Keeps learning without forgetting, going stale or growing.
An agent that runs for months needs five things at once: learn new work, keep earlier work, take corrections, stay a fixed size and know what it has not learned. Ten kinds of request were learned one after another, then the policy changed for four of them.
- 5 / 5Requirements for learning in production, all met on every runCache and retrieval averaged 2.2 / 5 · online fine-tuning 4 / 5
- 0%Stale answers after a policy changeCache and retrieval 72%
- FixedFootprint over timeCache and retrieval grew to ≈1,300 rows
Five requirements, three approaches.
Step 1 of 6: Learns
A correction reaches every wording.
One policy change, and four ways of asking for the same thing.
Step 1 of 4: Before
Four wordings of the same request, all answered with a full refund.
The same size, however long it runs.
How much each approach grows as the work goes on.
Step 1 of 3: The start
At the start, cache and retrieval held 128 rows.
Ten tasks later, it still knows the first one.
Accuracy right after learning each kind of request, and after all ten.
Step 1 of 2: Just learned
Right after learning each request type, all three were accurate.
How it was run.
- The work
- Ten customer-support and banking request types, from password resets to declined payments, learned one after another.
- The test
- Every question in new wording, about details not seen in training.
- Policy change
- Four of the ten types switched to a new right action; corrections arrived shuffled, the way a reviewer sends them.
- Pass marks
- Set before running, and the same for every approach.
- Scale
- One main run, then ten independent runs in a second setup.
- Compared with
- Cache and retrieval, a nearest-match store at its best tuning, and online fine-tuning, all with the same inputs.
What it means for your work.
Corrections hold everywhere.
After a policy change, Duplex gave no stale answers on any of ten runs.
Earlier work stays intact.
It carried all of its earlier work through, run after run.
It stays the same size.
Duplex added nothing to its footprint; a memory store grew to about 1,300 rows.
See it on your own work.
From your own assistant to every agent your company runs, a pilot puts Duplex on your own work, with the measures agreed before it begins.