Research papers
A complete result: the question asked, the protocol followed, what was measured, and above all the limits of what can be concluded from it.
Everything the lab makes public lands here — results as well as failures, protocols as well as conclusions. Each entry points back to a dated experiment and its protocol.
Our first works will appear once our large model, currently in training, is finished. They will be deposited here as they come, together with their protocol and reference data.
In the meantime, the home page describes the areas we work on and the method we follow.
A lab that published only its successes would be telling a false story. We make four kinds of document public, and the last matters as much as the first.
A complete result: the question asked, the protocol followed, what was measured, and above all the limits of what can be concluded from it.
Shorter, on one precise point: a measurement, an observation, a useful implementation detail — with no claim to generality.
Published before the experiment and timestamped: acceptance criteria, thresholds, analysis strata. This is what stops us rewriting our expectations after the fact.
The leads that came to nothing, documented as carefully as the rest. Knowing which saving does not work saves everyone time, starting with us.
Every figure comes with its measure of uncertainty and a breakdown by type of text. An average on its own does not tell you whether a gain is real.
We publish results, protocols and evaluation methods — not implementation recipes. Enough to judge our work and build on it, not enough to copy it.