Pointer-style models pick answers from input text instead of generating them

The article argues that many LLM tasks, such as labeling a ticket or extracting an invoice number, do not require the model to write its answer token by token. Instead, a pointer approach scores the candidate labels or text spans already present in the input, so the output can only be something the input contains. This idea dates back to Pointer Networks, SQuAD and BERT, and is now offered by ScaleDown through four small task models. ScaleDown's Classify and Extract models point at answers, while its Summarize model rewrites text and its Compress model's method is unspecified. The author notes that pointed answers can be verified against character offsets, but that pointing cannot handle questions requiring computation, and that confidence scores still need calibration on one's own data.