problem definition

Input: an Area Corpus , finalized Survey Outline , Section Matches , Evidence Roles , and Summarized Corpus

Output: an evidence-grounded latex survey and compiled PDF that follow , preserve its hierarchy, support claims with section-assigned papers, and include reviewed figures and tables

current system

evidence preparation

Goal: give each outline section the evidence needed to fulfill its responsibility within the survey.

  • Build one Section Evidence Packet for each leaf section in .
  • Validate that the outline, matches, roles, summaries, and Area Corpus refer to compatible source artifacts.
  • Keep evidence assignments connected to their intended sections.

drafting and assembly

Goal: write and assemble the complete evidence-grounded survey while preserving the outline.

  • Write each top-level Body Branch in parallel from its outline subtree and assigned evidence.
  • Preserve every outline node, its order, and its nesting when combining the branches.
  • Write the introduction and conclusion in separate calls after the body is complete.
  • Generate the title and abstract from the completed draft.
  • Assemble the title, abstract, introduction, body, conclusion, and bibliography into one native-latex document.

latex repair

Goal: produce a cleanly compiling base document before adding visuals.

  • Give the assembled document to a dedicated latex repairer agent.
  • It compiles and inspects the full document, then patches syntax and formatting errors without changing its meaning or structure.
  • It repeats this process until the document compiles successfully.

visuals

Goal: research, review, build, and integrate useful evidence-grounded figures and tables.

  • Give the repaired survey and its evidence to a visual-program agent.
  • It proposes and researches figures and tables, then selects the visuals that add useful information to the survey.
  • It delegates evidence review and visual construction to specialized subagents, revising or removing unsupported visuals.
  • It inspects the completed visuals, integrates them into the survey, and compiles the visualized document.

document revision and final validation

Goal: improve whole-document consistency and readability while preserving the grounded content and reviewed visuals.

  • Give the complete visualized survey to a dedicated document revision agent.
  • It improves global consistency, readability, formatting, and cross-section coherence without changing the grounded content, structure, citations, or visuals.
  • Compile the revised document and verify that:
    • the latex build succeeds
    • every accepted visual remains unchanged and appears exactly once
    • citation and outline structure remain valid

current limitations

  • Surveys remain too long; length control must preserve the outline rather than merge or drop sections.
  • Parallel Body Branch writers cannot see one another, so the initial draft can contain inconsistent terminology, notation, repeated explanations, and weak cross-section transitions.
  • Final revision can repair presentation and consistency but cannot add new factual synthesis or correct evidence without returning to an evidence-owning stage.
  • High-value figures require substantial cross-paper research and synthesis. Earlier runs produced only a few simple visuals, motivating the new mandatory research step for every visual candidate.
  • Visual grounding review checks whether frozen claims are supported, but it does not by itself guarantee that a figure is insightful or that the overall survey is scientifically strong.