Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring Generative Process Reward Modeling for Semi-Structured Data: A Case Study of Table Question Answering

Lei Tang, Wei Zhou, Mohsen Mesgar

TL;DR

This work evaluates generative PRMs for table question answering (TQA), highlighting challenges like irrelevant table content and weak step dependencies. It compares GenPRMs with textual and code verification, Multi-Dimensional Analysis, and Rephrase–React–Analysis against baselines across WTQ, TabFact, CRT, and SciTab. Findings show code verification improves both answer accuracy and step verification, but generalization to out-of-domain data remains a major hurdle, and better alignment between step verification and final answers is needed. The study provides actionable insights for building robust, process-aware verifiers in semi-structured reasoning tasks and suggests directions for stronger process-level evaluation.

Abstract

Process reward models (PRMs) improve complex reasoning in large language models (LLMs) by grading candidate solutions step-by-step and selecting answers via aggregated step scores. While effective in domains such as mathematics, their applicability to tasks involving semi-structured data, like table question answering (TQA) remains unexplored. TQA poses unique challenges for PRMs, including abundant irrelevant information, loosely connected reasoning steps, and domain-specific reasoning. This work presents the first systematic study of PRMs for TQA. We evaluate state-of-the-art generative PRMs on TQA from both answer and step perspectives. Results show that PRMs that combine textual and code verification can aid solution selection but struggle to generalize to out-of-domain data. Analysis reveals a weak correlation between performance in step-level verification and answer accuracy, possibly stemming from weak step dependencies and loose causal links. Our findings highlight limitations of current PRMs on TQA and offer valuable insights for building more robust, process-aware verifiers.

Exploring Generative Process Reward Modeling for Semi-Structured Data: A Case Study of Table Question Answering

TL;DR

This work evaluates generative PRMs for table question answering (TQA), highlighting challenges like irrelevant table content and weak step dependencies. It compares GenPRMs with textual and code verification, Multi-Dimensional Analysis, and Rephrase–React–Analysis against baselines across WTQ, TabFact, CRT, and SciTab. Findings show code verification improves both answer accuracy and step verification, but generalization to out-of-domain data remains a major hurdle, and better alignment between step verification and final answers is needed. The study provides actionable insights for building robust, process-aware verifiers in semi-structured reasoning tasks and suggests directions for stronger process-level evaluation.

Abstract

Process reward models (PRMs) improve complex reasoning in large language models (LLMs) by grading candidate solutions step-by-step and selecting answers via aggregated step scores. While effective in domains such as mathematics, their applicability to tasks involving semi-structured data, like table question answering (TQA) remains unexplored. TQA poses unique challenges for PRMs, including abundant irrelevant information, loosely connected reasoning steps, and domain-specific reasoning. This work presents the first systematic study of PRMs for TQA. We evaluate state-of-the-art generative PRMs on TQA from both answer and step perspectives. Results show that PRMs that combine textual and code verification can aid solution selection but struggle to generalize to out-of-domain data. Analysis reveals a weak correlation between performance in step-level verification and answer accuracy, possibly stemming from weak step dependencies and loose causal links. Our findings highlight limitations of current PRMs on TQA and offer valuable insights for building more robust, process-aware verifiers.
Paper Structure (22 sections, 13 figures, 6 tables)

This paper contains 22 sections, 13 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Examining PRMs on TQA.
  3. Problem Formulation.
  4. Examined PRMs.
  5. Training PRMs for TQA.
  6. Baselines and Evaluation.
  7. Results and Analysis
  8. GenPRMs achieve promising performance compared to other methods, while falling short of oracle for in-domain settings.
  9. Incorporating code verifier improves PRM performance in both answer and step correctness.
  10. Fine-tuned PRMs with code verification are not robust on out-of-domain datasets.
  11. Process-level evaluation highlights the competitiveness of GenPRM.
  12. Weak correlation between PRM verification and answer correctness in TQA.
  13. Conclusions
  14. Appendix
  15. Implementation and Adaptation Details
  16. ...and 7 more sections

Figures (13)

  • Figure 1: Comparing the verification process for a math (left) and a TQA problem (right). Examples draw from zhao2025genprmscalingtesttimecompute and zhou2025efficientmultiagentcollaborationtool. Compared to math, not all information in TQA is relevant.
  • Figure 2: EM measured by Best@N across varying numbers of paths. Left: WTQ. Right: Tabfact. Oracle always selects correct answers if they exist. Majority selects the most frequent predictions as answers.
  • Figure 3: Counts per confidence bin (bars, left axis) and Exact Match (line, right axis). Bins are PRM score intervals: $[0,0.2),[0.2,0.4),[0.4,0.6),[0.6,0.8),[0.8,1.0]$. It shows that average step scores are not strongly correlated with answer correctness.
  • Figure 4: A chain with many locally correct steps but an incorrect final answer, inflating the average step score. Labels here are human-annotated.
  • Figure 5: An example of a chain with local errors that does not affect the final answer; in this case, the chain’s score is over-penalized. Labels here are human-annotated.
  • ...and 8 more figures