See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL

A new training stage called Visual Evidence Pre-Alignment (VEPA) aims to improve how multimodal large language models use visual information, addressing a weakness in current systems where responses can be inconsistent with the images they process [1]. Multimodal large language models (MLLMs) combine text reasoning with visual inputs, but their answers do not always match the underlying images, signaling poor use of visual evidence during inference [1]. The standard training approach begins with large-scale caption-based pretraining for general alignment, then moves to supervised fine-tuning and reinforcement learning for instruction following and complex reasoning [1]. This pretraining provides only weak visual grounding because short, coarse captions steer models toward obvious objects while ignoring finer visual details [1]. VEPA is designed as an intermediate stage between pretraining and post-training. It uses a sufficiency-driven objective paired with Group Relative Policy Optimization (GRPO) to optimize question-conditioned visual evidence descriptions [1]. The method pushes the model to generate descriptions of visual evidence that are sufficient to answer a given question, rather than relying on superficial correlations learned from captions [2]. Experiments across multiple benchmarks show that VEPA consistently improves performance on visually demanding evaluations and works alongside standard supervised post-training [1]. The gains come from strengthened, transferable visual grounding, not from extra task-specific training [1]. This means a model that undergoes VEPA can apply its improved visual understanding to new tasks without having been explicitly trained on them. The work appears as a preprint on arXiv, submitted in June 2026, and is categorized under Computer Vision and Pattern Recognition [1]. The paper's code and data links are hosted through the arXivLabs framework, which supports community-developed features on the preprint server [1]. The authors frame VEPA as a response to a structural limitation in how MLLMs are currently built—where the pretraining phase, despite its scale, leaves models with an incomplete grasp of the visual world that later fine-tuning struggles to fully correct [2]. By inserting a dedicated visual-evidence alignment step, the researchers argue that models can learn to consult images more faithfully before generating an answer, a shift summarized in the paper's title: "See First, Answer Later" [1].