The 23-Stage Pipeline

Takes a user-provided topic prompt and generates a structured research plan including target conference, research questions, and expected contributions. Emphasizes novelty and alignment with recent conference trends.

Decomposes the research topic into concrete sub-problems, defines evaluation criteria, and identifies the key technical challenges to address.

Generates targeted search queries from the research plan, selects which APIs to query (OpenAlex, Semantic Scholar, arXiv), and defines inclusion/exclusion criteria.

Executes multi-source literature search with intelligent caching, circuit breakers, and rate limiting. Deduplicates results across sources and injects seminal papers from the seed library.

Gate Stage. LLM reviews each collected paper for relevance, quality, and domain match. Cross-domain false positives are explicitly rejected. Papers below threshold are filtered out.

Extracts structured knowledge from screened papers: key contributions, methods, results, limitations, and open questions. Builds a knowledge graph for synthesis.

Clusters extracted knowledge by topic, identifies research gaps, and synthesizes trends. Produces a structured literature review summary that informs hypothesis generation.

Generates novel, testable hypotheses that address gaps not covered by existing literature. Each hypothesis includes expected outcomes, evaluation metrics, and ablation dimensions.

Gate Stage. Designs the complete experimental methodology: baselines, ablations, metrics, datasets, and statistical tests. Requires modern benchmarks and real datasets (CIFAR-10, etc.).

Generates complete Python experiment code (main.py) with dataset loading, model definition, training loop, evaluation, and results output. Includes security validation, import checking, and code review.

Estimates GPU memory requirements, training time, and compute budget. Configures Docker sandbox resource limits and timeout values based on available hardware.

Runs generated code inside an isolated Docker container with NVIDIA GPU passthrough. Captures stdout metrics, timing data, and exit codes. Pre-cached datasets available at /workspace/data.

If experiment fails or produces poor results, automatically diagnoses issues and generates refined code. Checks ablation effectiveness (>5% difference from baseline). Up to 3 refinement iterations.

Parses experiment outputs, computes statistical significance, generates comparison charts, and produces structured results summaries. Detects result contradictions and null findings.

Evaluates experiment results and decides: Proceed (results support hypothesis), Refine (re-run with improvements), or Pivot (discard hypothesis, generate new one). Max 2 pivots to prevent infinite loops.

Creates a detailed paper outline with section-by-section arguments, key claims, and figure placements. Follows conference template structure (abstract, intro, related work, method, experiments, conclusion).

Generates the complete paper in Markdown/LaTeX with structured writing rules: 150-200 word abstract, no number repetition across sections, proper citation of original papers for all discussed techniques.

Multiple LLM reviewer personas evaluate the paper: one technical reviewer, one methodology expert, and one clarity/presentation reviewer. Each provides structured feedback with scores.

Addresses reviewer comments systematically: fixes technical issues, improves writing clarity, adds missing comparisons, and strengthens the narrative. Produces a revised draft.

Gate Stage. Comprehensive quality assessment scoring the paper on novelty, soundness, significance, clarity, and reproducibility. Papers below threshold are sent back for rewriting.

Stores research findings, methodology, and results in the persistent knowledge base for future reference and cross-project knowledge transfer. Non-critical: failure doesn't abort pipeline.

Converts the paper to conference-ready LaTeX, compiles to PDF, generates BibTeX bibliography, and produces the final output package (paper.pdf, main.tex, references.bib, charts/).

Verifies every cited paper exists in real academic databases. Checks DOI via CrossRef, title via OpenAlex, arXiv ID via arXiv API, and falls back to Semantic Scholar. Non-critical: failure doesn't abort pipeline.