PaperlayerView exact revision on arXiv

Source-linked AI summary

CGPT: Cluster-Guided Partial Tables with LLM-Generated Supervision for Table Retrieval

Tsung-Hsiang Chou, Chen-Jui Yu, Shui-Hsiang Hsu, Yao-Chung Fan

arXiv:2601.15849v1cs.IR

TL;DR

Table retrieval suffers when general-purpose embeddings compress structured tables into single vectors, while prior synthetic-query methods use heuristic partial-table selection and rarely fine-tune the embedder. CGPT uses K-means-guided partial tables, LLM-generated queries, and hard-negative contrastive fine-tuning, achieving consistent benchmark gains and retaining effectiveness with smaller LLMs.

  • Problem

    General-purpose embeddings remain suboptimal for table retrieval because whole-table encoding causes semantic compression, while prior methods often select partial tables heuristically and seldom use synthetic queries to train embeddings.

  • Method

    CGPT clusters table instances with K-means, samples semantically diverse partial tables, generates synthetic queries, and uses them in hard-negative contrastive fine-tuning.

  • Results

    16.54% mean R@1 gain over retrieval augmentation baselines is reported across MimoTable, OTTQA, FetaQA, and E2E-WTQ.

  • Takeaways & Limitations

    CGPT provides an effective table-retrieval framework that improves table representations through semantically guided partial-table construction and LLM-generated contrastive supervision.

  • Takeaways & Limitations

    Single-vector table representations remain constrained by excessive semantic compression, especially for large tables.

Abstract

from arXiv · show

General-purpose embedding models have demonstrated strong performance in text retrieval but remain suboptimal for table retrieval, where highly structured content leads to semantic compression and query-table mismatch. Recent LLM-based retrieval augmentation methods mitigate this issue by generating synthetic queries, yet they often rely on heuristic partial-table selection and seldom leverage these synthetic queries as supervision to improve the embedding model. We introduce CGPT, a training framework that enhances table retrieval through LLM-generated supervision. CGPT constructs semantically diverse partial tables by clustering table instances using K-means and sampling across clusters to broaden semantic coverage. An LLM then generates synthetic queries for these partial tables, which are used in hard-negative contrastive fine-tuning to refine the embedding model. Experiments across four public benchmarks (MimoTable, OTTQA, FetaQA, and E2E-WTQ) show that CGPT consistently outperforms retrieval baselines, including QGpT, with an average R@1 improvement of 16.54 percent. In a unified multi-domain corpus setting, CGPT further demonstrates strong cross-domain generalization and remains effective even when using smaller LLMs for synthetic query generation. These results indicate that semantically guided partial-table construction, combined with contrastive training from LLM-generated supervision, provides an effective and scalable paradigm for large-scale table retrieval. Our code is available at https://github.com/yumeow0122/CGPT.

1 Introduction

CGPT addresses semantic compression and heuristic partial-table selection in table retrieval by combining cluster-guided table construction with LLM-generated supervision. Across four benchmarks, it consistently improves retrieval over augmentation baselines, with smaller LLMs retaining effectiveness.

  • General-purpose embedding models struggle with table retrieval because encoding an entire table into one vector causes semantic compression.
  • QGpT generates synthetic queries from the first k table rows, but this selection may miss relevant information in later rows.
  • CGPT uses K-means clustering and cross-cluster sampling to build partial tables with broader semantic coverage.
  • CGPT uses synthetic queries as supervision in hard-negative contrastive fine-tuning to improve retrieval when relevant information is sparsely distributed across rows.
  • 16.54% mean R@1 gain is achieved over retrieval augmentation baselines across MimoTable, OTTQA, FetaQA, and E2E-WTQ.
  • CGPT remains effective with smaller LLMs without sacrificing retrieval effectiveness.

2 Related Work

Prior table-retrieval methods use table-aware encoders or synthetic queries, but representing whole tables with single vectors can compress semantics. CGPT instead constructs semantically diverse partial tables through K-means clustering and sampling across clusters.

  • TAPAS introduces table-aware positional encodings, while DTR separately encodes queries and tables for dense retrieval.
  • Single-vector representations can cause excessive semantic compression for large tables.
  • QGpT selects the first 10 rows and generates synthetic queries to create focused table representations for zero-shot retrieval.
  • CGPT clusters table instances with K-means and samples across clusters to form partial tables that more comprehensively cover the original table’s semantic space.

3 Method

CGPT constructs semantically representative partial tables through adaptive K-means clustering and sampling, generates synthetic queries for them, and fine-tunes embeddings with hard negatives. The training objective contrasts each query’s positive partial table against similar incorrect tables.

  • CGPT uses a four-stage pipeline: clustering-based partial-table generation, synthetic query generation, hard-negative sampling, and contrastive fine-tuning.
  • K-means produces multiple K-means Partial Tables by embedding instances, clustering them, and sampling instances from each cluster.
  • The procedure yields k partial tables per original table, with each KPT reflecting the semantics of a specific cluster.
  • An LLM generates synthetic queries for each KPT using queries conditioned on the partial table’s content.
  • Hard-negative sampling selects the top-h most similar incorrect KPTs from other tables for each synthetic query.
  • The InfoNCE objective trains on a synthetic query, its positive KPT, and selected hard negatives using cosine similarity and a temperature parameter.

4 Experiment

CGPT is evaluated across four benchmarks, ablations, cross-strategy transfer, unified multi-domain retrieval, language-model choices, and instance-sampling strategies. Results consistently favor cluster-guided construction, with hard-negative fine-tuning especially beneficial for top-1 precision.

  • CGPT is evaluated on MimoTable, OTTQA, FetaQA, and E2E-WTQ using retrieval comparisons and ablations.
  • Ablation Results: 52.74% and 57.14% R@1 are achieved by CGPT w/o FT on two compared subsets, improving over QGpT by 2.14 and 6.48 points.The result isolates the contribution of K-means-based partial-table construction without fine-tuning.
  • Ablation Results: CGPT w/o HNS attains the best R@5 and R@10 scores on MimoTable (CH), while full CGPT provides the strongest R@1 performance.The comparison indicates a trade-off between simplified negative sampling and hard-negative sampling across retrieval depths.
  • Main Results: 60.13% R@1 on MimoTable (EN) surpasses QGpT by 9.47 points, with similar improvements reported on OTTQA and E2E-WTQ.Hard negatives may slightly reduce top-5/10 performance but improve top-1 precision through more informative contrastive signals.
  • Cross-Strategy Transfer: 59.28% R@1 on MimoTable (EN) and 53.54% on MimoTable (CH) show transfer from CGPT training to the original QGpT dataset.These results assess transfer across differing partial-table construction schemes.
  • Unified Multi-Domain Evaluation: 57.20% R@1 and 78.18% R@5 are achieved on MimoTable (CH) when CGPT is combined with KPT in the unified multi-domain setting.CGPT also reaches 57.79% R@1 on MimoTable (EN), an 18.33-point increase over BGE-M3, and 55.03% on CH, improving by 16.49 points.
  • LLM Robustness: R@1 varies by only 0.6 percentage points across Llama-3.1-8B-Instruct, GPT-OSS-20B, and Qwen3-4B.The comparison evaluates CGPT’s retrieval performance on MimoTable (EN) under different synthetic-query generators.
  • Instance Sampling: 51.62% (CH) and 57.51% (EN) R@1 from CB Selection fall 5.18 and 2.62 points below CGPT, while S Selection reaches 57.03% on English.The results support preserving semantic variation through random intra-cluster sampling across languages.

5 Conclusion

CGPT addresses limited semantic coverage in zero-shot table retrieval by combining K-means-informed partial tables with hard-negative contrastive fine-tuning. Across multiple benchmarks, it consistently surpasses QGpT, achieving an average R@1 gain of 16.54%.

  • CGPT combines K-means clustering with hard-negative contrastive fine-tuning to improve table representations.
  • 16.54% average R@1 gain is reported over QGpT across multiple benchmarks.

A Prompts

The appendix specifies the prompt used to generate diverse synthetic questions for each table chunk. It requires content-grounded, answerable questions in JSON format.

  • The prompt asks an LLM to generate diverse questions that retrieve a specific table chunk.
  • Requested question types include entity-specific, temporal, comparison or ranking, aggregation, and complex reasoning queries.
  • Questions should use natural language, reference actual table values when possible, and remain answerable from the displayed chunk.
  • The prompt includes a language field and requires JSON-only output with a questions array.
Loading 2601.15849v1…