AI Tool Analyzes 60 Award Papers to Generate Full Conference Drafts from Abstracts

The open-source project Best-Paper-Writing-Skill addresses a gap in AI-assisted academic writing: generic prompts can mimic style but lack the fine-grained argumentation design needed for conference papers. The author analyzed 60 award-winning papers from CVPR, ICCV, and NeurIPS (2021–2025) — 22 CVPR, 12 ICCV, 26 NeurIPS — performing sentence-level annotation on abstracts and introductions.

Corpus Analysis and Extracted Patterns

The study annotated 1,929 sentences and distilled:

475 retrievable exemplars (concrete sentence-level templates)

20 writing metrics (e.g., participle-ending frequency, gap-explicitness)

8 argumentation routes mapped to paper types: method improvement, new capability, scaling laws, re-examination, mechanistic analysis, theory, dataset, and measurement.

Key statistical findings:

100% of introductions first affirm prior work concretely.

~98% then state the research gap and its cause.

~78% present an observation about the problem itself before introducing the method.

Generic AI drafts use participle-ending clauses ( ..., making ...) in ~32% of sentences; award papers only ~8% .

Agent Workflow and Output

The Skill packages these patterns for Claude Code and Codex . Two input modes are supported:

Offline info table — researcher fills contribution, task, method, results; exported as JSON.

Existing abstract only — the agent detects missing information, asks targeted follow-up questions once, then writes the full draft.

Output includes title, abstract, introduction, method, and experiments sections, pre-formatted in the CVPR LaTeX template and compilable . Missing numbers, baselines, and citations are marked

[TODO]</strong> to avoid hallucination. The Skill also plans Figure 1 and method-overview diagrams based on 44 award papers' figure conventions; result figures requiring real data are left as placeholders.</p>
<h2>Case Study: D4RT (CVPR 2026 Best Paper)</h2>
<p>Using <strong>only the original abstract</strong> — no full text, no web search — the agent generated a <strong>~6-page CVPR-formatted draft</strong> for D4RT. The same pipeline was demonstrated on O-Voxel (Best Student Paper) and SAM 3D (Honorable Mention).</p>
<h2>Self-Check and Evaluation</h2>
<p>The repository ships two check scripts:</p>
<ul>
<li>Abstract/Introduction writing-issue detector.</li>
<li>Figure caption, cross-reference, label, and placeholder-image validator.</li>
</ul>
<p>Both produce a <code>report.md</strong> summarizing core claims, claim–evidence alignment, and pending materials. In the author's self-check, a generic-prompt version triggered <strong>2 warnings</strong>; the Skill version triggered <strong>0</strong>.</p>
<h2>Multi-Model Evaluation</h2>
<p>Four generators — <strong>Claude Opus, Claude Sonnet, GPT-Sol, GPT-Astra</strong> — each produced drafts with and without the Skill. Three judges — <strong>Claude Sonnet, Claude Opus, GPT-Astra</strong> — scored blindly on a 1–5 scale. <strong>All four generators improved</strong>; the method and experiment sections rose from <strong>1.8 to 3.9</strong> on average.</p>
<h2>Access</h2>
<ul>
<li>GitHub: <code>https://github.com/gulucaptain/Best-Paper-Writing-Skill

Project page: https://gulucaptain.github.io/Best-Paper-Writing-Skill/