Ai2 Newsletter

PDFs are notoriously difficult to extract text from. Standard tools struggle with reading order, often jumbling text or missing content entirely, especially with scanned documents and handwritten text. We trained olmOCR on academic papers, technical documentation, and other reference content, and use a unique prompting technique to increase accuracy and decrease hallucinations. It outperforms leading PDF tools with a computing cost 1/32 of GPT-4o, and is open-source! Read the blog for how we built and evaluated olmOCR.

Over the past ten years, EarthRanger has evolved from an ambitious idea into a widely adopted platform that helps conservation organizations worldwide. But today, the focus isn’t on technology–it’s on the people. Rangers, ecologists, and park managers work tirelessly to safeguard nature, promote human-wildlife coexistence, and–just as importantly–ensure the safety of those on front lines.

Reasoning about images with rich text, such as charts and documents, is a critical application of vision-language models (VLMs). However, VLMs often struggle in these domains due to the scarcity of diverse text-rich vision-language data. We present CoSyn, a framework that leverages the coding capabilities of text-only LLMs to automatically create synthetic text-rich multimodal data.