Spark-to-Paper: Full Drafts With Real Experiments for $8.1, 99.5% Citation Validity
jiqizhixin · x · 2026-08-26
Spark-to-Paper breaks the research pipeline into 13 composable skills that reuse a coding agent's file-reading, code-execution, and web-search capabilities. It runs experiments, generates editable vector figures, and applies an adversarial review loop to catch hallucinated conclusions.
Key numbers:
- 99.5% citation validity, beating human preprints (97.8%) and AI Scientist (93%)
- 96.4% of figures are editable (humans: 58%, prior agents: 0%)
- Fake-conclusion detector catches 92% of injected errors, up from 14% for single-pass generation
- Average cost $8.1 per paper, 3.2 hours to produce
More from coding & agent
- 'Agent Experience' emerges as a new UX theme: some of your users are agents now — YvesMulkers · 2026-09-21
- Anacondahood launches: conda research + Robinhood MCP + onchain rails for AI trading agents — teoliphant · 2026-09-20
- rakazo: self-host AI bots with persistent Linux desktops via Docker — socialwithaayan · 2026-09-20
- Camera-only autonomous drone clears obstacle course with Jev judgment model, built in 15 minutes for $0.10 — alex_verem · 2026-09-20
- Spotify Turns Internal Coding-Tool Chaos Into Xirp, a Mac App for 50+ Parallel AI Sessions — shashib · 2026-09-20
- Jev launches and users are already using it to book flights, trade, and play Mario — alex_verem · 2026-09-20