4 ms·
LLM based OCR is a disaster, great potential for hallucinations and no estimate of confidence. Results might seem promising but you’ll always be wondering.
by deadbabe 2y ago
LLM based OCR is a disaster, great potential for hallucinations and no estimate of confidence. Results might seem promising but you’ll always be wondering.
- menaerus 2y agoCNN-based OCR also have "hallucinations" and Transformers aren't that much different in that respect. This is a problem solved with domain specific post-processing.
- leumon 2y agowell already in 2013 ocr systems used in xerox scanners (turned on by default!) randomly altered numbers, so its not an issue only occuring in llms.
- utkarshphirke 2y agoAbsolutely right - we tried estimating LLM confidence and the results are not great. Any process that requires reliability will struggle with LLM OCR. https://news.ycombinator.com/item?id=43350816 https://news.ycombinator.com/item?id=43350816