The problem
What an organisation knows about its own processes sits in standard operating procedures, policy manuals, process maps, regulations and historical records, and the quality, completeness and currency of that material vary widely. Before any of it can be automated, someone has to establish from these scattered, mixed-format documents what each process takes in, what it does and what it has to produce.
What the paper does
The paper names the pieces. A “Workflow Signal” is a single cue, in a document or a conversation, about a process’s input, its processing or its output. A “Workflow Intention” is a complete triple of the three. One set of documents can describe several objectives at once — a “Mixed Intention” — and those have to be separated before anything is built.
It then gives these objects a mathematical form, signals as vectors and intentions as tensors, and designs an attention-based model to extract them. Text, images and documents are encoded separately, attended to within each kind, fused across kinds, and decoded one intention at a time until a stopping rule decides there are no more, or that a new one merely repeats an earlier one. Training procedures and loss functions are given.
What it shows
This is a design and theory paper rather than an experiment. Its results are definitions and properties — an intention never carries more information about an input, process or output than the signals it came from, and any two intentions in a set differ in at least one of the three — together with an estimate of the model’s size, about 27.5 billion parameters in all. It expects a set of documents to yield between one and five intentions.
What it does not claim
It reports no evaluation. It is explicit that the cost of attention grows with the square of the input length, which is a problem in the fusion stage and above all for long documents, and it suggests ways round that.
Where it sits
Kingston is first author, with equal contribution from Théo Fagnoni. The paper formalises the intention introduced in the Large Work Model paper, and the Prompt Intention paper that follows tests the idea in practice.