Enhancing TableQA through Verifiable Reasoning Trace Reward

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 .

💡 Research Summary

Table Question Answering (TableQA) differs fundamentally from text‑ or image‑based QA because the answer cannot be extracted from a static input; instead, an agent must iteratively transform the table state through a series of actions. This multi‑step reasoning introduces a partially observable decision problem: the model only sees a snapshot of the table at each turn, and errors in early actions can cascade, leading to completely wrong answers. Existing approaches either rely on internal LLM signals, post‑hoc graph refinements, or test‑time scaling, all of which are fragile when the reasoning depth increases.

The paper proposes RE‑Tab, a plug‑and‑play framework that treats TableQA as a Partially Observable Markov Decision Process (POMDP) and equips the agent with explicit, verifiable rewards at two critical junctures: (1) State Transition – after each atomic table operation, and (2) Simulative Reasoning – when selecting among multiple candidate reasoning trajectories. The core contribution is a lightweight, training‑free reward metric called TAB‑ROUGE, which adapts the classic ROUGE evaluation to tabular data. TAB‑ROUGE measures (i) lexical coverage of query‑relevant tokens in the current table, (ii) precision by penalizing superfluous rows/columns, and (iii) structural integrity to ensure schema consistency. Unlike embedding‑based similarity scores (e.g., CLIPScore) that ignore table layout, TAB‑ROUGE directly reflects whether an intermediate table contains the exact information needed to answer the query.

During State Transition, the agent receives a scalar reward computed by TAB‑ROUGE for the newly generated table. This immediate feedback tells the agent “Is this action moving me toward the answer?” and allows early pruning of unpromising paths. In Simulative Reasoning, the agent generates several possible action sequences (trajectories). Each trajectory’s cumulative reward R(τ)=∑γ^l r_l (with discount γ) is evaluated, and the trajectory with the highest R(τ) is selected for final answer generation. This two‑phase search dramatically reduces the exploration space and aligns the agent’s policy with a verifiable progress signal.

The authors provide a theoretical justification (Proposition 3.1) showing that the reward reduces the conditional entropy of the true table state given the observation and reward, i.e., H(T_{t+1}|o_{t+1},r_t) < H(T_{t+1}|o_{t+1}), thereby making the decision policy more informative under partial observability. Because TAB‑ROUGE is rule‑based, it requires no labeled data and can be applied to any new schema without retraining.

Empirically, RE‑Tab is evaluated on multiple state‑of‑the‑art TableQA benchmarks (WikiTableQuestions, TabFact, WikiSQL, etc.) and across a range of LLM backbones (GPT‑4, LLaMA‑2‑70B, Claude‑2, etc.). The results are striking: average QA accuracy improves by up to 41.77 % relative to strong baselines, while the number of generated tokens during inference drops by 33.33 %, translating to roughly a 25 % reduction in overall inference cost. The gains are especially pronounced on tables with complex cell content (e.g., box‑score text), where TAB‑ROUGE’s structural awareness prevents the loss of critical information that would otherwise mislead the model.

In summary, RE‑Tab demonstrates that (1) explicit, verifiable stepwise rewards can stabilize multi‑turn table reasoning, (2) a training‑free, schema‑aware reward metric can be seamlessly integrated into existing pipelines, and (3) such a design yields both accuracy and efficiency improvements across diverse models and datasets. Future work may explore extending TAB‑ROUGE to support more sophisticated table operations (joins, pivots), learning adaptive reward functions that incorporate human feedback, and applying the same reward‑enhanced paradigm to other structured‑data tasks such as knowledge‑graph querying or spreadsheet automation.