Table-R1: Inference-Time Scaling for Table Reasoning

Motivation

• Inference-time scaling for table reasoning remains largely unexplored despite challenges involving diverse cells, cross-table alignment, aggregation, numerical operations, and long inputs.These capabilities support applications such as data analysis, scientific reporting, and decision-support systems.Source: §1 · pp. 1–2

Approach

• The study develops Table-R1-SFT from DeepSeek-R1 reasoning traces and Table-R1-Zero using task-specific verifiable rewards with GRPO.The distillation dataset contains verified reasoning traces, while RLVR is tailored to table reasoning.Source: §1 · p. 2

Results

• Across short-form QA, fact verification, and free-form QA, RLVR shows better performance and generalization than distillation.The models are evaluated on both in-domain and out-of-domain table reasoning benchmarks.Source: §1 · p. 2
• Table-R1-Zero is competitive with GPT 4.1 and DeepSeek R1 while using only a 7B-parameter Qwen2.5-7B backbone.Qualitative analysis also identifies multi-step reasoning, reflection, semantic understanding, information extraction, and arithmetic computation.Source: §1 · p. 2

2.1 Inference-Time Scaling

• Inference-time scaling uses longer reasoning, search, process rewards, supervised traces, and RLVR to improve reasoning across multiple domains.Prior applications include multimodal reasoning, translation, tool use, information retrieval, mathematics, games, vision, and agents.Source: §2.1 · pp. 2–3
• RLVR optimizes language models with rule-based rewards, enabling autonomous discovery of reasoning strategies without explicit intermediate supervision.Later work improves stability and sample efficiency through dynamic sampling, token-level policy gradients, and reward normalization.Source: §2.1 · p. 2

2.2 Table Reasoning

• Table reasoning spans short-form question answering, fact verification, and free-form question answering, with research progressing from task-specific models toward general-purpose LLMs and agents.Despite these advances, inference-time scaling for table reasoning remains largely underexplored.Source: §2.2 · p. 3
• Table-R1 shows that 7B-scale LLMs with inference-time scaling can match frontier-model performance such as GPT-4.1 on table reasoning.This result motivates inference-time scaling as a direction for structured reasoning.Source: §2.2 · p. 3

3 Table-R1 Models

• Table-R1 defines two variants: Table-R1-SFT, trained on frontier-model reasoning traces, and Table-R1-Zero, trained with table-specific RLVR.Both variants target inference-time scaling for table reasoning.Source: §3 · p. 3

3.1 Training Data Collection

• The training data combines short-form QA, fact verification, and free-form QA, whose ground truths support verifiable reward computation.These tasks require precise answers, entailment judgments, or open-ended responses grounded in tables.Source: §3.1 · p. 3, Table 1 · p. 3

3.2 Table-R1-SFT via Supervised Finetuning

• Table-R1-SFT uses DeepSeek-R1 to generate step-by-step traces and filters incorrect answers, yielding 33,601 high-quality training examples.The resulting dataset is used to fine-tune the SFT model.Source: §3.2 · p. 3

3.3 Table-R1-Zero via RL Training

• The RLVR procedure adapts GRPO with token-level loss computation and asymmetric clipping, while omitting the original KL penalty term.The method samples response groups, computes group-normalized advantages, and optimizes token-level objectives.Source: §3.3 · pp. 3–4
• Table-R1-Zero combines task-specific accuracy rewards with format rewards that encourage outputs matching a structured think-and-answer template.Accuracy uses exact match for TQA, classification correctness for TFV, and normalized BLEU and ROUGE-L for FF-TQA.Source: §3.3 · p. 4
• The combined reward design is intended to make training interpretable and robust, mitigating instability and reward hacking.The authors state that the resulting formulation enables stable and effective RL fine-tuning for table reasoning.Source: §3.3 · p. 4

4.3 Main Findings

• Inference-time scaling alone is not consistently advantageous for table reasoning, because reasoning models outperform general-purpose counterparts mainly on TQA but show mixed results on TFV and FF-TQA.The finding motivates table-specific adaptation rather than relying solely on general reasoning capability.Source: §4.3 · p. 6
• Table-R1-Zero-8B reaches 32.7 BLEU on FeTaQA, while Table-R1-SFT-8B reaches 91.1 accuracy on TabFact and Table-R1 models achieve 83.8 on WTQ and 81.8 on HiTab.These results are competitive with or exceed the strongest reported comparison scores of 26.2, 91.9, 85.4, and 84.7 on the corresponding tasks.Source: §4.3 · p. 6
• Table-R1 models generalize across out-of-domain benchmarks, with Table-R1-Zero-7B performing best overall, while SFT and Llama initialization generalize less effectively.The comparison indicates stronger out-of-domain robustness for RLVR and Qwen-based models.Source: §4.3 · p. 6

4.4 Analysis of Training Dynamics

• Instruction-tuned models train more stably: response lengths first drop during format acquisition, then increase, whereas Llama-3.1-8B shows instability and fails to acquire the desired format.Instruction models also maintain higher accuracy, while base-model accuracy can converge after format adaptation; Llama still gains on most tasks except TabFact.Source: §4.4 · pp. 6–7

4.5 Reasoning Capacity Boundary

• RLVR increases pass@k through k=32 on HiTab and WTQ, improving both first-attempt correctness and the breadth of valid reasoning paths.The effect is consistent across datasets and model architectures.Source: §4.5 · p. 7

4.6 Qualitative Analysis

• RL training produces explicit multi-step, reflective, column-aware, semantic, arithmetic, and temporal reasoning abilities that are rare at initialization.The qualitative progression includes decomposing queries, checking earlier conclusions, interpreting question intent, and synthesizing information across columns.Source: §4.6 · pp. 7–8

4.7 Ablation Studies on RLVR Training

• Ablations show that instruction tuning improves adaptation, architecture trades in-domain accuracy against cross-domain robustness, and domain-specific distillation data outperforms official distilled models.Llama-3.1-8B-Instruct performs better in-domain, whereas Qwen2.5 generalizes better out of domain; Table-R1 distillation models surpass official and larger distilled models.Source: §4.7 · p. 9
• Task transfer is uneven: TQA-only training improves TFV but not FF-TQA, while removing format rewards reduces stability, in-domain performance, and generalization.The results indicate that free-form answering requires skills not adequately covered by TQA, and format supervision supports transferable reasoning.Source: §4.7 · p. 9

5 Conclusion

• Across 13 diverse benchmarks, inference-time scaling improves table reasoning, with RLVR producing stronger out-of-domain generalization.Ablations also support benefits from instruction tuning, model architecture, and task composition.Source: §5 · p. 9
• Qualitative analysis shows that RLVR fosters structured multi-step reasoning and table-specific capabilities.The paper highlights these capabilities as emerging during RL training.Source: §5 · p. 9

• SFT data comes exclusively from DeepSeek-R1, limiting the diversity of reasoning traces used for distillation.The authors suggest incorporating outputs from other reasoning models such as QwQ-32B.Source: Limitations · p. 9
• Data verification and filtering may have removed difficult or high-quality training examples.Source: Limitations · p. 9
• LLaMA3.1-8B initializations show unstable RLVR training, including inconsistent output formatting and fluctuating response lengths.Instruction tuning mitigates some issues, but the underlying causes remain underexplored.Source: Limitations · p. 9
• Future work should improve RLVR robustness and generalizability by investigating initialization sensitivity, reward sparsity, and optimization hyperparameters.Source: Limitations · p. 9

• Training data originates from WTQ, HiTab, TabFact, and FeTaQA, covering the paper’s table-reasoning tasks.Source: §A.1 · p. 14
• TabFact is filtered by difficulty, yielding 20,740 samples, and the combined RLVR dataset contains 48,563 samples after length filtering.Samples with pass rates ≤7/16 are retained, while 10,000 are sampled from the 8/16–15/16 range.Source: §A.1 · p. 14
• SFT traces are generated by DeepSeek-R1 and retained using exact-match criteria for WTQ, HiTab, and TabFact, or average BLEU and ROUGE-L ≥0.35 for FeTaQA.The resulting SFT dataset contains 33,601 samples.Source: §A.1 · p. 14
• Quality filtering reduces the overall response-to-prompt length ratio from 1.73 to 1.49.The paper associates this reduction with removing incorrect or lower-quality, longer reasoning traces.Source: §A.1 · p. 14

• Table 5 provides an overview of datasets used in training and evaluation.Source: Table 5 · p. 15
• The table concerns both training datasets and evaluation datasets.Source: Table 5 · p. 15
• The supplied caption establishes the table’s scope but does not provide dataset names or quantitative statistics.Source: Table 5 · p. 15

B.1 Baseline Systems

• Baseline systems include GPT-4.1, GPT-4.1 mini, Qwen2.5 and LLaMA3 models, DeepSeek-R1 variants, QwQ-32B, and table-oriented models.Source: §B.1 · p. 15

B.2 Inference Setup

• Open-source inference uses vLLM, closed-source inference uses the OpenAI API, and reasoning models receive an 18,000-token output limit.Most models use a 2,048-token limit; inference runs on four NVIDIA A100-80G GPUs with temperature 0.6 and top-p 0.95.Source: §B.2 · p. 15

B.3 Evaluated Model Configuration

• The evaluated model configuration includes the model list summarized in Table 6.Source: Table 6 · p. 15

C Prompts

• Prompts specify separate formats for short-answer QA, fact verification, and free-form QA, including structured JSON answer fields.Short-answer outputs use answer lists, fact verification outputs use entailed or refuted labels, and free-form QA outputs use a sentence.Source: C Prompts · p. 16

• Qualitative case examples are provided for TQA tasks.Source: Figure 6 · p. 18
• Qualitative case examples are also provided for TFV tasks.Source: Figure 7 · p. 19
• Pass@k performance is presented for the WTQ and HiTab datasets.Source: Figure 8 · p. 20
• The ablation study reports results across model-level, task-level, and format dimensions.Source: Table 7 · p. 20