Experiments
Replacing FWPRec's memory grids with a small private neural network: what we ran, and what held up. The professor's earlier results are on the Handoff page.
The idea
Not a run. The question the handoff asks, what changes, what stays the same, and what not to claim.
Read the ideaHow the memory works
Not a run. The grid and the network side by side: how each is built, written and read.
Read the methodImportant things to note
Not a run. Where the evidence stands after our rerun, and the next steps in order.
Read the notesRuns
The same six memories on MIND news
We repeated the Amazon design on MIND-large news, where readers skip most of what they see, splitting six models over three lab machines. The grid scored 0.378 and the merged network 0.378; plain averages with no learned memory reached 0.352, far closer than on Amazon.
OpenSix memories, eight seeds, on Amazon
We widened the Amazon test to eight seeds and six memories: the grid, three network versions, and two plain averages with no learned memory. The learned memories score around 0.26; the plain averages reach only 0.17. The merged network led the grid on 6 of eight seeds.
OpenRerunning the professor's Amazon test
We reran the professor's grid-versus-network comparison on Amazon video-game reviews with his unchanged code and settings. Tuning scores came back nearly identical. On held-out users, the network's lead shrank to under two percent and reversed on one of three seeds.
Open