RegioGraph revision guide

Align the paper’s argument with its evidence.

The opening is strong. The revision should preserve that problem statement while making the results and conclusion answer it in the same order.

01 · Diagnosis

Where the narrative currently breaks

The introduction frames graph substrate and traversal control as the central problem. The results begin with external retriever comparisons, and the conclusion foregrounds efficiency. The paper therefore opens and closes on different contributions.

Primary mismatch

The reader is led to ask whether AGraph is the right graph for agentic document retrieval, but the current clean results do not include the matched AGraph-versus-general-graph 2×2. At the same time, Phase 2 receives substantial method emphasis despite having no independently detected accuracy benefit.

Problem

Document structure carries information that chunk and entity representations can lose.

Method

AGraph makes typed regions and document-native relations directly navigable.

Evaluation

Separate graph substrate, traversal policy, state constraints, semantic augmentation, and retrieval-family context.

Conclusion

Identify the structural graph and bounded traversal interface as the empirical core.

02 · Core thesis

Use one sentence to organize the paper

Documents should be retrieved through their native structural regions; AGraph makes those regions navigable, agentic traversal exploits that graph, and explicit state constraints keep the search efficient and reliable.

AGraph is the representation contribution.
Agentic traversal is the retrieval-policy contribution.
State constraints are the efficiency and control contribution.
Phase 2 is an optional semantic augmentation whose independent accuracy benefit is not detected in the current evaluation.

Recommended positioning

Present RegioGraph as a compositional reference architecture for region-native agentic document retrieval. Prefer “formalize,” “decompose,” “operationalize,” and “reference architecture” over “standardize,” “SOTA,” or “prove synergy.”

03 · Claim hierarchy

Separate the claims by evidence type

Traversal-ready AGraph

Typed document regions and deterministic structural relations become first-class retrieval states. This is the primary design contribution.

AGraph under a matched policy

Under the exact matched-ablation assumption, AGraph outperforms the evaluated general graph by 3.36 EM points under the same agentic traversal policy. If the general graph also uses region nodes, this is not a chunk-versus-region comparison.

Bounded traversal control

State constraints reduce token use by 23.12% and 34.67% and sharply reduce budget exhaustion. This supports an operational claim, not an accuracy claim.

Structural core, optional semantics

The current Phase 2 does not outperform Phase 1 or the randomized control. The deterministic structural backbone is therefore the empirical core.

Interaction decision

The observed interaction is (76.0−73.8)−(72.6−72.1)=+1.7 EM points. The interaction claim depends on its confidence interval.

If the interval excludes zero

Claim a positive interaction: “AGraph and agentic traversal exhibit a positive interaction, supporting their joint design.”

If the interval includes zero

Do not claim synergy or say that the two components should be co-designed. Keep the matched AGraph comparison and frame joint evaluation as a methodological principle.

04 · Abstract

Close the abstract in six steps

Use the sequence problem → decomposition → implementation → architecture evidence → operational evidence → boundary finding.

Observation

Document organization encodes relations needed to interpret structure-dependent evidence.

Gap

Chunk and entity-centric retrieval remove or transform these relations, while adaptive agents still need an actionable document state space.

Formalization

RegioGraph separates graph substrate, traversal policy, framework-owned state, bounded actions and budgets, and synthesis.

Instantiation

AGraph uses typed regions and a deterministic structural backbone with optional semantic augmentation; the navigator operates under explicit state constraints.

Positive evidence

Report the matched AGraph-versus-general-graph comparison, followed by token and exhaustion reductions.

Boundary

Report the Phase 2 null result and identify the structural graph and bounded traversal interface as the empirical core.

Proposed abstract skeleton

Documents encode information through layout and structural relations, yet conventional retrieval systems flatten them into text chunks or entity triples. We formulate agentic document retrieval as the composition of a graph substrate, a traversal policy, framework-owned search state, bounded actions and budgets, and answer synthesis. RegioGraph instantiates this formulation with AGraph, whose typed region nodes and deterministic structural edges expose document-native relations as traversal actions, with optional semantic augmentation. In a matched factorial evaluation, AGraph outperforms the evaluated general graph under the same agentic traversal policy. State constraints substantially reduce token use and budget exhaustion without a detected EM difference. Meanwhile, the current semantic augmentation provides no independently detected benefit, identifying the structural graph and bounded traversal interface as the empirical core of the system.

Use absolute scores as supporting evidence

The 76.32/62.69 scores and the +7.08/+9.35 gains over dense retrieval may be added, but the abstract should end on the architectural finding. Do not imply a general advantage over BM25+ or Hybrid RRF.

05 · Introduction

Give each paragraph one job

P1 · Structure

Document structure is an information-encoding representation, not merely a presentation layer.

P2 · Flattening

Chunking removes section, page-local, and cross-region relations as explicit retrieval objects.

P3 · Prior gap

Entity GraphRAG makes relations searchable; agentic retrieval adapts search. Neither alone specifies a document-native state space and a controlled traversal interface.

P4 · Formalization

Define (G, π, M, C, B, S): graph, policy, state and memory, constraints, budget, and synthesis.

P5 · AGraph

Introduce typed regions and deterministic structural edges as the core. Present semantic links as an optional augmentation.

P6 · Control

Explain that the model selects the next permitted action while the framework owns state, duplicate blocking, valid actions, and stopping.

P7 · Evaluation

State the five research questions so the results can answer them in the same order.

Research questions

RQ1 · Graph substrate

Does AGraph outperform the evaluated general graph under the same traversal policy?

RQ2 · Policy relationship

Does the effect of the graph depend on the traversal policy, and what is the interaction interval?

RQ3 · State constraints

How do state constraints affect accuracy, tokens, search depth, and budget exhaustion?

RQ4 · Semantic augmentation

Does Phase 2 improve over the structural Phase 1 graph?

RQ5 · Retrieval context

How does the complete system compare with dense, sparse, and hybrid retrieval?

06 · Contributions

Make each contribution match an evidence layer

1. A compositional formulation of agentic document retrieval

Separate the graph substrate, traversal policy, framework-owned search state, action and evidence budgets, and answer synthesis through an explicit bounded traversal interface.

2. A traversal-ready document graph

Make typed layout regions first-class retrieval states and connect them through a deterministic structural backbone, with semantic links evaluated as an optional augmentation.

3. A controlled empirical decomposition

Use a matched factorial comparison to isolate graph substrate and traversal policy; use further controls to quantify constraint efficiency and test semantic augmentation.

Avoid unsupported wording

Do not use “outperforms all systems,” “Phase 2 improves accuracy,” “traversal and AGraph independently contribute,” “proves synergy,” or “standardizes Agentic GraphRAG.”

07 · Results order

Answer the architecture question first

5.1 · Protocol

Complete coverage, shared parser and evaluator, fixed backbone, judge-based EM, independent F1, and paired document-cluster bootstrap.

5.2 · Graph × policy

Present the 2×2 as the main architectural evidence. Report direct contrasts separately from the interaction.

5.3 · State constraints

Conclude that constraints improve efficiency and stability, not accuracy.

5.4 · Structural core

Compare Phase 1, Phase 1+2, and randomized Phase 2. State that the semantic-edge benefit is not detected.

5.5 · Navigation depth

Separate the benchmark-dependent search-depth trade-off from graph construction.

5.6 · Retrieval context

Report the clear dense-retrieval gain, the competitive sparse and hybrid baselines, and the additional model-mediated token cost.

If the 2×2 score differs from the current main score

Label it as a matched factorial suite and identify its snapshot, evaluator, and protocol. Use the paired contrasts within the 2×2; do not directly compare its 76.0 cell with the 76.32 score in the current main table.

08 · Evidence matrix

Map each claim to its support

Claim	Evidence	Status	Allowed wording
AGraph beats the evaluated general graph under agentic traversal.	Matched 2×2 direct contrast: +3.36 EM, CI [+1.14, +5.58].	Supported*	“Under the same agentic traversal policy…”
AGraph and agentic traversal exhibit synergy.	Observed interaction: +1.7; interval required.	Conditional	Use “positive interaction” only if the interval excludes zero.
State constraints improve accuracy.	Accuracy intervals include zero on both benchmarks.	Not detected	Do not claim an accuracy gain.
State constraints improve efficiency and stability.	Tokens fall 23.12%/34.67%; exhaustion falls to 4.72%/4.95%.	Supported	“Substantially reduce token use and budget exhaustion.”
Phase 2 semantic links improve accuracy.	Full-versus-Phase-1 and randomized-control intervals include zero.	Not detected	Describe Phase 2 as optional and report the null result.
RegioGraph broadly outperforms standard retrieval.	Clear gain over dense retrieval; BM25+ and Hybrid intervals reach or cross zero.	Partial	Claim the dense gain and call sparse and hybrid retrieval competitive.

* This assumes that the 2×2 holds policy, backbone, budget, evaluator, and coverage fixed and is explicitly re-adopted into the current manuscript.

09 · Conclusion

Close the paper in the same order it opens

Proposed conclusion core

RegioGraph formalizes agentic document retrieval as the composition of a traversal-ready graph substrate and bounded search control. In the matched factorial evaluation, AGraph outperforms the evaluated general graph under the same agentic traversal policy. State constraints substantially reduce token use and budget exhaustion without a detected EM change, while the present semantic augmentation does not improve over either the structural core or its randomized control. These findings identify the deterministic region-structural graph and bounded traversal interface—not semantic graph complexity alone—as the empirical core of RegioGraph.

Add a positive-interaction sentence only if the interaction interval excludes zero. Otherwise, do not use “joint design” or “co-designed” in the conclusion.