Abstract
The table reasoning task, crucial for efficient data acquisition, aims to answer questions based on the given table. Recently, reasoning large language models (RLLMs) with Long Chain-of-Thought (Long CoT) significantly enhance reasoning capabilities, leading to brilliant performance on table reasoning. However, Long CoT suffers from high cost for training and exhibits low reliability due to table content hallucinations. Therefore, we propose Row-of-Thought (RoT), which performs iteratively row-wise table traversal, allowing for reasoning extension and reflection-based refinement at each traversal. Scaling reasoning length by row-wise traversal and leveraging reflection capabilities of LLMs, RoT is training-free. The sequential traversal encourages greater attention to the table, thus reducing hallucinations. Experiments show that RoT, using non-reasoning models, outperforms RLLMs by an average of 4.3%, and achieves state-of-the-art results on WikiTableQuestions and TableBench with comparable models, proving its effectiveness. Also, RoT outperforms Long CoT with fewer reasoning tokens, indicating higher efficiency.
Introduction
Table reasoning is a dense, information-heavy task essential for modern data analysis. Recently, Reasoning Large Language Models (RLLMs) have leveraged Long Chain-of-Thought (Long CoT) to achieve impressive performance. However, this approach faces two significant hurdles: high computational costs for training and low reliability due to table content hallucinations.
As reasoning chains lengthen in standard RLLMs, models often "lose sight" of the original data, leading to errors where the model misquotes or ignores the actual rows. To address these issues, we propose Row-of-Thought (RoT), a training-free prompting strategy that structures the reasoning process through iterative row-wise table traversal.
Methodology: Row-of-Thought
RoT replaces the "black box" internal monologue of Long CoT with a structured, verifiable process. It consists of two primary components:
- Iterative Row-Wise Traversal: Instead of letting the model jump to conclusions, the prompt instructs it to assess information one row at a time. By updating intermediate results after each row, the model's attention remains anchored to the tabular structure, significantly reducing hallucinations.
- Dynamic Iteration: For complex, multi-hop questions where one pass is insufficient, the model can choose to reflect on its previous pass and initiate a new traversal. This mimics the deep-thinking capabilities of RLLMs but in a structured, verifiable way.
Because RoT relies on a one-shot prompt to demonstrate this behavior, it is entirely training-free and compatible with any standard instruction-tuned LLM.
Experimental Evaluation
The effectiveness of RoT was evaluated across three diverse benchmarks: WikiTableQuestions (WikiTQ), HiTab (which features hierarchical tables), and TableBench (covering a wide variety of question types).
Key Results
- Superior Performance: RoT using standard models (like Llama-3.1-70B) outperformed RLLMs using Long CoT by an average of 4.3%.
- State-of-the-Art (SOTA): RoT achieved SOTA results on WikiTQ (78.7) and TableBench (44.8) among models of comparable scale.
- High Efficiency: RoT reduced the number of reasoning tokens by 7% to 50% compared to Long CoT baselines. This proves that structured thinking is more efficient than the rambling monologues often found in RLLMs.
Insights and Discussion
A critical finding of this research is that structure beats scale. Forcing a model to follow a row-wise structure acts as a natural "attention anchor." Error analysis revealed that 66% of Long CoT failures were due to hallucinations; RoT mitigates this by forcing the model to explicitly "read" every relevant row sequentially.
Furthermore, the choice of traversal unit matters. While row-wise traversal is optimal for flat tables, the authors found that column-wise traversal performed better on hierarchical tables (HiTab), as it better captures the relationship between nested headers. Currently, this selection is manual, but it points toward a more flexible reasoning architecture.
Conclusion
Row-of-Thought (RoT) provides a robust, efficient, and training-free alternative to Long CoT for table reasoning. By grounding the model's reasoning in the physical structure of the data, RoT achieves superior accuracy with significantly fewer tokens. Future work will explore extending RoT to multi-turn conversational contexts and automating the selection of traversal units based on table structure.
Summary updated