Multimodal Large Language Models (MLLMs) often falter in complex reasoning tasks, treating visual input as a black box that leads to superficial pattern matching rather than deep inference. Existing methods struggle to bridge the gap between abstract logic and the continuous pixel space required for visual verification.
Bridging Logic and Pixels with V-tableR1
The V-tableR1 framework directly addresses this challenge by introducing process-supervised reinforcement learning tailored for multimodal domains. It leverages the deterministic structure of tables as an ideal testbed, enabling a specialized critic VLM to provide granular, step-level feedback on the explicit visual chain-of-thought generated by a policy VLM. This approach fundamentally shifts multimodal inference from opaque pattern matching to a verifiable logical derivation process.
Process-Guided Alignment for Robust Inference
Optimizing this system requires a novel approach, leading to the development of Process-Guided Direct Alignment Policy Optimization (PGPO). This RL algorithm integrates process rewards, decoupled policy constraints, and length-aware dynamic sampling. Extensive evaluations confirm that the V-tableR1 framework effectively penalizes visual hallucinations and shortcut guessing. The result is state-of-the-art accuracy among open-source models on complex tabular benchmarks, with the 4B parameter model outperforming models up to 18 times its size and significantly improving over its SFT baseline.