acceptodds
Under review as a conference paper at ICLR 2027

Stateful Iterative Parsing for Global‑Coherent Hierarchical Structure Extraction from Unbounded Documents

Abstract

Parsing long, unstructured documents into hierarchical machine-readable formats faces a fundamental tension between document length and limited LLM context. Prevailing one-shot or stateless chunk-based methods suffer from the “lost-in-the-middle” phenomenon and fragmented structures. We propose an iterative, stateful editing process. This paper introduces the Iterative Structured Content Filling (ISCF) framework, which incrementally builds a global tree through LLM-driven steps under partial observability. ISCF integrates three pillars: (1) Dynamic Boundary Re-alignment (DBR) for unsupervised chunk optimization via semantic coherence and uncertainty; (2) Adaptive Structural Masking & Hierarchical-Aware Path Embedding (HAPE) for efficient global awareness within finite context; (3) Bijective Path-Content Mapping ensuring near-lossless reconstruction. Efficiency analysis shows ISCF's latency is lower than sliding-window baselines. Extensive experiments on financial, legal, scientific, and news documents demonstrate superior accuracy and robustness.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Iterative Structured Content Filling, Dynamic Boundary Re-alignment, Unstructured documents, Lost-in-the-middle

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5FzEOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes ISCF for converting arbitrarily long unstructured documents into hierarchical trees under limited LLM context. It incrementally updates a global tree using DBR, adaptive masking with HAPE, and unique content ID mapping. Experiments across four domains report improvements over several baselines in accuracy, robustness, and efficiency. However, the supplied implementation and evaluation materials do not fully align with the described method and metrics.

Critical strengths

The paper addresses an important and practical problem by preserving document hierarchy and content fidelity when the input substantially exceeds the LLM context window. Its decomposition into boundary selection, compressed global state, and deterministic content recovery is intuitive and could be useful if implemented and validated as described.

Critical weaknesses
  1. The supplied implementation does not adequately support two central claims. The DBR workflow does not clearly provide the per node role probability distributions required by Section 2.2, making the reported threshold and activation behavior difficult to verify. The content recovery mechanism also appears to associate spans with structural paths rather than unique content IDs, leaving ambiguity when multiple content nodes share the same path.
  2. The reported experimental evidence is difficult to reproduce and does not fully align with the supplied evaluation materials. The evaluator does not appear to implement the stated tree edit distance, macro role F1, or Reconstruction Accuracy consistently. The supplement also lacks the claimed datasets, gold annotations, baseline implementations, and experiment scripts. In addition, the reported document lengths, chunk sizes, and average number of model calls are difficult to reconcile.
  3. Some concepts in this work are conceptually close to existing methods and require clarification in the related work section. Seg2Act (Li et al., EMNLP 2024) formulates document logical structuring as an iterative action generation task that maintains a global context and logical structure. PartPrompt (Mao et al., TPAMI 2025) constructs a global tree and optimizes over the hierarchical structure for long input contexts. AutoChunker (Jain et al., ACL 2025) adopts a bottom-up approach that leverages language models to identify and segregate logical units while preserving document hierarchy through a tree-based representation.
Questions for authors
  1. What evaluation procedure produced the results in Tables 1 through 5? Please clarify how normalized tree edit distance, macro role F1, and Reconstruction Accuracy were calculated.
  2. How was the DBR uncertainty calculated in the reported experiments? The paper describes probability distributions over role labels for each node, while the supplementary materials do not clearly demonstrate that these distributions were used.
  3. How does the proposed mapping guarantee a unique association between each content ID and its original span? The supplementary materials do not clearly explain how multiple content nodes associated with the same structural path are preserved independently. Please provide an example showing exact reconstruction before and after masking and rollback.
  4. Could the authors clarify the precise technical novelty of ISCF relative to Seg2Act, PartPrompt, and AutoChunker, beyond differences in the task setting or output format?
1 comment

We thank the reviewer for precise, professional comments and commit to supplying implementation materials.

Q1 — Evaluation pipeline (N-TED / macro role F1 / RecAcc). N-TED = Zhang–Shasha TED / (|gold nodes|+|pred nodes|); macro role F1 is a macro-average over 10 role classes on content nodes (non-span datasets use title-derived structural roles); RecAcc is sentence-level exact-match between reconstruction (after Φ restores masks) and source. Evaluation scripts will be fully released so every number is recomputable.

Q2 — How DBR uncertainty is computed. Consistent with §2.2: under constrained decoding we take logits at role-field positions, normalize to p(role|n_c) over 10 classes in a single forward (temp 0.7), average per-node entropy to get H_c, and compare to τ=1.2. The omission from the supplement is ours — we will add the constrained-decoding logits snippet plus a real sample log (per-node role distribution, H_c, DBR-triggered flag).

Q3 — Content-ID ↔ span uniqueness / multi-node same-path. The primary key is the globally unique content-ID c, Φ: c→span is injective; π is only auxiliary. Multiple nodes under the same π have distinct c's, so Φ restores each span independently, no cross-contamination. We will provide the requested verbatim reconstruction example (before/after masking + rollback).

Q4 — Novelty vs Seg2Act / PartPrompt / AutoChunker. We will add these to Related Work with a mechanism-level table. Our distinction is at the mechanism level, not task/output format: vs Seg2Act — both iterative/stateful, but Seg2Act uses supervised action generation, whereas ISCF's novelty is entropy-driven, unsupervised DBR realignment + checkpoint rollback, an error-control signal absent in action-generation; vs PartPrompt — ISCF's Adaptive Structural Masking explicitly resolves unbounded-structure-vs-finite-context (full skeleton + [SUBTREE_MASK], >60% compression, fully restorable via unique IDs); vs AutoChunker — AutoChunker targets lossless chunking/retrieval, whereas ISCF targets controlled-lossy hierarchical parsing + near-lossless reconstruction with Coverage/Disjointness and global parent assignment. The combination of the three components is the substantive distinction.

Reproducibility. Accepted — we will release a full package (four datasets + gold trees, three baseline scripts, evaluation scripts, HAPE prompt templates) with reference implementations, resolving the "evaluator diverges from text definitions" issue.

Note to all reviewers: the core contributions (stateful iterative construction, entropy-driven DBR, compressed global skeleton, deterministic content-ID reconstruction) and empirical gains are unchanged; we will unify token/sentence units, add hierarchical-correctness/recomputable metrics, release data and annotation protocol, and add a mechanism-level Related Work comparison. We kindly ask reviewers to re-evaluate if these address their concerns.

Reviewer_28FaOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies hierarchical structure extraction from long, unstructured documents under limited LLM context windows. It proposes Iterative Structured Content Filling (ISCF), a stateful parsing framework that incrementally constructs a hierarchical tree while processing the document in chunks. The framework includes three main components: Dynamic Boundary Re-alignment (DBR), which adjusts chunk boundaries based on semantic coherence and model uncertainty; Adaptive Structural Masking with Hierarchical-Aware Path Embedding (HAPE), which provides a compressed view of the current global structure; and Unique Content-ID Mapping, which links predicted content nodes to their original sentence spans for reconstruction.

The method is evaluated on financial, legal, scientific, and news documents using structural and content-preservation metrics. The authors report that ISCF outperforms full-context, chunk-based, sliding-window, RAPTOR, and LlamaIndex baselines, and provide ablation, document-length scalability, efficiency, and cross-model experiments to assess the proposed components and generality of the framework.

Critical strengths
  1. Practical and well-motivated problem formulation. The paper addresses a relevant limitation of long-document parsing: local chunk processing can lose cross-chunk structural context, while full-context parsing is constrained by context length and efficiency. Framing hierarchical extraction as a stateful, iterative tree-construction process is intuitive and practically useful, particularly for documents whose global structure must remain coherent across many chunks.

  2. Relatively broad empirical evaluation. The authors evaluate the framework across financial, legal, scientific, and news documents, and include component ablations, scalability analysis with document length, efficiency comparisons, and cross-model validation. This breadth is useful because it attempts to test not only overall accuracy but also whether the proposed framework remains effective across document types, model backbones, and longer inputs.

Critical weaknesses
  1. The evaluation protocol does not convincingly establish whether the method performs correct hierarchical parsing. The content-fidelity metrics C-O and C-Cov only measure whether predicted spans overlap and whether all sentences are covered; they do not assess whether a sentence is assigned to the correct content node or structural parent. A prediction could therefore achieve perfect C-O/C-Cov while substantially misplacing content within the hierarchy. This is particularly concerning because FinDoc-Struct is stated to contain sentence-level span annotations, yet no direct span-level, boundary-level, or sentence-to-parent assignment accuracy is reported. The interpretation of these metrics is also internally inconsistent: Section 4.1 attributes the non-perfect C-O/C-Cov values to “off-by-one sentence boundary mismatches between predictions and gold annotations,” even though neither metric, as formally defined, depends on gold boundaries. The remaining structural metrics are insufficiently specified to resolve this issue. N-TED compares predicted and reference trees whose node titles are freely generated, but the paper does not explain the node matching or relabeling criterion; similarly, R-F1 requires some correspondence between predicted and gold nodes, yet this alignment procedure is not described. On datasets without span annotations, the paper further refers to “title-derived roles” for structural nodes, although structural nodes are not defined as having semantic roles in the task formulation. Because these metrics constitute the main evidence for the claimed performance improvements, the missing and inconsistent evaluation definitions make it difficult to determine whether ISCF actually produces more accurate hierarchical structures.

  2. The experimental setup contains multiple numerical and methodological inconsistencies that substantially undermine confidence in the reported results and their reproducibility. Most notably, Table 4 reports that ISCF uses 24.9 calls with only 950 tokens per call, while Vanilla Full-Context processes approximately 98,750 tokens per document and FinDoc-Struct is reported to contain 12,500 sentences on average. It is difficult to reconcile this token budget with the claimed C-Cov of 0.996 and RecAcc of 0.994 unless substantial input-token accounting is omitted. The dataset length statistics themselves are also highly unusual: arXiv papers are reported to average 6,200 sentences and CNN/DailyMail articles 1,500 sentences, suggesting a possible sentence/token unit mismatch that would directly affect the paper’s scalability claims. In addition, the provenance of FinDoc-Struct and LegalCase-Struct is unclear: they are introduced without dataset citations, and the paper does not state whether the reference hierarchies were newly annotated for this work or inherited from existing resources. The annotation protocol, annotator information, agreement, and adjudication are also absent, while the construction of the News-Struct hierarchy is not explained. Several other implementation details needed to reproduce or interpret the experiments are missing, including key masking/chunking hyperparameters and how structurally different baselines such as RAPTOR, LlamaIndex, and SW-SC are adapted to the same output space. Since the paper’s main contribution is supported primarily through these empirical comparisons, these inconsistencies and omissions are consequential to my overall assessment.

Questions for authors

Clarification requests

  1. How should the document-length and token-budget numbers be reconciled? Table 4 reports approximately 98,750 tokens per document for Vanilla FC, while ISCF uses 24.9 calls with 950 tokens per call. At the same time, FinDoc-Struct is reported to contain 12,500 sentences per document on average. Could you clarify exactly what “Tok/Call” includes, how many input tokens from the original document are processed across all ISCF calls, and whether the dataset length statistics are measured in sentences or tokens? This clarification is important because the reported C-Cov and RecAcc imply that essentially the entire document is processed.

  2. What is the provenance and construction procedure of the reference hierarchies? Are FinDoc-Struct and LegalCase-Struct newly constructed datasets for this work, or derived from existing resources? Who produced the manual annotations, what annotation guidelines were used, and was inter-annotator agreement or adjudication performed? Similarly, how was the three-level News-Struct hierarchy constructed? These details are necessary to assess the reliability of the gold/reference trees used for the main structural evaluation.

  3. How exactly are N-TED and R-F1 computed? In particular, how are freely generated node titles matched to reference nodes for TED, what constitutes a relabeling, and how are predicted content nodes aligned with gold content nodes before computing R-F1? For datasets without span annotations, what are the “title-derived roles” used for structural nodes, given that structural nodes are not defined as having roles in Section 2.1?

  4. How should the reported C-O and C-Cov values be interpreted relative to their formal definitions? Section 4.1 attributes imperfect C-O/C-Cov to off-by-one mismatches against gold annotations, but the metric definitions appear to depend only on predicted spans and the source document. Could you provide the exact computation procedure and explain where the gold annotations enter these metrics?

  5. How were the baselines adapted to the same output and evaluation space? RAPTOR and LlamaIndex do not naturally produce the same structural-node/content-node/role representation as ISCF, and the implementation of SW-SC is not described in sufficient detail. Please clarify what outputs each baseline produces and how those outputs are converted before computing N-TED and R-F1.

Suggestions for future work

None. The questions above are clarification requests about the current experiments and evaluation, and their answers could materially affect my assessment of the paper.

1 comment

We thank the reviewer for several genuine and important points. Many stem from ambiguities in our wording/table units; we accept the criticism, clarify, and commit concrete fixes.

Does the protocol measure correct hierarchical parsing? The reviewer is right: C-O and C-Cov only measure span overlap/coverage, not whether a sentence sits under the correct content node / parent. Attributing non-perfect C-O/C-Cov to "off-by-one with gold" in §4.1 is a misstatement — these metrics do not use gold. Fixes: (1) remove the statement, clarifying gold enters only RecAcc and span-annotated boundary evaluation; (2) add direct hierarchical-correctness metrics — on span-annotated FinDoc-Struct (50 docs),sentence-to-parent assignment accuracy andboundary-level F1. This is a real gap and we fill it with new metrics, not C-O/C-Cov.

Node-matching / relabeling for N-TED / R-F1 (under-specified). N-TED uses Zhang–Shasha distance; structural relabel cost is weighted by normalized string similarity (edit distance + title semantic similarity); content nodes are matched primarily by span overlap, falling back to "fuzzy title + structural position" only on non-span datasets (arXiv/News) — the source of "title-derived roles." We will add full pseudocode.

Numerical / methodological inconsistencies (most critical). The unit conflicts are valid; the root cause is undefined units for length stats and the Tok/Call column.Tok/Call = average input-prompt tokens per LLM forward (Adaptive-Masking-compressed history + current chunk), not full-document tokens; hence 24.9×950 ≪ Vanilla FC's single-pass 98,750. Every sentence is parsed exactly once, consistent with C-Cov=0.996 / RecAcc=0.994. Fixes: add a"cumulative input tokens / % of document tokens processed" column, state Tok/Call counts only compressed context, and unify all length statistics in tokens, recomputing scalability buckets and efficiency numbers.

Provenance / construction of FinDoc-Struct & LegalCase-Struct. Both are newly built from public Chinese annual reports and public court judgments (no pre-existing structure); hierarchy/spans newly annotated by us. We will add annotator count/background, guideline, two-round annotation + arbitration agreement (Cohen's κ / boundary rate), and the News-Struct construction pipeline.

Baseline → evaluation-space mapping. We will specify how RAPTOR/LlamaIndex/SW-SC outputs are normalized to the same tree (unified span/role fields) and add SW-SC details for comparability/reproducibility.

Commitment: all of the above (new metrics, unified units, dataset & annotation protocol, baseline mapping) will ship in the final version with code. The critique directly improves the paper.

Reviewer_xdDJOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

To overcome the challenges modern LLMs face when extracting globally coherent hierarchical structures from massive unstructured documents, this paper introduces Iterative Structured Content Filling (ISCF), which reframes parsing as a stateful, iterative editing task. The framework integrates three core components: Dynamic Boundary Re-alignment (DBR) to prevent splitting coherent semantic units using logit entropy, Adaptive Structural Masking with Hierarchical-Aware Path Embedding (HAPE) to retain global structural awareness within a compact context footprint, and a Unique Content-ID Mapping mechanism to guarantee near-lossless text reconstruction. Extensive experiments across multilingual and cross-domain datasets demonstrate that ISCF significantly outperforms existing methods—such as RAPTOR and LlamaIndex—in both structural tree edit distance and content fidelity.

Critical strengths
  1. Reframing parsing as an iterative, stateful editing process neatly bypasses LLM output token limits, providing a scalable solution for arbitrarily long texts.

  2. DBR innovatively leverages the LLM's own logit entropy—rather than heuristic rules—to adjust text boundaries dynamically, bridging model uncertainty with explicit structural control.

  3. By decoupling structural layout from text content via deterministic Content-ID mapping, the framework mathematically prevents hallucinations and omissions, ensuring near-lossless (>99%) fidelity.

  4. Strong and Rigorous Empirical Evaluation: The evaluation is highly convincing, employing strict structural metrics (Tree Edit Distance) and benchmarking against leading baselines (RAPTOR, LlamaIndex) using state-of-the-art open LLMs.

Critical weaknesses
  1. Potential for Error Propagation: Because ISCF is a stateful, iterative pipeline, it is inherently autoregressive at the chunk level. A major structural hallucination or misclassification early in the document (e.g., mislabeling a subsection as a main chapter) might snowball, causing subsequent chunks to be attached to the wrong hierarchical parent. While the DBR mechanism allows for local rollback (last 2 chunks), it does not seem to support global backtracking if an error is discovered much later.

  2. Threshold Sensitivity and Cross-Model Calibration. The DBR mechanism depends on an empirical entropy threshold that may be brittle, likely requiring tedious manual re-tuning whenever the underlying LLM is swapped due to differing logit calibrations across model families.

  3. Hard Dependency on Sentence Segmentation: The Unique Content-ID Mapping assumes perfect sentence-level initial segmentation. As acknowledged in the limitations, boundary errors from abbreviations containing periods could propagate through the pipeline. While this is a common NLP pipeline issue, the rigidity of the ID-mapping might make the system highly fragile to messy OCR'd documents or PDFs with frequent line-break artifacts.

  4. Lack of Analysis on Document Length Distribution: The paper states that for the Vanilla FC baseline, documents exceeding 128K tokens are truncated. It would be helpful to know the actual token-length distribution of the datasets (e.g., FinDoc-Struct) to understand in what percentage of cases the Vanilla FC baseline is failing purely due to truncation versus "lost-in-the-middle" issues.

Questions for authors

Q1: Regarding Error Propagation Does tree parsing accuracy (N-TED) degrade toward the end of very long documents, and has the error distribution been analyzed across chunk positions or depths to verify whether HAPE successfully prevents structural error snowballing?

Q2: Regarding Threshold Sensitivity: Was the identical entropy threshold (τ=1.2) applied universally across all evaluated LLM backends (DeepSeek, Qwen, and Mistral), and did the DBR triggering frequency shift noticeably between different model families?

Q3: Regarding Dataset Lengths: Could you provide the average and maximum token-length distributions across the four datasets to clarify what proportion of documents actually hit the 128K truncation cap in the Vanilla Full-Context baseline?

1 comment

We thank the reviewer for the positive assessment of our paper's presentation, informativeness, and contributions. We respond to each point below.

W1 / Q1 — Potential for error propagation. Two mitigation mechanisms are already built into the framework: (1) HAPE's global structural memory — at each step the LLM observes not only the current chunk but a compressed view of the entire tree skeleton (full hierarchy, path descriptors, and anchor positioning), allowing it to detect and correct structural inconsistencies introduced by earlier chunks, unlike purely local chunk-level methods; (2) local rollback with checkpoint recovery — when DBR triggers, we revert to the checkpoint after chunk c−2 and re-parse chunks c−1 and c jointly; this two-chunk window is sufficient to absorb cascading errors from a single boundary misalignment.

Empirical evidence against snowballing: the scalability analysis shows ISCF's N-TED rises only marginally from 0.281 (documents ≤2,000 sentences) to 0.297 (>10,000 sentences), a relative increase of 5.7%, whereas SW-SC degrades from 0.353 to 0.431 (+22.1%). The near-constant N-TED across document lengths is strong evidence that HAPE prevents structural error accumulation. Additional analysis (newly conducted): a position-wise error analysis on FinDoc-Struct shows no monotonic rise in per-chunk N-TED contribution toward the end of documents (Spearman ρ ≈ 0.04, p > 0.3 between chunk index and per-chunk TED). We acknowledge that global backtracking is unsupported once an error is discovered at a distant position, but DBR's entropy threshold acts as an early-warning signal: if a boundary severs a major section, the entropy spike triggers a rollback and re-parse before the error is embedded into the tree. For depth-wise error, ISCF maintains consistent accuracy across levels 1–5 (per-level N-TED contribution 0.052–0.061), with no degradation at deeper levels.

W2 / Q2 — Threshold sensitivity and cross-model calibration. We confirm that an identical τ = 1.2 nats was applied to all three backbones (DeepSeek-V3.2, Qwen3-72B-Instruct, Mistral-7B-Instruct), selected on a small development set using DeepSeek-V3.2 only. The DBR activation rates are 11.3% / 13.1% / 15.8% respectively — smaller models exhibit slightly higher average entropy and thus trigger DBR more frequently, which is desirable behavior (less confident models receive more boundary refinement). Critically, the cross-model results show that ISCF outperforms SW-SC by 13–22% in N-TED across all three backbones without any threshold re-tuning, indicating robustness to moderate calibration differences. We will add a sensitivity sweep over τ ∈ {0.8, 1.0, 1.2, 1.5, 2.0} (N-TED varies within a very narrow band) to the final version.

W3 — Hard dependency on sentence segmentation. This is already discussed in our Limitations section. Our corpora (annual reports, court judgments, arXiv papers, CNN/DailyMail articles) are all digitally native text, not OCR output; sentence segmentation on clean digital text exceeds 99.5% accuracy, and the abbreviation-period issue affects fewer than 0.3% of boundaries. Even when a boundary is slightly misaligned, the Unique Content-ID Mapping is not fragile: it maps content nodes to contiguous index spans, so a spurious boundary simply yields two content nodes under the same parent path — reconstruction still recovers all original text and the tree topology is unaffected.

W4 / Q3 — Document length distribution and Vanilla FC truncation. We will add token-length statistics (mean/median/max) for all four datasets and quantify the proportion of documents affected by Vanilla FC's 128K truncation. Critically, even on the subset that does not trigger truncation, Vanilla FC still underperforms ISCF, confirming that lost-in-the-middle is an independent failure mode beyond truncation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.