Three literature syntheses: document AI, few-shot AR and agent reliability

Each synthesis starts from a search across major venues and preprints, then charts every work in the same columns: the method, what it achieved, where it falls short, and which gaps keep reappearing across papers. The topic comes out of the chart, not the other way round.

33 works on extracting structured fields from invoices, receipts and forms. The recurring theme: accuracy on benchmarks is high, but knowing when an extraction is wrong under distribution shift, and routing cheaply between small and large models, is not solved.

31 papers on teaching on-device systems to recognise new objects from a handful of examples. The open gaps: personalization that survives cluttered scenes, continual “teachable” recognition, and open-set detection of things that are none of the user’s objects.

33 papers (9 read in full) on making small agents reliable across repeated runs (pass^k) rather than just accurate once, with attention to what can be post-trained on a single GPU.