RoadMapper: A Multi-Agent System for Roadmap Generation of Solving Complex Research Problems

Haihong E; Haocheng Gao; Jiacheng Liu; Linwei Jia; Rongjin Li; Ruofei Bai; Shiyao Peng; Xinyi Hu; Xueyuan Lin; Zhongjun Yang

arxiv: 2604.27616 · v1 · submitted 2026-04-30 · 💻 cs.CL · cs.MA

RoadMapper: A Multi-Agent System for Roadmap Generation of Solving Complex Research Problems

Jiacheng Liu , Zichen Tang , Zhongjun Yang , Xinyi Hu , Xueyuan Lin , Linwei Jia , Ruofei Bai , Rongjin Li

show 3 more authors

Shiyao Peng Haocheng Gao Haihong E

This is my paper

Pith reviewed 2026-05-07 05:32 UTC · model grok-4.3

classification 💻 cs.CL cs.MA

keywords roadmap generationmulti-agent systemslarge language modelsresearch problem solvingtask decompositionbenchmark evaluation

0 comments

The pith

RoadMapper is a multi-agent LLM system that generates higher-quality roadmaps for complex research problems by adding knowledge augmentation and iterative critique.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces RoadMap, a benchmark for testing how well large language models create hierarchical step-by-step guides to solve difficult research questions. It identifies three recurring LLM weaknesses: missing professional knowledge, unbalanced task splits, and tangled logical order. To overcome these, the authors built RoadMapper, a multi-agent setup that runs three stages in sequence: a first draft, an external knowledge injection step, and repeated rounds of critique, revision, and scoring. Experiments on the benchmark show the system lifts average scores by more than eight percent while cutting the time human experts need by eighty-four percent. This points to a practical way for AI agents to support structured planning in scientific work.

Core claim

RoadMapper is an LLM-based multi-agent system that decomposes roadmap generation into initial generation, knowledge augmentation, and iterative critique-revise-evaluate stages. On the RoadMap benchmark it raises average performance more than eight percent above single-model baselines and reduces the time required by human experts by eighty-four percent.

What carries the argument

RoadMapper, a multi-agent framework that decomposes roadmap generation into initial generation, knowledge augmentation, and iterative critique-revise-evaluate stages.

If this is right

Single LLMs can be upgraded to produce more usable research plans without full retraining.
The three-stage pipeline reduces the need for human experts to supply initial structure and domain facts.
Iterative critique loops can be reused for other hierarchical planning tasks such as project scheduling.
Knowledge augmentation from external sources becomes a standard fix for LLM knowledge gaps in structured output.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the benchmark problems are representative, similar multi-agent patterns could shorten the planning phase of many AI-assisted research pipelines.
Adding live search or database lookup inside the knowledge-augmentation stage would likely raise performance further on current topics.
The same critique-revise loop might transfer to domains outside research, such as policy design or engineering project roadmaps.

Load-bearing premise

The RoadMap benchmark tasks and scoring criteria accurately capture the quality and real-world usefulness of roadmaps for complex research problems.

What would settle it

Running RoadMapper on a fresh collection of research problems outside the RoadMap benchmark and measuring both roadmap quality and expert time savings against the same human baselines.

Figures

Figures reproduced from arXiv: 2604.27616 by Haihong E, Haocheng Gao, Jiacheng Liu, Linwei Jia, Rongjin Li, Ruofei Bai, Shiyao Peng, Xinyi Hu, Xueyuan Lin, Zhongjun Yang, Zichen Tang.

**Figure 1.** Figure 1: We propose the roadmap generation task. Prior answer formats and methods face multiple challenges when solving complex research problems, but RoadMapper effectively addresses these challenges through an iterative “critique-revise-evaluate” process. sulting professional knowledge, meticulously designing, and iteratively conducting reviews. However, this manual process is both time-consuming and resource-… view at source ↗

**Figure 2.** Figure 2: Overview of RoadMap. ICE: Information and Communication Engineering, CS: Computer Science, EST: Electronic Science and Technology, CSE: Control Science and Engineering, EE: Electrical Engineering, Agr: Agronomy, Econ: Economics, Edu: Education, ME: Mechanical Engineering, Lit: Literature. Our main contributions are as follows: • We define the roadmap generation task and construct RoadMap, providing extensi… view at source ↗

**Figure 3.** Figure 3: Overview of RoadMapper, a multi-agent system consisting of six agents: Init Agent (I), Knowledge Agent (K), Logic Critique Agent (L), Granularity Critique Agent (G), Revise Agent (R), and Evaluate Agent (E). 4.1.4 Granularity Critique Agent (G) This agent is responsible for critiquing the granularity of nodes and outputting revision suggestions: G : GCt = G(yt). (5) Here, yt denotes the roadmap to be crit… view at source ↗

**Figure 4.** Figure 4: Quantitative analysis between enforced N iterations of “critique-revise” and performance. view at source ↗

**Figure 5.** Figure 5: Cost comparison between methods on the roadmap generation task. DP denotes Direct Prompting. Time and Token Efficiency. As shown in Figure 5, we comparatively analyzed the time and token consumption of different methods in roadmap design. In terms of time, all LLM-based automated methods significantly outperformed human experts, among which RoadMapper requires the least time in complex systems, saving 3… view at source ↗

**Figure 6.** Figure 6: Distribution of the publication year of our view at source ↗

**Figure 7.** Figure 7: Template for extracting core research problems and skill points from dissertations. view at source ↗

**Figure 8.** Figure 8: Prompt of the K agent for generation of internal knowledge view at source ↗

**Figure 9.** Figure 9: Prompt of the K agent for knowledge augmentation view at source ↗

**Figure 10.** Figure 10: Prompt of the E agent view at source ↗

**Figure 11.** Figure 11: Case study of RoadMapper view at source ↗

**Figure 12.** Figure 12: Example of Golden-Roadmap view at source ↗

read the original abstract

People commonly leverage structured content to accelerate knowledge acquisition and research problem solving. Among these, roadmaps guide researchers through hierarchical subtasks to solve complex research problems step by step. Despite progress in structured content generation, the roadmap generation task has remained unexplored. To bridge this gap, we introduce RoadMap, a novel benchmark designed to evaluate the ability of large language models (LLMs) to construct high-quality roadmaps for solving complex research problems. Based on this, we identify three limitations of LLMs: (1) lack of professional knowledge, (2) unreasonable task decomposition, and (3) disordered logical relationships. To address these challenges, we propose RoadMapper, an LLM-based multi-agent system that decomposes the research roadmap generation task into three key stages (i.e., initial generation, knowledge augmentation, and iterative "critique-revise-evaluate"). Extensive experiments demonstrate that RoadMapper can improve LLMs' ability for roadmap generation, while enhancing average performance by more than 8% and saving 84% of the time required by human experts, highlighting its effectiveness and application potential.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

RoadMapper defines a new roadmap-generation task and ships a three-stage multi-agent fix plus its own benchmark, but the 8% gain and 84% time-saving numbers rest on evaluation details that are still missing.

read the letter

The paper's main contribution is defining roadmap generation as a distinct task for LLMs and releasing the RoadMap benchmark to measure it, together with a three-stage multi-agent system that first generates a draft, then augments it with external knowledge, and finally runs iterative critique-revise-evaluate loops. That decomposition directly targets the three LLM weaknesses the authors list—missing professional knowledge, poor task breakdown, and tangled logic—and the pipeline is a straightforward, concrete extension of existing multi-agent patterns rather than a wholly new mechanism. The time-saving claim is the part that stands out most: if the 84% figure holds against real expert effort, it would be a practical win for anyone using LLMs to structure research plans. What the work does cleanly is name the limitations and tie each one to a stage, which makes the design easy to understand and replicate in principle. The soft spots sit in the evaluation. The abstract reports an average performance lift above 8% and the large time reduction, yet it gives no rubric for scoring roadmap quality, no count of problems or domains tested, no list of baselines, no inter-annotator numbers, and no description of how human-expert time was actually measured. Because the benchmark is new and author-created, those omissions leave open the possibility that the gains are partly an artifact of the scoring criteria or task selection. The stress-test note is accurate on this point: without those details the central claims cannot be checked. This paper is aimed at people already working on LLM agents for research assistance or automated planning. A reader in that niche could pull the stage structure and the task definition for their own experiments, but only after seeing the full results section and any human validation data. I would bring it to a reading group to talk through whether the three-stage split is novel enough to justify the new benchmark. I would not cite it yet. It deserves peer review because the task is fresh and the architecture is explicit, but the authors will need to supply the missing experimental protocol and external checks before the performance numbers can be treated as solid evidence.

Referee Report

3 major / 2 minor

Summary. The paper introduces the RoadMap benchmark to evaluate LLMs on generating hierarchical roadmaps for complex research problems and proposes RoadMapper, a multi-agent system that decomposes the task into initial generation, knowledge augmentation, and iterative critique-revise-evaluate stages. It identifies three LLM limitations (lack of professional knowledge, unreasonable decomposition, disordered logic) and claims that RoadMapper improves average LLM performance by more than 8% while saving 84% of the time required by human experts.

Significance. If the evaluation protocol proves robust, the work could advance AI-assisted research planning by providing a structured multi-agent framework for roadmap generation and a dedicated benchmark in an underexplored area. The introduction of a benchmark and the explicit targeting of LLM weaknesses via critique-revise cycles are constructive contributions; the reported time savings, if independently verifiable, would have clear practical value for researchers.

major comments (3)

[Experimental evaluation / Results] The abstract and evaluation sections claim >8% average performance gains and 84% human time savings, but provide no details on experimental design: number of problems/domains tested, exact scoring rubric for roadmap quality (completeness, coherence, usefulness), inter-annotator agreement, statistical significance tests, or how human expert time was measured and compared. These omissions make it impossible to assess whether the gains are reliable or generalizable.
[RoadMap benchmark construction] RoadMap is a newly introduced benchmark created by the authors, creating a circularity risk where performance is measured against a self-defined standard. The manuscript must detail task curation criteria, validation against external or real-world outcomes, and whether tasks were selected post-hoc to favor the proposed knowledge-augmentation approach.
[Baselines and ablation studies] Baseline comparisons are insufficiently specified: it is unclear which exact systems (vanilla GPT-4, other agent frameworks, or ablations of RoadMapper) were used, whether LLM-as-judge scoring was calibrated against human experts, and how the three-stage pipeline was ablated to isolate the contribution of each component.

minor comments (2)

[Abstract] The abstract refers to 'extensive experiments' without any quantitative details or pointers to specific tables/figures; add a brief summary of key metrics and sample sizes.
[Introduction / Benchmark] Clarify the precise definition of 'roadmap quality' early in the paper and ensure the scoring criteria are reproducible (e.g., provide the full rubric in an appendix).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments identify important areas where the manuscript requires greater transparency and elaboration. We address each major comment point by point below and will incorporate the necessary clarifications and expansions in the revised version to improve verifiability and robustness.

read point-by-point responses

Referee: [Experimental evaluation / Results] The abstract and evaluation sections claim >8% average performance gains and 84% human time savings, but provide no details on experimental design: number of problems/domains tested, exact scoring rubric for roadmap quality (completeness, coherence, usefulness), inter-annotator agreement, statistical significance tests, or how human expert time was measured and compared. These omissions make it impossible to assess whether the gains are reliable or generalizable.

Authors: We acknowledge that the initial submission omitted critical details on the experimental protocol, making independent assessment difficult. We will revise the evaluation section to provide a complete description of the design, including the number of problems and domains tested, the precise scoring rubric (with explicit criteria and scales for completeness, coherence, and usefulness), inter-annotator agreement statistics, results of statistical significance testing, and the exact protocol used to measure and compare human expert time (including how scratch generation time versus review-and-edit time was recorded). These elements were part of our internal experimental process but were not fully documented in the original manuscript; the revision will rectify this. revision: yes
Referee: [RoadMap benchmark construction] RoadMap is a newly introduced benchmark created by the authors, creating a circularity risk where performance is measured against a self-defined standard. The manuscript must detail task curation criteria, validation against external or real-world outcomes, and whether tasks were selected post-hoc to favor the proposed knowledge-augmentation approach.

Authors: We share the concern regarding potential circularity and will add a dedicated subsection on benchmark construction. This will specify the task curation criteria (e.g., requirements for multi-stage complexity and domain coverage), the validation process against external or real-world outcomes (including expert review and alignment checks with published research trajectories), and explicit confirmation that the benchmark was finalized prior to any RoadMapper experiments, with no post-hoc filtering or selection of tasks to advantage the knowledge-augmentation component. We will also include representative examples and documentation of the curation workflow. revision: yes
Referee: [Baselines and ablation studies] Baseline comparisons are insufficiently specified: it is unclear which exact systems (vanilla GPT-4, other agent frameworks, or ablations of RoadMapper) were used, whether LLM-as-judge scoring was calibrated against human experts, and how the three-stage pipeline was ablated to isolate the contribution of each component.

Authors: We agree that the descriptions of baselines and ablations were insufficiently detailed. In the revision we will explicitly list all baseline systems (including vanilla GPT-4, chain-of-thought variants, and other agent frameworks), clarify that primary evaluation relies on human experts while also reporting LLM-as-judge results with calibration details against human annotations on a held-out subset, and present full ablation results that isolate the contribution of each stage in the three-stage pipeline (initial generation, knowledge augmentation, and critique-revise-evaluate). Quantitative performance drops for each ablation will be reported in tables. revision: yes

Circularity Check

0 steps flagged

No significant circularity; evaluation uses standard new-benchmark protocol without self-referential reduction.

full rationale

The paper introduces the RoadMap benchmark to evaluate LLM roadmap generation, identifies three LLM limitations from it, and proposes the RoadMapper multi-agent pipeline (initial generation, knowledge augmentation, critique-revise-evaluate) to address them. Performance claims (>8% average improvement, 84% human time savings) are empirical results measured on the new benchmark against baselines. No mathematical derivation chain, equations, or first-principles results exist that reduce to inputs by construction. No self-citations are invoked as load-bearing uniqueness theorems or ansatzes. The benchmark and system are presented as novel contributions for an unexplored task; measuring system performance on an author-introduced benchmark is standard practice and does not equate the output to the input definitionally. The central claims remain independently falsifiable via the reported experimental protocol.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 2 invented entities

The central claim depends on the validity of the new RoadMap benchmark as a proxy for real research problem-solving utility and on the assumption that multi-agent critique cycles reliably address the three stated LLM limitations. No free parameters are explicitly named, but the 8% and 84% figures are empirical outcomes that function as fitted results on the authors' benchmark.

free parameters (1)

RoadMap benchmark scoring criteria
The metric for 'average performance' improvement is not defined in the abstract and appears to be defined by the authors' evaluation protocol.

axioms (2)

domain assumption LLMs exhibit the three specific limitations (lack of professional knowledge, unreasonable task decomposition, disordered logical relationships) in roadmap generation.
Stated as identified limitations without supporting evidence or citation in the abstract.
ad hoc to paper Iterative critique-revise-evaluate cycles in a multi-agent setup can systematically overcome these limitations.
Core design choice of RoadMapper presented without prior justification or external validation.

invented entities (2)

RoadMap benchmark no independent evidence
purpose: To evaluate LLM ability to construct high-quality roadmaps for complex research problems.
Newly created evaluation resource; no independent evidence of its validity or correlation with actual research success is provided.
RoadMapper multi-agent system no independent evidence
purpose: To decompose roadmap generation into initial generation, knowledge augmentation, and iterative critique-revise-evaluate stages.
Newly proposed architecture whose effectiveness is claimed via internal experiments only.

pith-pipeline@v0.9.0 · 5528 in / 1958 out tokens · 81836 ms · 2026-05-07T05:32:02.880331+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 3 canonical work pages · 3 internal anchors

[1]

InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4762–4779, Flo- rence, Italy

COMET: Commonsense transformers for auto- matic knowledge graph construction. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4762–4779, Flo- rence, Italy. Association for Computational Linguis- tics. Philip E Burian, Lynda Rogerson, and Francis R Maf- fei III. 2010. The research roadmap: A primer to the ap...

2010
[2]

Training verifiers to solve math word prob- lems.Preprint, arXiv:2110.14168. DeepSeek-AI, Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chen- hao Xu, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, and 245 others. 2025. Deepseek-v3.2: Pushin...

work page internal anchor Pith review arXiv 2025
[3]

A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration

STRUCTSUM generation for faster text com- prehension. InProceedings of the 62nd Annual Meet- ing of the Association for Computational Linguis- tics (Volume 1: Long Papers), pages 7876–7896, Bangkok, Thailand. Association for Computational Linguistics. Bingxuan Li, Yiwei Wang, Jiuxiang Gu, Kai-Wei Chang, and Nanyun Peng. 2025a. METAL: A multi-agent framewo...

work page internal anchor Pith review arXiv 2023
[4]

gpt-oss-120b & gpt-oss-20b Model Card

TriageAgent: Towards better multi-agents col- laborations for large language model-based clinical triage. InFindings of the Association for Computa- tional Linguistics: EMNLP 2024, pages 5747–5764, Miami, Florida, USA. Association for Computational Linguistics. Yaojie Lu, Qing Liu, Dai Dai, Xinyan Xiao, Hongyu Lin, Xianpei Han, Le Sun, and Hua Wu. 2022. U...

work page internal anchor Pith review arXiv 2024
[5]

Extraction must be based strictly on the content of the provided paper
[6]

Extract at most five of the most innovative and valuable skill points
[7]

Extract the core research task of the entire paper
[8]

Answer in JSON format and follow the template provided below
[9]

Enclose the extracted results with```json```
[10]

core_research_question

Answer in English. Template: { "core_research_question": "" # The core research question in a concise question format like "How to ...?" (no more than 30 words). "skill_points": [ { "problem_description": "", # Describe the specific problem this skill point addresses. "skill_point_name": "", # Concise name of the skill point. "skill_point_description": ""...
[11]

Analyze the research problem and output several key skills that are essential for solving the problem
[12]

-`name`: The name of the skill point (brief description)

Output format requirements: - Use JSON list format for output, with each item being a skill point including name and description fields. -`name`: The name of the skill point (brief description). -`description`: The detailed explanation of the skill point (explains the role of the skill point in solving the problem). - Enclose the output in```json```
[13]

name": "xxx

The skills should be answered in {split_language}. Example output format: ```json [ {{"name": "xxx", "description": "xxx"}}, {{"name": "xxx", "description": "xxx"}}, ... ] ``` Figure 8: Prompt of theKagent for generation of internal knowledge. Prompt of theKAgent for Knowledge Augmentation You are an experienced research expert, specialized in optimizing ...
[14]

Carefully analyze each skill point in the skill point repository (each skill point includes its name and description) according to the research problem and the current roadmap
[15]

Determine which skill points are helpful for improving the roadmap, insert the names of helpful skill points as a step node into appropriate positions in the roadmap (names can be adapted to the problem as needed), and ignore unhelpful skill points
[16]

Ensure the correct format of the roadmap during insertion, maintaining the logical order and continuity of node indices
[17]

- Use different heading levels (#, ##, ###, etc.) to indicate the node level and the hierarchical structure of the roadmap

Roadmap format requirements, make sure to strictly follow: - Use Markdown format for output, with each line as a node (including level, index, and title, separated by spaces). - Use different heading levels (#, ##, ###, etc.) to indicate the node level and the hierarchical structure of the roadmap. - Use indices like 1.1.1 to indicate the node's position ...
[18]

Enclose the output in```markdown```
[19]

Example output format: ```markdown # 1 [Main Step] ## 1.1 [Sub-step] ## 1.2 [Sub-step] # 2 [Main Step] ## 2.1 [Sub-step] ### 2.1.1 [Sub-step] ### 2.1.2 [Sub-step]

The roadmap content should be answered in {split_language}. Example output format: ```markdown # 1 [Main Step] ## 1.1 [Sub-step] ## 1.2 [Sub-step] # 2 [Main Step] ## 2.1 [Sub-step] ### 2.1.1 [Sub-step] ### 2.1.2 [Sub-step] ... ``` Figure 9: Prompt of theKagent for knowledge augmentation. Prompt of theEAgent You are an experienced research expert, speciali...
[20]

Logic Structure: Evaluate the logical coherence of the roadmap, including whether the dependencies between nodes are reasonable, whether the step order is logical, and whether there are conflicts or contradictions between different parts
[21]

Granularity Degree: Evaluate the rationality of task decomposition granularity, including whether there are too macroscopic or too microscopic nodes, whether the sub-task division is balanced, etc
[22]

Topic Relevance: Evaluate the relevance of the roadmap to the input research problem, including whether the roadmap fully covers the core content required to solve the problem, whether there are redundant nodes unrelated to the topic, and whether the key technical points are fully reflected, etc
[23]

Based on the four dimensions above, provide an overall assessment using the following 100-point scale (percentage system, 0-100): 0-19: Very poor quality, cannot be used

Completeness: Evaluate the richness and completeness of the roadmap content, including whether there are missing key nodes or steps, and whether all the content required to solve the problem is fully covered. Based on the four dimensions above, provide an overall assessment using the following 100-point scale (percentage system, 0-100): 0-19: Very poor qu...

[1] [1]

InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4762–4779, Flo- rence, Italy

COMET: Commonsense transformers for auto- matic knowledge graph construction. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4762–4779, Flo- rence, Italy. Association for Computational Linguis- tics. Philip E Burian, Lynda Rogerson, and Francis R Maf- fei III. 2010. The research roadmap: A primer to the ap...

2010

[2] [2]

Training verifiers to solve math word prob- lems.Preprint, arXiv:2110.14168. DeepSeek-AI, Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chen- hao Xu, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, and 245 others. 2025. Deepseek-v3.2: Pushin...

work page internal anchor Pith review arXiv 2025

[3] [3]

A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration

STRUCTSUM generation for faster text com- prehension. InProceedings of the 62nd Annual Meet- ing of the Association for Computational Linguis- tics (Volume 1: Long Papers), pages 7876–7896, Bangkok, Thailand. Association for Computational Linguistics. Bingxuan Li, Yiwei Wang, Jiuxiang Gu, Kai-Wei Chang, and Nanyun Peng. 2025a. METAL: A multi-agent framewo...

work page internal anchor Pith review arXiv 2023

[4] [4]

gpt-oss-120b & gpt-oss-20b Model Card

TriageAgent: Towards better multi-agents col- laborations for large language model-based clinical triage. InFindings of the Association for Computa- tional Linguistics: EMNLP 2024, pages 5747–5764, Miami, Florida, USA. Association for Computational Linguistics. Yaojie Lu, Qing Liu, Dai Dai, Xinyan Xiao, Hongyu Lin, Xianpei Han, Le Sun, and Hua Wu. 2022. U...

work page internal anchor Pith review arXiv 2024

[5] [5]

Extraction must be based strictly on the content of the provided paper

[6] [6]

Extract at most five of the most innovative and valuable skill points

[7] [7]

Extract the core research task of the entire paper

[8] [8]

Answer in JSON format and follow the template provided below

[9] [9]

Enclose the extracted results with```json```

[10] [10]

core_research_question

Answer in English. Template: { "core_research_question": "" # The core research question in a concise question format like "How to ...?" (no more than 30 words). "skill_points": [ { "problem_description": "", # Describe the specific problem this skill point addresses. "skill_point_name": "", # Concise name of the skill point. "skill_point_description": ""...

[11] [11]

Analyze the research problem and output several key skills that are essential for solving the problem

[12] [12]

-`name`: The name of the skill point (brief description)

Output format requirements: - Use JSON list format for output, with each item being a skill point including name and description fields. -`name`: The name of the skill point (brief description). -`description`: The detailed explanation of the skill point (explains the role of the skill point in solving the problem). - Enclose the output in```json```

[13] [13]

name": "xxx

The skills should be answered in {split_language}. Example output format: ```json [ {{"name": "xxx", "description": "xxx"}}, {{"name": "xxx", "description": "xxx"}}, ... ] ``` Figure 8: Prompt of theKagent for generation of internal knowledge. Prompt of theKAgent for Knowledge Augmentation You are an experienced research expert, specialized in optimizing ...

[14] [14]

Carefully analyze each skill point in the skill point repository (each skill point includes its name and description) according to the research problem and the current roadmap

[15] [15]

Determine which skill points are helpful for improving the roadmap, insert the names of helpful skill points as a step node into appropriate positions in the roadmap (names can be adapted to the problem as needed), and ignore unhelpful skill points

[16] [16]

Ensure the correct format of the roadmap during insertion, maintaining the logical order and continuity of node indices

[17] [17]

- Use different heading levels (#, ##, ###, etc.) to indicate the node level and the hierarchical structure of the roadmap

Roadmap format requirements, make sure to strictly follow: - Use Markdown format for output, with each line as a node (including level, index, and title, separated by spaces). - Use different heading levels (#, ##, ###, etc.) to indicate the node level and the hierarchical structure of the roadmap. - Use indices like 1.1.1 to indicate the node's position ...

[18] [18]

Enclose the output in```markdown```

[19] [19]

Example output format: ```markdown # 1 [Main Step] ## 1.1 [Sub-step] ## 1.2 [Sub-step] # 2 [Main Step] ## 2.1 [Sub-step] ### 2.1.1 [Sub-step] ### 2.1.2 [Sub-step]

The roadmap content should be answered in {split_language}. Example output format: ```markdown # 1 [Main Step] ## 1.1 [Sub-step] ## 1.2 [Sub-step] # 2 [Main Step] ## 2.1 [Sub-step] ### 2.1.1 [Sub-step] ### 2.1.2 [Sub-step] ... ``` Figure 9: Prompt of theKagent for knowledge augmentation. Prompt of theEAgent You are an experienced research expert, speciali...

[20] [20]

Logic Structure: Evaluate the logical coherence of the roadmap, including whether the dependencies between nodes are reasonable, whether the step order is logical, and whether there are conflicts or contradictions between different parts

[21] [21]

Granularity Degree: Evaluate the rationality of task decomposition granularity, including whether there are too macroscopic or too microscopic nodes, whether the sub-task division is balanced, etc

[22] [22]

Topic Relevance: Evaluate the relevance of the roadmap to the input research problem, including whether the roadmap fully covers the core content required to solve the problem, whether there are redundant nodes unrelated to the topic, and whether the key technical points are fully reflected, etc

[23] [23]

Based on the four dimensions above, provide an overall assessment using the following 100-point scale (percentage system, 0-100): 0-19: Very poor quality, cannot be used

Completeness: Evaluate the richness and completeness of the roadmap content, including whether there are missing key nodes or steps, and whether all the content required to solve the problem is fully covered. Based on the four dimensions above, provide an overall assessment using the following 100-point scale (percentage system, 0-100): 0-19: Very poor qu...