MLPs, Encoder-Free MLLMs, and Self-Evolving Vision Models Lead New arXiv Batch

A cluster of newly posted arXiv papers takes aim at the architectural and training assumptions behind today's multimodal and vision-language systems, with several works questioning how much of the standard stack is actually necessary.

One line of work asks whether attention is being fully exploited in multimodal language models, with a paper titled "Just MLPs: Efficient Visual State Reconstruction for Multimodal Language Models" posing the question of whether the attention mechanism is really used to its full extent. A related paper, "How Far Are We from Removing the Visual Encoder? Scaling Laws for Encoder-Free Multimodal Pretraining," notes that most modern multimodal large language models (MLLMs) rely on a pretrained visual encoder that supplies a strong visual prior, while encoder-free MLLMs instead learn visual representations directly from raw pixels, offering a simple and unified architecture — though the paper states their scaling behavior has not been systematically studied.

A second thread targets efficiency in generation and embodied navigation. "WaveFront Decoding: Parallelized Self-Speculative Decoding for Looped Language Models" introduces WaveFront Decoding for looped language models. "DroneWAM: Efficient World Action Model for Drone Visual Navigation" presents DroneWAM, described as an efficient world-action model for drone visual navigation that adopts JEPA-based predictive learning to model future states directly in representation space, focusing computation on predictive scene structure and motion cues most relevant to navigation. "VLA-Precision: Asymmetric Co-Bootstrapping for Efficient Real-World Online RL of Vision-Language-Action Models" carries arXiv identifier 2609.04355.

A third group addresses how vision-language models generate and verify their own training signals. "Program-Verified Self-Evolution for Vision-Language Models" observes that self-evolving vision-language models learn from questions they generate from unlabeled images, but the labels they use are often wrong, with human checks showing 24% of majority-vote labels and 18% of model-judge labels are incorrect; the proposed VQS approach has the model turn each image into a structured form instead.

Finally, "SolveEdit: Benchmarking Visual Problem Solving in Generative Models" offers a benchmark for visual problem solving in generative models, with a repository at https://github.com/WenjieShu/SolveEdit.