A paper that writes papers, for $8.10 and 3.2 hours
Top of HuggingFace papers today with 176 upvotes: Spark-to-Paper, an end-to-end research paper generator built as thirteen composable skills inside a coding assistant. Literature retrieval, experiment design, evidence-based claim revision, figure generation, the whole pipeline. Average cost per manuscript is $8.10, average time 3.2 hours, average token spend 11.9 million.
Ignore the headline capability for a second, because the architecture is the actually transferable idea. The system separates model-based judgment from deterministic operations, and separates experiment planning from result reporting. Everywhere something can be checked mechanically it is, and the model only gets to decide the things that genuinely require judgment. The results follow directly from that discipline: 99.5% citation validity, 96.4% figure editability, and fabrication detection going from 14% in a single pass to 92% with the full system. That last number is a 6.5x improvement from structure alone, no better model involved.
The most honest thing in the paper is a named failure mode they call the Self-Refutation Loop, where repeated experiments end up rejecting the original research objective and the system keeps going anyway. Anyone who has run a long agent loop recognizes it. The agent optimizes the process it was given while the goal quietly dies underneath, and nothing in the loop is checking whether the destination still exists. Naming it is genuinely useful.
The thirteen-skills framing is the other thing to take away. This isn't a bespoke research agent, it's a skill library aimed at a domain, running in a general coding assistant. Which lines up with where the whole skills ecosystem has been heading all month, and quietly makes the case that domain expertise is now something you write down as files rather than train into weights.
No public code repo. arXiv 2608.11924, authors Zhuoyang Qian and colleagues.
← Back to all articles
Ignore the headline capability for a second, because the architecture is the actually transferable idea. The system separates model-based judgment from deterministic operations, and separates experiment planning from result reporting. Everywhere something can be checked mechanically it is, and the model only gets to decide the things that genuinely require judgment. The results follow directly from that discipline: 99.5% citation validity, 96.4% figure editability, and fabrication detection going from 14% in a single pass to 92% with the full system. That last number is a 6.5x improvement from structure alone, no better model involved.
The most honest thing in the paper is a named failure mode they call the Self-Refutation Loop, where repeated experiments end up rejecting the original research objective and the system keeps going anyway. Anyone who has run a long agent loop recognizes it. The agent optimizes the process it was given while the goal quietly dies underneath, and nothing in the loop is checking whether the destination still exists. Naming it is genuinely useful.
The thirteen-skills framing is the other thing to take away. This isn't a bespoke research agent, it's a skill library aimed at a domain, running in a general coding assistant. Which lines up with where the whole skills ecosystem has been heading all month, and quietly makes the case that domain expertise is now something you write down as files rather than train into weights.
No public code repo. arXiv 2608.11924, authors Zhuoyang Qian and colleagues.
Comments