Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing TableQA through Verifiable Reasoning Trace Reward

Tung Sum Thomas Kwok, Xinyu Wang, Hengzhi He, Xiaofeng Lin, Peng Lu, Liheng Ma, Chunhe Wang, Ying Nian Wu, Lei Ding, Guang Cheng

TL;DR

The paper addresses the challenge of multi-turn reasoning in TableQA and introduces RE-Tab, a plug-and-play framework that uses a verifiable state-level reward TabROUGE to steer table transformations. It frames reasoning as a Partially Observable Markov Decision Process ($POMDP$) and employs a two-phase trajectory search to maximize the state-trajectory reward $R(\tau)$, incorporating turn-level rewards $r_t$. The key contributions are the TabROUGE-based state reward, a training-free RE-Tab framework applicable to Chain-of-Tables and Tree-of-Tables, and extensive cross-dataset, cross-model validation showing improved QA accuracy and reduced inference cost. The approach enhances reliability and efficiency of TableQA in realistic settings, with potential for extension to broader tabular challenges and reinforcement learning.

Abstract

A major challenge in training TableQA agents, compared to standard text- and image-based agents, is that answers cannot be inferred from a static input but must be reasoned through stepwise transformations of the table state, introducing multi-step reasoning complexity and environmental interaction. This leads to a research question: Can explicit feedback on table transformation action improve model reasoning capability? In this work, we introduce RE-Tab, a plug-and-play framework that architecturally enhances trajectory search via lightweight, training-free reward modeling by formulating the problem as a Partially Observable Markov Decision Process. We demonstrate that providing explicit verifiable rewards during State Transition (``What is the best action?'') and Simulative Reasoning (``Am I sure about the output?'') is crucial to steer the agent's navigation in table states. By enforcing stepwise reasoning with reward feedback in table transformations, RE-Tab achieves state-of-the-art performance in TableQA with almost 25\% drop in inference cost. Furthermore, a direct plug-and-play implementation of RE-Tab brings up to 41.77% improvement in QA accuracy and 33.33% drop in test-time inference samples for consistent answer. Consistent improvement pattern across various LLMs and state-of-the-art benchmarks further confirms RE-Tab's generalisability. The repository is available at https://github.com/ThomasK1018/RE_Tab .

Enhancing TableQA through Verifiable Reasoning Trace Reward

TL;DR

The paper addresses the challenge of multi-turn reasoning in TableQA and introduces RE-Tab, a plug-and-play framework that uses a verifiable state-level reward TabROUGE to steer table transformations. It frames reasoning as a Partially Observable Markov Decision Process () and employs a two-phase trajectory search to maximize the state-trajectory reward , incorporating turn-level rewards . The key contributions are the TabROUGE-based state reward, a training-free RE-Tab framework applicable to Chain-of-Tables and Tree-of-Tables, and extensive cross-dataset, cross-model validation showing improved QA accuracy and reduced inference cost. The approach enhances reliability and efficiency of TableQA in realistic settings, with potential for extension to broader tabular challenges and reinforcement learning.

Abstract

A major challenge in training TableQA agents, compared to standard text- and image-based agents, is that answers cannot be inferred from a static input but must be reasoned through stepwise transformations of the table state, introducing multi-step reasoning complexity and environmental interaction. This leads to a research question: Can explicit feedback on table transformation action improve model reasoning capability? In this work, we introduce RE-Tab, a plug-and-play framework that architecturally enhances trajectory search via lightweight, training-free reward modeling by formulating the problem as a Partially Observable Markov Decision Process. We demonstrate that providing explicit verifiable rewards during State Transition (``What is the best action?'') and Simulative Reasoning (``Am I sure about the output?'') is crucial to steer the agent's navigation in table states. By enforcing stepwise reasoning with reward feedback in table transformations, RE-Tab achieves state-of-the-art performance in TableQA with almost 25\% drop in inference cost. Furthermore, a direct plug-and-play implementation of RE-Tab brings up to 41.77% improvement in QA accuracy and 33.33% drop in test-time inference samples for consistent answer. Consistent improvement pattern across various LLMs and state-of-the-art benchmarks further confirms RE-Tab's generalisability. The repository is available at https://github.com/ThomasK1018/RE_Tab .
Paper Structure (49 sections, 3 theorems, 35 equations, 10 figures, 8 tables, 2 algorithms)

This paper contains 49 sections, 3 theorems, 35 equations, 10 figures, 8 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. TableQA
  4. World-Model Perspectives in TableQA
  5. Build Internal Reward Models in Multi-Turn Table Reasoning
  6. Problem Formulation
  7. Tool-based Table Reasoning State Transitions
  8. Optimal Trajectory Search in Simulative Reasoning
  9. RE-Tab Reward Modeling
  10. State-Based Reward Modeling
  11. Verifiable Query2Table Precision Reward: TabROUGE
  12. Two-Phase Implementation of Reward Models
  13. Two-Phase Search
  14. Optimization Objective
  15. Experiments
  16. ...and 34 more sections

Key Result

Proposition 3.1

From Section sub: problem-formulation, let the tabular reasoning task be defined as a POMDP where $T_t \in \mathcal{T}$ represents the high-dimensional underlying table state and $o_t = \phi(T_t)$ represents the observation via a projection function $\phi: \mathcal{T} \to \mathcal{O}$. This projecti The agent's decision policy $\pi(a_t \mid o_{1:t})$ must consequently infer task progress under unc

Figures (10)

  • Figure 1: Illustration of the comparison between (a) Tabular agent without rewards (Chain-of-Tables wang2024chainoftable), (b) VLM agent with world model rewards (VAGEN wang2025vagen), and (c) Our proposed RE-Tab. Given a complex table where the score is embedded in the form of box score text, (a) retrieves correctly but lacks an explicit reward to reason correctly; (b) provides correct reasoning but falls short in accurately retrieving the table; (c) RE-Tab iteratively provides verifiable reward as feedback to the agent to evaluate transformation correctness, and conducts multiple thought processes to search for the most rewarded answer, allowing consistent improvements in both reasoning and retrieval tasks.
  • Figure 2: Action-token confidence can be high even when the reasoning trajectory drifts, whereas the state-based, verifiable reward measured by TabROUGE better reflects whether an intermediate table state contains the right information to answer the query by rewarding query-relevant lexical coverage and penalizing excessive low-relevance table content (low precision). Specifically, TabROUGE facilitates evaluation in both the step (action) and trajectory levels, enabling RE-Tab’s two-phase trajectory search.
  • Figure 3: Results on WikiTQ with GPT-3.5. We faithfully replicate both Chain-of-Tables and Tree-of-Tables and obtain similar results as reported in wang2024chainoftableji2025treeoftable. After incorporating RE-Tab, both frameworks show substantial improvementsz with notable reduction in inference cost.
  • Figure 4: The two panels plot mean and standard deviation accuracy versus the number of sampled reasoning chains under different reward weightings. Annotated values mark the chains required to reach 95% accuracy. State-based rewards from TabROUGE both accelerate convergence and reduce variance.
  • Figure 5: Comparison of reward designs on three TableQA benchmarks. VLM-based rewards improve multi-step reasoning but degrade on large-table lookup, while embedding-based rewards are unstable due to truncation-induced structural loss.
  • ...and 5 more figures

Theorems & Definitions (5)

  • Proposition 3.1: Reward Model Significance in Multi-turn LLM Interactions
  • Proposition 3.2: Metric Optimality of Table States
  • Proposition 1.1: Variance growth with interaction length
  • proof
  • proof