Two New Systems Push AI Toward Autonomous Science Writing and Long-Form Story Memory

Three cents shy of a third of a dollar: that's what it now costs to produce a complete, submission-ready scientific manuscript using Prompt-to-Paper, a multi-agent framework detailed in a paper posted to arXiv on July 5, 2026. The system doesn't just draft text. It runs real computational biology experiments through an autonomous coding agent, replacing the placeholder numbers and synthetic outputs that have plagued earlier attempts at AI-generated research papers with genuine, executed results.

The architecture leans on deterministic retrieval-augmented generation, pulling from a working corpus of 60 to 100 papers per project. A section-aware relevance scorer decides which sources matter where, while a snowball citation expansion method lets the system chase down related work the way a human researcher would follow a citation trail. To judge whether a draft is actually good, Prompt-to-Paper uses an eight-dimensional automated quality scorer, calibrated against approximate reference statistics pulled from published literature and fitted with explicit penalties for hallucinated claims.

That scorer feeds a quality-driven improvement loop, which routes each revision cycle to one of three possible researcher actions and triggers a deeper research pass every ten iterations. Across five bioinformatics case studies used to validate the system, all five runs produced fully formatted submission PDFs with zero citations pointing outside the source corpus. The improvement loop lifted manuscript quality by an average of 17.96 points on a 0-100 scale, with the best case gaining 26.04 points. A human reviewer who evaluated the five finished manuscripts scored them at an average of 7.0 out of 10, a respectable mark for fully machine-generated scientific writing, even if it stops short of publication-ready polish.

A Parallel Approach to Long-Form Memory

A day later, on July 6, a separate arXiv submission tackled a different but related problem: how AI systems track the internal logic of long, evolving narratives. The paper, "Narrative World Model: Narratology-Grounded Writer Memory for Long-Form Fiction," introduces NWM, a system built to answer the kind of multi-hop questions that trip up most retrieval setups. Who knows a particular secret, and when did they find out? Did an event actually happen before it was narrated to the reader? Did an earlier setup pay off later, and how did a relationship shift across the story's arc?

NWM answers these by combining a typed temporal-state graph, grounded in narratology theory, with a query-conditioned hybrid retrieval mechanism. The researchers used Opus 4.8 as a fixed reader model across all evaluations, comparing NWM against Graphiti/Zep, a temporal-knowledge-graph memory framework from Rasmussen et al. (2025), as well as against GraphRAG and flat retrieval baselines. Across both test corpora, NWM came out significantly ahead on multi-hop narratological question-answering, outperforming every baseline tested.

The 23-page paper, with 9 pages of main text and the rest in appendix material plus four figures, is filed under AI, computational linguistics, and information retrieval categories on arXiv. Together, the two papers point to a broader pattern: retrieval-augmented systems are being adapted well beyond chatbot Q&A, into domains as specific as autonomous scientific experimentation and the structural memory demands of long-form fiction.