LensVLM
A 9-billion-parameter vision-language model, released with its evaluation code, that scans compressed page images and selectively expands only the relevant ones to their uncompressed form using learned tools, for answering questions over long documents.
LensVLM is a vision-language model that handles long documents by keeping every page compressed by default and expanding only the pages it judges relevant to a question, via a learned tool-calling loop. Given a set of page images and a question, it identifies which pages carry the answer, reads their text through a page-reading tool, and produces a multi-turn trajectory that connects the evidence across pages.
The repository includes scripts to prepare evaluation data from Hugging Face datasets at several compression ratios, run the model on a custom document, and evaluate its answers against an LLM judge served through any OpenAI-compatible endpoint. It is released as Python code with a requirements file for installation.