, C., Xie, L., Zhong, C., Chen, H., Zhu, Y., Du, Y., Gao, Y., Huang, L., Wang, B., Tan, F.,
-
-and Zou, P. (2025). Co-sight: Enhancing llm-based agents via conflict-aware meta-verification and trustworthy
-
-reasoning with structured facts.
-
-Zhou, A., Yan, K., Shlapentokh-Rothman, M., Wang, H., and Wang, Y.-X. (2023). Language agent tree search unifies
-
-reasoning acting and planning in language models. arXiv preprint arXiv:2310.04406.
-
-Zhu, H., Qin, T., Zhu, K., Huang, H., Guan, Y., Xia, J., Yao, Y., Li, H., Wang, N., Liu, P., Peng, T., Gui, X., Li, X., Liu, Y.,
-
-Jiang, Y. E., Wang, J., Zhang, C., Tang, X., Zhang, G., Yang, J., Liu, M., Gao, X., Liu, J., and Zhou, W. (2025). Oagents:
-
-An empirical study of building effective agents.
-
-A PlanFactory Details
-
-We detail the established planning system in PlanFactory as follows:
-
-• Co-Sight
-
-Co-Sight establishes a cross-check net topology, specifically engineered to resolve epistemic discrepancies
-
-through mutual verification. The system is initialized via an inconsistency trigger, where the planning process
-
-is activated only upon detecting conflicting information or divergent perspectives among internal modules.
-
-Navigation is executed through conflict resolution, utilizing trustworthy reasoning with structured facts
-
-to systematically eliminate cognitive biases across the agent collective. For its adaptation mechanism, the
-
-framework employs meta-verification, conducting high-level assessments of the underlying verification logic to
-
-ensure the integrity of the process of building consensus.
-
-15
-
-### Page 16
-
-• AgentOrchestra
-
-AgentOrchestra adheres to an orchestration hierarchy topology, establishing a structured command chain
-
-for multi-agent coordination. The system initiates through role definition, where functional identities are
-
-assigned to activate the environment. During this phase, a planning agent leverages its global perspective to
-
-decompose complex objectives into manageable sub-tasks. Navigation is facilitated via centralized routing, with
-
-the planning agent dispatching specific instructions to specialized sub-agents based on their designated roles.
-
-The framework’s adaptation is driven by environment feedback, where the system dynamically re-calibrates the
-
-plan by synthesizing execution data, aggregating feedback loops, and monitoring cumulative progress toward
-
-the final objective.
-
-• OAgents
-
-OAgents employs a modular graph topology, representing the global objective as a web of decoupled yet
-
-interdependent modules. The framework initiates via SOP configuration, where the agent decomposes the
-
-primary task into sub-tasks interconnected by edges that define prerequisite dependencies. Navigation is driven
-
-by dynamic programming, which, at each discrete step, identifies and dispatches the set of candidate nodes
-
-whose dependencies have been fully satisfied. The system’s adaptation mechanism relies on critic-loop feedback
-
-for periodic refinement: every N steps, intermediate results are cross-referenced against global constraints
-
-to verify alignment with the objective, triggering a re-sequencing of sub-tasks based on novel observations.
-
-Furthermore, trajectories from prior execution attempts are distilled into heuristic guidance and integrated
-
-into the planning module as soft constraints or behavioral preferences, dynamically biasing sub-task selection
-
-toward proven success paths.
-
-• JoyAgent
-
-JoyAgent utilizes a collective hierarchy topology, structuring its multi-agent system to balance global oversight
-
-with local flexibility. the system is initialized through hybrid planning, which implements a supervisor agent
-
-based on a plan-and-execute framework to maintain global coherence while concurrently deploying multiple
-
-single agents utilizing react to ensure step-level responsiveness. navigation is governed by joint deliberation,
-
-where outputs from the diverse agent pool are aggregated and processed through consensus voting to determine
-
-the optimal execution path. the framework’s adaptation is achieved through the intrinsic react loops of the
-
-individual agents, allowing for real-time adjustments based on localized feedback without compromising the
-
-overarching trajectory.
-
-• Flash-Searcher
-
-Upon receiving a request, Flash-Searcher decomposes the task into a parallel Directed Acyclic Graph (DAG),
-
-where nodes denote granular sub-tasks and edges represent their dependencies. The system instantiates this
-
-structure through dependency parsing, mapping out the prerequisite constraints to initialize the graph’s nodes
-
-and edges. Navigation is governed by aggressive parallelization. A node is dispatched to a concurrent execution
-
-pool as soon as its predecessors are satisfied or when partial execution results provide sufficient auxiliary
-
-validation. To maintain system agility, the framework performs workflow pruning at defined step intervals,
-
-where it summarizes progress to excise resolved nodes and re-evaluates the dependencies of pending tasks,
-
-dynamically injecting new decomposition branches if environmental contingencies arise.
-
-• FlowSearch
-
-FlowSearch conceptualizes task resolution through a thought graph topology, representing the reasoning
-
-process as an evolving network of cognitive states. The framework employs flow construction for incremental
-
-instantiation; starting from the root task, a knowledge flow planner iteratively evaluates whether active
-
-nodes require further decomposition or supplemental context. This process generates descendant nodes that
-
-encapsulate sub-problems, intermediate reasoning steps, and required evidentiary grounding while concurrently
-
-establishing dependency edges to preserve logical consistency and structural integrity. Navigation is managed by
-
-a knowledge collector, which identifies and dispatches nodes that exhibit the highest execution readiness based on
-
-satisfied dependencies. The system’s adaptation is realized through dynamic expansion via a knowledge refiner,
-
-which leverages newly acquired insights to perform structural transformations on the flow. By synthesizing
-
-16
-
-### Page 17
-
-current knowledge contexts with execution states, the refiner dynamically executes atomic operations including
-
-the addition, deletion, or modification of nodes and edges to optimize the graph’s trajectory toward the goal.
-
-• OWL
-
-OWL adopts a dual hierarchy topology that formally segregates the strategic management layer from the tactical
-
-execution layer. Upon task arrival, the system undergoes planner decomposition, where a high-level planner
-
-analyzes task complexity against the latent capabilities of available worker nodes to instantiate a structured
-
-task list. Navigation is facilitated via dynamic dispatch, managed by a coordinator that evaluates real-time
-
-agent profiles to map specific sub-tasks to the most suitable worker nodes. The framework’s adaptation logic is
-
-driven by manager intervention triggered by decentralized failure detection: individual workers autonomously
-
-monitor their execution status, broadcasting failure signals to a dedicated task channel upon impasse. This
-
-channel acts as an observation primitive, prompting the planner to perform reactive re-planning and inject
-
-revised sub-tasks based on the contextual feedback from the failed execution.
-
-B Datasets
-
-The five datasets used in this study are described as follows: (1) GAIA (Mialon et al., 2023) consists of 165 tasks,
-
-categorized into 53 Level-1, 86 Level-2, and 26 Level-3 problems. (2) WebWalkerQA (Wu et al., 2025a) evaluates an
-
-agent’s capability in handling complex, multi-turn web interactions. It comprises 680 real-world queries across four
-
-domains and spans over 1, 373 webpages. We sample a subset of 170 queries for evaluation. (3) xBench-DeepSearch
-
-(xBench-DS) (Chen et al., 2025) contains 100 tasks assessing agentic planning, tool use, and reasoning. (4) TaskCraft(Shi
-
-et al., 2025a) is a synthetic benchmark generated via an autonomous data pipeline, we collect 300 queries as a valid
-
-subset.(5) DeepSearchQA (Google, 2025) targets the long-horizon research capabilities of agents, we collect 50 queries
-
-as a valid subset.
-
-C Case Study
-
-To provide a concrete and intuitive understanding of the planning architectures synthesized by TodoEvolve, we
-
-visualize three representative systems generated for distinct query types, as shown in Figures 5 to 7. These examples
-
-demonstrate how our meta-planner moves beyond static templates, dynamically tailoring the control flow—ranging
-
-from linear sequential logic to complex parallel graph structures—to match the specific cognitive impedance and
-
-dependency requirements of the task. By autonomously configuring the topology initialization, execution navigation,
-
-and adaptation triggers, TodoEvolve ensures robust performance across varying levels of problem complexity.
-
-17
-
-### Page 18
-
-Figure 5 Linear Sequential Planning for Multi-Criteria Filtering. For a query requiring strict multi-stage filtering and
-
-calculation (identifying countries based on migration thresholds followed by crime index analysis), TodoEvolve instantiates a linear
-
-execution topology. The system prioritizes a sequential “fetch-and-filter” pipeline to manage data dependencies, incorporating a
-
-periodic adaptation trigger to validate intermediate retrieval results before proceeding to the final synthesis and verification stage.
-
-This structure minimizes branching overhead for tasks where step-wise logical progression is paramount.
-
-18
-
-### Page 19
-
-Figure 6 State-Aware Graph Topology for Structured Data Extraction. Addressing a structured retrieval task involving
-
-sorting and ranking constraints, the meta-planner constructs a Knowledge Flow Graph. This topology decomposes the problem
-
-into granular nodes (acquisition, filtering, and finalization). The navigation strategy employs a state-aware routing mechanism that
-
-dynamically selects between parallel extraction or sequential reasoning based on the current node status ("pending" vs. "success"),
-
-allowing the system to efficiently prune the search space while adhering to numerical constraints.
-
-19
-
-### Page 20
-
-Figure 7 High-Breadth Parallel Planning for Complex Entity Resolution. Faced with a complex entity resolution task
-
-requiring the retrieval of nested attributes for multiple subjects simultaneously, TodoEvolve evolves a highly parallelized graph
-
-architecture. The system identifies independent sub-goals (e.g., retrieving data for different players concurrently) and activates a
-
-“Parallel Executor” module to minimize latency. The adaptation layer monitors the synchronization of these concurrent streams,
-
-ensuring that the graph topology is only updated and merged when specific dependency conditions are met.
-
-20