V-tableR1: Process-Supervised Reinforcement Learning Ushers in the Verifiable Era of Multimodal Table Reasoning

This article introduces the V-tableR1 framework, which uses a specialized evaluator VLM to provide dense step-level feedback and combines optimization with the PGPO algorithm. It allows multimodal large models to move from black-box pattern matching to verifiable logical deduction, achieving the best performance among open-source models on complex table reasoning benchmarks. This framework marks a major shift in the multimodal reasoning paradigm from black-box pattern recognition to transparent, verifiable logical deduction.

Current multimodal large models (MLLMs) adopt an end-to-end training paradigm, with internal mechanisms lacking transparency. They tend to rely on statistical correlations to guess answers rather than logical reasoning. In table reasoning, tasks such as locating cells, extracting values, performing calculations, and verifying logic are required; the black-box model easily leads to uncontrollable error rates. Supervised Fine-Tuning (SFT) and outcome-oriented reinforcement learning only focus on the final answer, which incentivize shortcut learning and hallucinations, resulting in intermediate steps with logical errors.

V-tableR1 introduces a process supervision mechanism where the evaluator model verifies each step of the reasoning chain. Tables, due to their structured nature, serve as an ideal testbed, allowing explicit visual chains of thought (e.g., locating columns, extracting values, calculating growth rates). In the dual VLM architecture, the policy VLM generates the reasoning chain and answer, while the evaluator VLM provides feedback on each step (whether it is reasonable, has errors, or is in the correct direction), offering dense step-level learning signals.

The research team proposed the Process-Guided Direct Alignment Policy Optimization (PGPO) algorithm: 1. Process reward integration: Convert evaluator feedback into fine-grained step rewards (positive reward for correct positioning, penalty for incorrect extraction); 2. Decoupled policy constraints: Balance exploring new strategies and maintaining basic capabilities; 3. Length-aware dynamic sampling: Adaptively adjust the length of the reasoning chain (short chains in the early stage to build a foundation, long chains in the later stage to promote rigor).

V-tableR1 performs excellently on complex table reasoning benchmarks: 1. V-tableR1-4B (4 billion parameters) outperforms models with 18 times more parameters, proving that the improvement in reasoning ability does not rely on scale expansion; 2. It significantly improves accuracy and reasoning quality compared to the SFT baseline; 3. Ablation experiments show that process supervision effectively suppresses hallucinations (fictional values) and shortcut learning (skipping steps).

The contributions of V-tableR1 include: 1. A general framework for verifiable reasoning (applicable to structured visual tasks such as chart understanding and geometric proof); 2. A specialized path for evaluator models (providing more accurate feedback); 3. Paradigm shift: From statistical pattern matching to logical deduction, providing a premise for AI deployment in high-risk scenarios (medical, finance).

V-tableR1 has limitations: 1. It relies on structured table input; expanding to open-domain images requires new reasoning chain expression and verification; 2. The training cost of the evaluator model is high (step-level annotation is expensive); 3. A trade-off between reasoning chain length and efficiency is needed; 4. There is a lack of theoretical basis for the effectiveness of process supervision. Future directions include semi-automatic evaluator model training, adaptive reasoning depth, and interdisciplinary theoretical research.

V-tableR1 has broad application prospects: financial data analysis (auditable verifiable reasoning chains), scientific research assistance (quick analysis of experimental tables), and business intelligence (natural language interfaces for non-technical users). Conclusion: V-tableR1 is a milestone in multimodal reasoning, proving that process-supervised reinforcement learning can turn models into transparent logical reasoning engines, promoting the evolution of AI toward rigor and reliability.