ZICQ
中 Log in / Sign up
Newsroom Research & Papers #RICoT #Page Stream Segmentation #Document Parsing #LLM Reasoning #FinancePSS-205

RICoT Adds Recurrent Interleaved Reasoning to Page Stream Segmentation

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published:

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:RICoT reframes page stream segmentation as sequential decision-making over an entire document stack rather than isolated pairwise boundary classification. Within a sliding window, the model reasons page by page and commits each boundary decision; across windows, it carries forward a compact state representing what has already been read. This keeps each model call bounded without requiring the complete document to remain in context. The paper also introduces FinancePSS-205, a planned public benc


Key Takeaway

RICoT is not a new foundation model. It is a recursive prompting framework for splitting multi-page files. Instead of treating segmentation as an independent boundary-classification problem over neighboring pages, RICoT turns the page stack into a sequential decision process. The model reads and reasons page by page, commits each boundary decision, and carries a compact record of previously read content into the next window.

The central engineering benefit is bounded per-call context. The model does not need to retain the complete document in every request, yet it can use a structured memory to preserve long-range document structure. The accompanying FinancePSS-205 dataset is intended to contain 205 real-world, de-identified financial page streams, providing a more production-relevant test than synthetic documents.

From Local Boundary Classification to Sequential Segmentation

Page stream segmentation must identify where one constituent document ends and another begins. Pairwise prompting compares a page with one or a few neighbors, while conventional windowed prompting processes several pages simultaneously. Both can lose long-range evidence: visually similar pages may belong to different files, while an apparently isolated continuation page may still belong to an earlier document.

RICoT treats each page as a node in a sequential decision process. Let the pages be x_i, the cross-window state be s_k, and the boundary decision for a page be b_i. When the current window advances, RICoT updates the state with the document structure, continuation cues, and decisions established so far. The model therefore performs stateful reading rather than a collection of independent page-pair classifications.

Mechanism: Interleaved Reasoning Within and Recurrence Across Windows

An RICoT call can be decomposed into four stages:

  1. Provide the model with the current sliding window, local page relationships, and the state produced by the previous window.
  2. Generate intermediate reasoning and a boundary decision page by page, committing each decision before advancing.
  3. At the end of the window, extract and update a compact structural state.
  4. Pass that state and the following pages to the next call.

If a window contains W pages, each page averages L tokens, the cross-window state occupies S tokens, and reasoning plus decisions require R output tokens, an engineering approximation is:

C_call ≈ C(W·L + S + R)

This is an analytical estimate, not a reported measurement. With non-overlapping advancement, the number of calls is approximately N/W, while total input processing remains broadly linear in the number of pages. Overlapping windows add repeated page tokens. RICoT bounds the context of each call; it does not automatically reduce total tokens, latency, or cost across the full pipeline.

Committing decisions before advancing can reduce repeated reinterpretation across windows. It also introduces error propagation: if an early boundary is wrong, the compact state may preserve and extend that mistake. Compared with placing the entire file in one long-context request, RICoT gives up some parallelism in exchange for lower peak context usage and more controlled cross-window behavior.

Engineering Trade-offs

Throughput versus consistency. A single long-context request can process pages in parallel, but it raises input-token count, KV Cache usage, and memory pressure while exposing the model to irrelevant pages. RICoT turns one large request into multiple sequential requests. Peak context is easier to control, but orchestration overhead and page-level reasoning output increase end-to-end latency.

State compression versus information loss. The cross-window state is the framework's critical component. A short summary may omit continuation evidence; a long summary increases the cost of every subsequent call. Production implementations should define explicit fields, such as current document type, observed titles, table continuity, pending boundary decisions, and confidence, rather than relying only on free-form recapulation.

Auditability and privacy. Interleaved reasoning can identify which page contributed to a mistaken decision, but sensitive page content may also enter reasoning traces, state caches, and logs. Financial-document deployments need explicit retention, access-control, and redaction policies, and long reasoning traces should not automatically be exposed to end users.

Distillation into smaller models. The paper reports that a 2B student trained on a larger teacher's decisions performs similarly with and without page-level reasoning and remains within 0.05 document F1 of the strongest prompted model on FinancePSS-205. This suggests that final boundary decisions can transfer much of the useful signal without preserving the teacher's full reasoning process. A practical architecture could use the 2B model for routine routing and escalate low-confidence or structurally inconsistent samples to a larger model. The abstract does not provide absolute scores, distillation data volume, or training cost, so the production savings remain unquantified.

Reported Results and Evaluation Boundaries

| Experimental setting | Reported result | Supported conclusion | Still unverified | |---|---|---|---| | FinancePSS-205 with three open-weight backbones | RICoT matches or exceeds pairwise and conventional windowed prompting with the same backbone | Under a controlled backbone comparison, recursive state and interleaved decisions may outperform simpler local prompting | Model identities, absolute scores, window size, and variance across runs | | TABME test subset with GLM-5 | Document F1 of 0.847 and boundary F1 of 0.915 | Boundary localization and document grouping reach comparatively high levels and exceed published zero-shot results | Baseline values, statistical significance, test-set size, and annotation agreement | | Distilled 2B student | Similar performance with and without page-level reasoning; within 0.05 document F1 of the strongest prompted model | Page-level reasoning can be distilled into a lighter decision model | Inference-token savings, training compute, and generalization across teacher-student pairs |

Same-backbone comparisons are one of the stronger aspects of the evaluation because they partially separate prompting effects from foundation-model capability. FinancePSS-205 also uses real financial streams, which are more likely than synthetic files to expose difficult layouts, scanned pages, and cross-page tables. Aggregate F1 scores, however, do not reveal whether errors come from missed splits, false splits, incorrect merges, or incorrect groupings. The gap between boundary F1 and document F1 suggests that boundary localization and complete document grouping may not be equally difficult, although the summary is insufficient to identify the cause.

Important evidence is still missing: data-split construction, boundary annotation protocol, annotator agreement, test-set size, random seeds, confidence intervals, per-page cost, and end-to-end latency. FinancePSS-205 is described as “to be released publicly,” so it should not yet be treated as a downloadable reproducibility artifact. Strong results in finance also do not guarantee transfer to legal, medical, manufacturing, or heavily OCR-corrupted archives.

Deployment Guidance

  1. Define a stable boundary taxonomy. Specify how document starts, continuation pages, blank pages, attachments, and ambiguous pages should be labeled. Preserve document identifiers instead of reducing the task to a binary boundary label alone.
  2. Build a representative validation set. Stratify files by length, layout, scan quality, cross-page tables, and multi-document concatenation. Use double annotation and report agreement so that one financial template does not masquerade as general capability.
  3. Establish strong baselines. Compare against pairwise prompting, conventional windowed prompting, full-context prompting, commercial document APIs, layout-feature models, and deterministic rules.
  4. Tune the window and state together. Search window size, overlap stride, state schema, and maximum reasoning-token budget on validation data. Reserve the test set for final evaluation to prevent benchmark-specific overfitting.
  5. Measure the full cost curve. Track input and output tokens per thousand pages, request count, p50 and p95 latency, peak memory, retry rate, and cost per thousand pages alongside F1.
  6. Use tiered inference. Let a 2B student handle first-pass routing and escalate low-confidence, state-conflicting, or table-heavy samples to a larger model. Preserve a human-review path for production failures.
  7. Verify the public artifacts. Until the code, prompt templates, state structure, and FinancePSS-205 are actually available, reproducibility claims should remain provisional, and the finance benchmark should not be applied directly to external customer data.

Conclusion

RICoT's main contribution is not a larger model but a different inference interface for long-document segmentation. Page decisions are interleaved within bounded windows, while a compact state carries selected evidence across those windows. The same-backbone comparisons and the distilled 2B student are the most persuasive results; the largest gaps are the absence of confirmed public artifacts, absolute baselines, statistical stability, and production cost measurements. If FinancePSS-205 and the complete implementation become available, RICoT could become a useful reference for document-level routing and boundary detection in long-document processing pipelines.


Source: Hugging Face Daily Papers (2026-11-17)

— END —

Tags: #RICoT #Page Stream Segmentation #Document Parsing #LLM Reasoning #FinancePSS-205

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…