Spark-to-Paper Implements Full Research-Paper Generation as 13 Skills Inside an Existing Coding Assistant — No Agent Platform Required
arXiv 2608.11924 (submitted 12 Aug 2026, 87 upvotes on HuggingFace today) argues that end-to-end paper generation needs no separate orchestration service: thirteen composable skills inside a coding assistant suffice, provided you separate model judgment from deterministic checkable operations, and separate experiment *planning* from reporting so required evidence is specified before results are seen. They name a failure mode worth knowing — the Self-Refutation Loop, where repeated experiments keep rejecting the original objective — and bound it explicitly. Reported results: 99.5% citation validity, 96.4% figure editability across eight controlled topics, and an ablation showing fabrication detection rising from 14% for a single-pass draft to 92% with the full integrity and review stack.
↳ Follow the thread