← Back to Paper Radar

TableMind: An Autonomous Programmatic Agent for Tool-Augmented Table Reasoning

Chuang Jiang, Mingyue Cheng, Xiaoyu Tao, Qingyang Mao, Ouyang Jie, Qi Liu
Web Search and Data Mining | 2025
TableMind trains a lightweight LLM to act as an autonomous, multi-turn programmatic agent for table reasoning. It internalizes planning, code-based tool use and reflection through a two-stage pipeline: supervised fine-tuning (SFT) on curated reasoning trajectories, then reinforcement learning (RL) with a multi-perspective reward and a new optimization objective. This replaces both single-pass reasoning over flattened tables and hand-designed training-free agent workflows.

Problem Statement

Most LLM table-reasoning methods serialize the whole table into the prompt and answer in one forward pass. That approach overflows the context on large tables, handles continuous numerical values poorly, and offers no explicit tool use or self-correction. Training-free agent workflows partly address this, but they rely on prompt engineering and large models rather than learned agentic behavior, which limits robustness and efficiency.

Key Novelty

  • A tuning-based (not prompt-only) autonomous table agent that learns a plan–act (code execution)–reflect loop over multiple turns inside a lightweight LLM.
  • A two-stage training recipe: SFT on constructed and quality-filtered multi-turn table-reasoning trajectories to bootstrap structured behavior, followed by RL to refine it.
  • A multi-perspective reward scheme plus a novel RL optimization objective aimed at precise, executable code generation for table manipulation and numerical computation.

Evaluation Highlights

  • Evaluated on diverse table-reasoning benchmarks (likely table QA, fact verification and numerical reasoning); the abstract reports consistent gains over previous baselines.
  • Results support the claim that training an autonomous agent beats both single-turn LLM reasoning and training-free workflow designs. The abstract gives no specific numbers or ablation details.

Signal Assessment

5/10 This is a solid, well-motivated application of the now-standard SFT-then-RL agentic training recipe to table reasoning, with domain-specific reward design and a modified objective. It is not a paradigm shift, since code-based table agents and RL for tool-using LLMs are established directions, but it shows that small models can internalize multi-turn tool use for structured data.

Methodology

  1. Formulate table reasoning as multi-turn interaction: the agent plans, writes code (e.g., Python/pandas) that runs against the table in a sandbox, observes the results, and reflects or revises until it answers.
  2. Build a high-quality SFT dataset by generating multi-turn reasoning-and-code trajectories (likely distilled from a stronger model) and filtering them for correctness and quality, then fine-tune a lightweight LLM to bootstrap the agentic format.
  3. Run RL on the SFT model with a multi-perspective reward (likely covering answer correctness, format compliance and code executability or quality) and a novel policy optimization objective to sharpen code generation and multi-turn decisions.
  4. Benchmark against single-turn LLM baselines, prompt-based table agents and other trained models across multiple table-reasoning datasets.

System Components

Programmatic Agent Loop

A multi-turn plan → code action → execution feedback → reflection cycle. It lets the model query and compute over tables through code rather than reading a flattened table in its context.

Code Execution Environment

A sandboxed interpreter that runs agent-generated code on the table and returns outputs or errors as observations. This gives exact numerical computation and keeps large tables out of the context window.

SFT Data Construction & Filtering

A pipeline that builds and filters high-quality multi-turn reasoning trajectories to cold-start the agent's structured reasoning and tool-use format.

Multi-Perspective Reward Scheme

A composite RL reward that scores trajectories along several dimensions (e.g., final-answer accuracy, output format, code validity) to give richer signal than outcome-only rewards.

Novel RL Optimization Objective

A modified policy-optimization objective, likely a GRPO/PPO-style variant, tailored to stabilize and improve learning of precise code generation in multi-turn agent rollouts.

Lightweight Backbone LLM

A small-to-mid-sized open model fine-tuned to internalize agentic behavior, targeting efficiency compared with large proprietary-model workflows.

Results

Metric/Benchmark Baseline This Paper Delta
Accuracy across diverse table-reasoning benchmarks Prior single-turn LLM methods TableMind (SFT+RL agent) Consistently higher (exact numbers not given in abstract)
vs. training-free agent workflows Prompt-engineered multi-step table agents Learned autonomous agent Reported improvement (qualitative)
Large-table robustness Flattened-table prompting (context overflow) Code-based table access Qualitatively better scalability
Numerical precision Implicit numerical reasoning in text Executed code computation Qualitatively more precise
Contribution of RL stage SFT-only model SFT + RL with multi-perspective reward Implied gain (ablation details not in abstract)

Key Takeaways

  • For structured-data tasks, offloading computation to executed code instead of serializing tables into the prompt sidesteps context limits and LLMs' weak arithmetic. This is a pattern worth adopting for spreadsheet, database and analytics agents.
  • The SFT-cold-start → RL recipe, with filtered trajectories and composite rewards that cover format, executability and correctness, transfers well to domain-specific tool-using agents and lets small models compete with large prompted systems.
  • Reward design matters for code-generating agents. Multi-perspective rewards give denser and more targeted signal than answer-only rewards, which can improve stability and code quality during multi-turn RL.

Abstract

Table reasoning requires models to jointly perform comprehensive semantic understanding and precise numerical operations. Although recent large language model (LLM)-based methods have achieved promising results, most of them still rely on a single-turn reasoning paradigm that processes flattened tables in a single forward pass. This paradigm suffers from inherent limitations, including context overflow on large tables, weak sensitivity to continuous numerical values, and the absence of explicit tool-use and reflection. In this paper, we propose TableMind, a tuning-based autonomous programmatic table agent that simulates the human-like cognitive schema of multi-turn interaction within a lightweight LLM. Instead of adopting a training-free workflow design, TableMind learns to internalize planning, action, and reflection through a principled two-stage training strategy. To bootstrap structured table reasoning capabilities, we construct and filter high-quality reasoning data for the supervised fine-tuning (SFT) stage. To enable precise code generation, we introduce a designed multi-perspective reward scheme and a novel optimization objective in the reinforcement learning (RL) stage. Extensive experiments on diverse benchmarks demonstrate that TableMind consistently outperforms previous baselines, validating the effectiveness of training autonomous agents to improve overall performance.

Generated automatically from available metadata and abstract on 2026-11-09.