Rethinking Table Pruning in TableQA: From Sequential Revisions to Gold Trajectory-Supervised Parallel Search

TL;DR

The paper tackles the challenge of pruning tables for TableQA by addressing the unreliability of critique signals and the rigidity of sequential pruning. It introduces TabTrim, which uses gold pruning trajectories derived from intermediate sub-tables of gold SQL executions to supervise a trajectory-aware pruner and a loss-aware verifier, and employs Parallel Trajectory Search to explore multiple pruning paths. Empirical results across WikiTQ, TabFact, and TableBench show TabTrim achieving state-of-the-art performance, with TabTrim-8B attaining an average accuracy of 73.5% (79.4% on WikiTQ and 61.2% on TableBench), and demonstrated plug-and-play gains for open-source reasoners. The work provides a scalable, robust front-end pruning module that preserves answer-critical data while reducing computational noise, though it notes limitations due to computational resources at 4B/8B scales and suggests future study on larger models and broader deployments.

Abstract

Table Question Answering (TableQA) benefits significantly from table pruning, which extracts compact sub-tables by eliminating redundant cells to streamline downstream reasoning. However, existing pruning methods typically rely on sequential revisions driven by unreliable critique signals, often failing to detect the loss of answer-critical data. To address this limitation, we propose TabTrim, a novel table pruning framework which transforms table pruning from sequential revisions to gold trajectory-supervised parallel search. TabTrim derives a gold pruning trajectory using the intermediate sub-tables in the execution process of gold SQL queries, and trains a pruner and a verifier to make the step-wise pruning result align with the gold pruning trajectory. During inference, TabTrim performs parallel search to explore multiple candidate pruning trajectories and identify the optimal sub-table. Extensive experiments demonstrate that TabTrim achieves state-of-the-art performance across diverse tabular reasoning tasks: TabTrim-8B reaches 73.5% average accuracy, outperforming the strongest baseline by 3.2%, including 79.4% on WikiTQ and 61.2% on TableBench.

Figures (5)

• Figure 1: Illustration of (a) Program-based and (b) LLM-based pruning paradigms. represents table headers, represents gold cells related to answer and represents incomplete gold cells.
• Figure 2: Overview of TabTrim. TabTrim constructs gold sub-table trajectories to provide supervision for two core components: a Trajectory-supervised Pruner and a Loss-aware Verifier. During inference, TabTrim performs Parallel Trajectory Search with a step-wise generate–score–select procedure: the pruner generates candidate pruned sub-tables, the verifier assesses sub-tables with loss-aware scores, and the search retains the top candidates, finally outputting the highest-scoring sub-table for downstream reasoning.
• Figure 3: Accuracy comparison of TabTrim, Table-Critic, and TALON across different difficulty levels in the WikiTQ dataset.
• Figure 4: Scaling behavior of TabTrim-8B with increased inference-time search budget on WikiTQ and TableBench. Left: varying maximum depth $D_{\max}$ with $k=b=2$. Right: varying beam width $k$ with $D_{\max}=4$ and $b=2$.
• Figure 5: TabTrim vs. Best-of-$N$ sampling under a matched compute budget. We fix $k=b=2$ for TabTrim and set Best-of-$N$ to $N=k \cdot b=4$ and vary the maximum depth $D_{\max}$. We report accuracy on WikiTQ and TableBench.