AI-Assisted Authoring for Transparent, Data-Driven Documents

ICLR 2026 Conference Submission17505 Authors

Keywords: Large language models (LLMs); Data provenance; Interpretability; Scholarly communication

TL;DR: An agent-based LLM framework for transforming scholarly articles into interactive, data-driven documents linked to their underlying data sources

Abstract: We introduce the idea of *transparent documents*, web-based data-driven scholarly articles which allow readers to explore the relationship to the underlying data by hovering over fragments of text. We present an agent-based LLM framework for authoring transparent documents, building on recent developments in data provenance for general-purpose programming languages. Our implementation uses Fluid, an open source functional programming language with a provenance-tracking runtime, as a target platform. Our tool consists of two LLM agents which support a human author during the creation of a transparent document. A SuggestionAgent helps identify fragments of text which could plausibly be computed from data, including numerical values selected from records or computed by aggregations like sum and mean, comparatives and superlatives like “better than” and “largest”, trend-adjectives like “growing”, and other idiomatic quantitative or semi-quantitative phrases. An InterpretationAgent, given such a fragment, then attempts to synthesise a suitable Fluid query over the data which will generate the target string. The resulting expression is spliced into the source code for an interactive web page, turning the static text fragment into an interactable data-driven element able to reveal the data that underwrites the natural language claim. We evaluate our approach on a subset of SciGen, an open source dataset consisting of tables from scientific articles and their corresponding descriptions, which we extend with hand-generated counterfactual test cases to evaluate how well machine-generated expressions generalise in the presence of changes to the data. Our results show that gpt4o is often able to synthesise compound expressions extensionally compatible with our gold solutions.

Primary Area: interpretability and explainable AI

Submission Number: 17505