RICoT paper: LLMs read page streams like humans for document segmentation, boundary F1 0.915

Prashant Khatri · hf · 2026-11-20

The RICoT paper (Recurrent Interleaved Chain-of-Thought) reframes Page Stream Segmentation (PSS)—splitting multi-page files into constituent documents—as a sequential process over the whole stream, unlike prior LLM work that only classifies boundaries over neighboring pages.

The prompting framework works like a reader: within a sliding window it reasons about each page and commits a boundary decision; across windows it carries a compact record of what was read, keeping per-call cost bounded. It also introduces FinancePSS-205, a public benchmark of 205 real-world de-identified financial page streams.

Results: RICoT matches or beats pairwise and windowed prompting with the same backbone across three open-weight LLMs; on a TABME subset with GLM-5 it reaches document F1 0.847 and boundary F1 0.915, above published zero-shot results. A 2B student distilled from a larger teacher's decisions comes within 0.05 document F1 of the strongest prompted model.

Original post →

More from Research

Research channel →