QLIP: A Dynamic Quadtree Vision Prior Enhances MLLM Performance Without Retraining

Multimodal Large Language Models (MLLMs) encode images into visual tokens, aligning visual and textual signals within a shared latent space to facilitate crossmodal representation learning. The CLIP model is a widely adopted foundational vision language model whose vision encoder has played a critical role in the development of MLLMs such as LLaVA. However, the CLIP vision encoder suffers from notable limitations including being constrained to only handling fixed input resolutions and a failure to produce separated embeddings for dissimilar images. Replacing the vision encoder of an existing model typically incurs substantial computational costs because such a change often necessitates retraining the entire model pipeline. In this work, we identify two factors which underlie the limitations of the CLIP vision encoder: mesoscopic bias and interpolation bias. To address these issues, we propose QLIP, a drop-in replacement for CLIP that can be seamlessly integrated with existing MLLMs with only a few lines of code and can enhance both coarse-grained and fine-grained visual understanding, without re-training. QLIP is designed around an image quadtree which replaces the standard uniform grid patches with a novel content aware patchification. Our experimental results demonstrate that QLIP improves the general visual question answering accuracy of the LLaVA v1.5 model series across various model sizes--without requiring retraining or fine-tuning of the full MLLM. Notably, QLIP boosts detailed understanding performance on the challenging $V^{\ast}$ benchmark by up to 13.6 percent.

A Quasilinear Algorithm for Computing Higher-Order Derivatives of Deep Feed-Forward Neural Networks

The use of neural networks for solving differential equations is practically difficult due to the exponentially increasing runtime of autodifferentiation when computing high-order derivatives. We propose $n$-TangentProp, the natural extension of the TangentProp formalism \cite{simard1991tangent} to arbitrarily many derivatives. $n$-TangentProp computes the exact derivative $d^n/dx^n f(x)$ in quasilinear, instead of exponential time, for a densely connected, feed-forward neural network $f$ with a smooth, parameter-free activation function. We validate our algorithm empirically across a range of depths, widths, and number of derivatives. We demonstrate that our method is particularly beneficial in the context of physics-informed neural networks where \ntp allows for significantly faster training times than previous methods and has favorable scaling with respect to both model size and loss-function complexity as measured by the number of required derivatives. The code for this paper can be found at https://github.com/kyrochi/n\_tangentprop.