Documentation

Analysis workflow

Eight connected views transform one manuscript into a research-grounded revision workspace.

Dependency-aware execution.Later stages receive structured evidence from earlier stages. The full review uses the research landscape, method analysis, and macro-logic findings; it is not an isolated prompt.

1. Paper structure

The ingestion stage creates a normalized paper representation containing title, abstract, section hierarchy, body text, references, figures, tables, and source metadata.

PDF path

MinerU runs in an on-demand Docker container. Its Markdown and extracted assets are preserved under the project directory. A constrained reconstruction pass fixes layout artifacts such as hard line wraps, column-order errors, and stray figure text without rewriting scientific claims.

LaTeX path

The selected main file is expanded through safe local inputs, while hierarchy and source-file boundaries are retained. Appendix and references are classified so they can be treated differently in later analysis.

2. Research landscape

The landscape begins with manuscript signals: title, abstract, introduction, related work, in-text citations, and the reference list. It creates several research directions rather than one broad query.

  1. Generate diverse queries for the problem, method family, benchmarks, competing techniques, and adjacent approaches.
  2. Search general web results plus academic providers such as arXiv, OpenReview, and DBLP.
  3. Analyze the first round and issue follow-up queries for gaps, ancestors, and closely competing work.
  4. Resolve citations from related work against the reference list and retrieve cited papers when possible.
  5. Build evidence-backed direction graphs in which nodes are papers and edges describe influence, extension, comparison, or shared lineage.
  6. State where the submitted paper sits and list novelty risks with uncertainty.

The visible graph is an explanatory map, not a complete bibliometric citation graph. Every strong positioning claim should remain traceable to a retrieved source.

3. Compliance review

Mechanical review is intentionally separated from scientific review. Venue policies provide the expected review scale and stable requirements; local PDF/source checks inspect page count, anonymity, font embedding, paper dimensions, and main-file content.

Cycle-specific policies can change. CciipScope labels requirements that still require verification against the official call for papers.

4. Method analysis

This stage reads the complete method-relevant manuscript context and receives the research landscape for comparison. It examines:

5. Macro logic and consistency

Macro analysis follows the paper's argument from motivation to conclusion. It checks whether the problem leads to the method, the experiments test the method's claims, and the conclusions stay within the evidence.

A dedicated consistency pass compares numbers and claims across the abstract, main text, equations, figures, tables, experiment sections, conclusions, and appendices. Potential hard contradictions are kept distinct from presentation improvements.

6. Full review

Four specialized review agents work in parallel:

PerspectivePrimary evidenceFocus
Background and originalityPaper + research landscapeMotivation, related work, novelty, significance
MethodologyPaper + method analysisCorrectness, assumptions, alternatives, technical value
Logic and consistencyPaper + macro analysisClaim support, contradictions, scope of conclusions
Results and writingComplete paperExperiments, data integrity signals, clarity, reproducibility

Each perspective must produce an internal score, independent questions, strengths, weaknesses, and evidence. A deliberation stage then reconciles disagreements and determines the decisive evidence. The final synthesizer preserves the stable user-facing format: summary, key scores, strengths, weaknesses, reviewer questions, required changes, and recommendation.

Scores use the selected conference's native scale. Repairable presentation or ablation gaps should not automatically become hard blockers; missing core evidence, invalid methodology, serious contradictions, or integrity concerns can.

7. Section analysis

Each non-appendix section receives its own analysis. The interface keeps the parsed original beside section-role and logic feedback so recommendations remain anchored to the manuscript. References and appendices are excluded from ordinary prose-polishing analysis.

The stage examines section purpose, paragraph flow, unsupported transitions, missing content, local contradictions, and sentences that obscure technical meaning. It should preserve author intent instead of rewriting the paper wholesale.

8. Revision checklist

The checklist consolidates findings from compliance, method, macro logic, full review, and section analysis. Duplicate findings are merged and every item receives a priority:

Items are sorted by importance and include the affected location, rationale, and a concrete action. User-authored checklist content is preserved separately from generated validation.