Prompt-to-Paper: An Agentic System That Writes (and Runs) Bioinformatics Research
Source: arXiv:2607.05456 — Kamran, Amjad, Mustansar, Shaukat, Sherbaz, Khan (NUST SINES, Islamabad) — submitted July 5, 2026
Summary
The paper targets a specific credibility problem with AI-generated scientific manuscripts: claims aren't grounded in verifiable literature, reported experimental results are often fabricated rather than actually run, and there's no standard way to score whether an AI-written paper meets real publication-quality rigor. Their system, built on a multi-agent pipeline called Research Landscape Explorer v4 (RLEv4), turns a research topic into a submission-formatted manuscript by combining a deterministic retrieval-augmented pipeline (grounding every claim against a corpus of 60–100 real papers via section-aware relevance scoring and snowball citation expansion) with an autonomous coding agent that executes real computational biology experiments instead of inventing plausible-looking numbers.
Key Takeaways
- Grounding is deterministic, not vibes-based: claims are matched against an actual retrieved corpus (60–100 papers) using section-aware relevance scoring plus snowball citation expansion (following citation chains outward), rather than relying on the model's parametric memory for citations.
- Experiments are actually executed: a coding agent runs real computational biology analyses and feeds genuine numerical results back into the manuscript — directly targeting the well-known failure mode where LLMs fabricate plausible-but-fake results tables.
- Iterative refinement with a quality ratchet: every 10 iterations, a "deep research cycle" has a leader agent read the full draft, identify a weak point, have the coding agent strengthen the underlying analysis, and rewrite the manuscript from the stronger output — with a "never-regress guard" ensuring manuscript quality is monotonically non-decreasing across iterations.
- Relevance beyond bioinformatics: the grounding + real-execution + monotonic-refinement pattern is a general template for using agents in any domain where "looks plausible" and "is actually true" need to be forced apart — worth watching as a design pattern independent of the bio use case.