Research map / Large language models / Reasoning and chain of thought
TaREx: Reinforcement Learning for Code-Driven Table Reasoning
Fangyu Lei, Jinxiang Meng, Yiming Huang, Shizhu He, Jun Zhao, Kang Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TaREx, a reinforcement-learning approach that trains models to solve table reasoning tasks by generating code, targeting structure comprehension, hallucination-prone text reasoning, and multi-step processing over tables.
Problem
Table reasoning is hard due to diverse hierarchical table structures, reliance on complex logical and numerical reasoning that makes text-only methods hallucinate, and the need for multi-step processing over multiple and lengthy tables.
Approach
Uses reinforcement learning to drive code-based table reasoning, so the model executes code over tables rather than reasoning purely in text. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DataFrame observation instead of raw text (ablation) | Table-RL · WikiTQ | score: +1.15 (81.00 vs. 79.03) | Improved |
| use DataFrame observation instead of raw text (ablation) | Table-RL · FinQA | score: +1.70 (69.61 vs. 67.91) | Improved |
| use DataFrame observation instead of raw text (ablation) | Table-RL · MultiHiertt | score: +1.17 (43.52 vs. 42.35) | Improved |
| add hybrid code-and-text action space (code execution) (ablation) | DF-Table-RL · TabFact | accuracy: +2.19 (90.76 vs. 88.57) | Improved |
| add hybrid code-and-text action space (code execution) (ablation) | DF-Table-RL · FinQA | score: +1.42 (78.97 vs. 77.55) | Improved |
| add multi-turn interaction (ablation) | DF-Table-Code-RL · WikiTQ | score: 84.12 vs. 81.00 | Improved |
| add multi-turn interaction (ablation) | DF-Table-Code-RL · TabFact | accuracy: 92.35 vs. 90.76 | Improved |
| add multi-turn interaction (ablation) | DF-Table-Code-RL · FinQA | score: 69.81 vs 69.61 | Improved |
| add difficulty-based data filtering (ablation) | TaREx-7B · WikiTQ | score: +1.16 (83.62 vs 82.46) | Improved |
| add difficulty-based data filtering (ablation) | TaREx-7B · MultiHiertt | score: +1.49 (47.82 vs 46.33) | Improved |
| add compact filtering of timeout and max-step trajectories (ablation) | TaREx · training dynamics | response length | Improved |
| add trajectory filtering of stagnant turns (ablation) | TaREx · training dynamics | gradient variance | Improved |
Closest papers by meaning
- TableMix: Enhancing Multimodal Table Reasoning in MLLMs from a Data-Centric Perspective (CVPR 2026 · similarity 0.86)
- Chain-of-Table: Evolving Tables in the Reasoning Chain for Table Understanding (ICLR 2024 · similarity 0.86)
- TaTToo: Tool-Grounded Thinking PRM for Test-Time Scaling in Tabular Reasoning (ICLR 2026 · similarity 0.86)
- TableBench: A Comprehensive and Complex Benchmark for Table Question Answering (AAAI 2025 · similarity 0.85)
- Policy Guided Tree Search for Enhanced LLM Reasoning (ICML 2025 · similarity 0.84)
- Multimodal Tabular Reasoning with Privileged Structured Information (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
