Multimodal large language models (MLLMs) can appear fluent in vision-language tasks, but they often stumble on problems that demand fine-grained visual perception. The authors of V-GIFT argue the root cause is not weak visual representations but under-utilization of visual information during instruction tuning: many tasks can be partially solved using language priors alone, so the model never fully learns to look.

Their solution is simple and training-light. They convert classical self-supervised pretext tasks—rotation prediction, color matching, and cross-view correspondence—into image-instruction-response triplets phrased as natural language. Because these tasks cannot be solved without consulting the image, they force the model to rely on visual evidence. The approach requires no human annotations, no architectural changes, and no extra training stages.

Across multiple models, training regimes, and benchmarks, the authors report that adding only a small fraction (3–10%) of these visually grounded instructions consistently improves performance on vision-centric evaluations. The work positions instruction tuning with self-supervised tasks as a practical lever for improving visual perception in MLLMs by simply adjusting the training data distribution. As a single-source digest, this summary reflects only the paper's own claims; no independent replication or contrasting results are reported here.