acceptodds
Withdrawn from ICLR 2027

Sheet-Raptor: Reconstructing Context for Spreadsheet Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spreadsheets are widely used in real-world data analysis. However, their multi-sheet organization and irregular table structures require understanding both structure and semantics within and across files, which conventional database techniques do not adequately support. Existing LLM-, RAG-, and agent-based methods struggle with lengthy and redundant spreadsheet content, fragmented context from row-level retrieval, and limited flexibility across analysis tasks. To address these challenges, we propose Sheet-Raptor, a unified system for context-critical spreadsheet analysis. First, it combines structure-aware preprocessing with layout-sensitive folding to reduce redundancy while preserving semantically meaningful table regions. Second, it integrates multi-granular spreadsheet embeddings, question rewriting, and progressive agentic retrieval to recover relevant context across files. Third, it decomposes queries into spreadsheet-specific atomic operations and dynamically coordinates semantic reasoning and programmatic execution. Experiments on four benchmarks show that Sheet-Raptor consistently outperforms state-of-the-art baselines, improving accuracy by up to 18.74%.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Spreadsheet Analysis, Data Analysis, Information Retrieval, Agent System

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_PYk1OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Sheet-Raptor, which answers questions over messy spreadsheets where the answer requires combining tables from different sheets or files.

Critical strengths

The problem is real and not much studied. Answering queries often needs data from other files, and the three failure cases mentioned in Figure 1 show this clearly.

Critical weaknesses

Sections 1–2 (Introduction)

  • W1. A lot of terms are inconsistent and symbols are reused. "Spreadsheet", "file" and "sheet" are used for different things, and several symbols are reused (ex. calligraphic F is the file collection in 2.1 but the set of folded regions in Algorithm 1, m is both the number of sheets and the number of columns in Eq. 1, etc.).
  • W2. M = (F, S, R) is not explained where it is first introduced (2.1). It becomes clear only in 3.1 that M is the table's address (file, sheet, range). Also, how header cells H are separated from data cells D is never stated.
  • W3. Figure 1 is hard to read. What do the S=0.9 / 0.85 / 0.8 labels mean, and what are the Step1–Step5 arrows? Also, the caption does not walk through the running example used in Section 2.2.

Section 3.1 (Preprocessing)

  • W4. The table representation is described only through words. There is no worked example of a serialized sheet and the table ranges the LLM predicts. Also, the snippet in Figure 2 is unreadable. Further, the paper never reports which LLM detects the tables, what prompts are used, or how accurate the detection is.
  • W5. Folding is unclear, and several parts in it are not specified properly. For example, what counts as a "repeated structure" becomes concrete only with the 36-row monthly example at the end. Similarly, W_max gets just an upper bound (< 10), and ρ_min is never given a value, why?

Section 3.2 (Retrieval)

  • W6. Eq. 2 uses L^f, L^s_i, L^t_{i,j} and L^{rc}_{i,j,k} without explaining them. Also, how are the summaries produced? LLM? prompt?
  • W7. The training data for the retriever is not clear. The paper only says "about five million task–context pairs drawn from open and enterprise QA data", but no dataset is named.
  • W8. The paragraph about question rewriting is long and hard to follow.

Section 3.3 (Agentic Analysis)

  • W9. There is no overview of the agents and operations. Except the routing agent, the other agents (QA, manipulation, visualization) are introduced only in 3.3. And the whole section reads like a list of separate facts.
  • W10. Further, claims are stated without examples. For example, "a structural edit can change the ranges needed by a subsequent edit" sounds plausible, but it is never shown.

Section 4.1 (Benchmark)

  • W11. The new benchmark is not well motivated or described. Why the existing benchmarks are not enough is covered in only a few lines, in 4.2. Who annotated it, and how? How were the context files chosen, and are the relevant tables labeled?

Section 4.2 (Main Results)

The headline gains are probably real, at least in direction. But several numbers don't add up, and the comparison setup favors the proposed system.

  • W12. The baseline numbers contradict each other. Plain Seed-1.6 scores 23.62 (LLM) and 24.62 (VLM) on ContextSheetBench QA in Table 1 and Figure 3. Appendix A.2 reports 12.23 and 14.39 for the same "vanilla" models, and Table 5 has a vanilla VLM without folding at 40.91. These cannot all be the same setup.
  • W13. Agent baselines look under-tuned. On RealHiTBench, SheetAgent, ReAcTable and TableZoomer are all far below plain Seed-1.6 (69.34), even though it is their own backbone.
  • W14. I am not sure the retrieval baselines are fair. For Vanilla RAG and DeepSearch, the retriever, chunking and context budget are not described, while Sheet-Raptor uses a retriever trained on about 5M in-domain pairs (see W7). It is also unclear how the agent baselines are given the auxiliary files on ContextSheetBench.
  • W15. The closest baselines are missing. ST-Raptor (the most similar prior system), SpreadsheetLLM and TableRAG are all discussed in related work, but none of them appears in the experiments.
Questions for authors

Check Weaknesses

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

SHEET-RAPTOR addresses cross-sheet and cross-file spreadsheet analysis by combining structure-preserving preprocessing, hierarchical context retrieval, and task-adaptive semantic and programmatic execution. A shared coordinate-aware representation connects retrieved evidence to spreadsheet operations. Experiments on four benchmarks report improvements over the compared baselines, supported by module ablations, backbone comparisons, and efficiency measurements.

Critical strengths
  1. The shared representation preserves source information and cell coordinates, connecting retrieved evidence to subsequent spreadsheet operations. This provides a coherent approach to cross-file dependencies in irregular spreadsheets.
  2. The evaluation covers four benchmarks, backbone comparisons, and efficiency measurements.
Critical weaknesses
  1. Evaluation and score aggregation need clarification. Output-specific scoring criteria are unclear, and human validation of the LLM judge is not reported. TableZoomer reports only QA on ContextSheetBench, yet its Total (10.83%) differs from QA (11.11%), contrary to the stated exclusion of unsupported tasks. Please clarify scoring rules, evaluation counts, and aggregation.
  2. Retrieval components’ contributions remain insufficiently isolated. Existing ablations and context controls do not isolate gains from hierarchical indexing, question rewriting, and progressive retrieval beyond those of the trained retriever. With the retriever, reranker, and budget limits fixed, compare key alternatives, such as flat versus hierarchical indexing or fixed rounds versus sufficiency-based stopping, and report each setting’s actual cost.
  3. Technical novelty is limited. Although the shared coordinate representation’s role is explained, comparisons with prior spreadsheet representation, retrieval, and agent methods remain high-level. Please identify the new mechanisms, adaptations of existing techniques, and research insights contributed by their integration.
Questions for authors
  1. What correctness criteria are used for QA, manipulation, and visualization, and has the LLM judge been validated against human judgments? How are total scores calculated, and what explains TableZoomer’s Total/QA discrepancy?
  2. With the retriever, reranker, and budget caps fixed, how do flat versus hierarchical indexing and fixed rounds versus sufficiency-based stopping compare in accuracy and actual cost?
  3. Which mechanisms are technically new relative to prior work, and what research insights does their integration provide beyond improved end-to-end performance?
Reviewer_scHqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents Sheet-Raptor, a framework for reasoning on spreadsheets involving multiple sheets and files. It combines structure-aware pre-processing, hierarchical context retrieval, and agent-based analysis with programmatic execution. Experiments on several benchmarks show improvements over existing methods.

Critical strengths
  • The paper addresses an important practical problem. Combining information across sheets and files is challenging, and the proposed framework provides a reasonable approach.

  • The experimental results are promising, with improvements across four benchmarks. The ablation and efficiency studies are also useful.

Critical weaknesses

The technical novelty is somewhat limited, as the framework mainly integrates existing techniques. More detailed ablations and comparisons with closely related methods, particularly ST-Raptor, would help establish the contributions.

Questions for authors
  1. Could the authors explain the differences for the baseline results between Table 1 and Appendix A.2, and how the LLM-based evaluation was validated?

  2. How much does each retrieval component contribute individually, esp. hierarchical indexing and progressive retrieval?

  3. What are the main technical differences from ST-raptor and other closely related methods?

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

Related papers

Discussion (0)

Sign in to comment.