Research & models Aug 9, 2026

A paper claiming a "third pretraining axis" ran into two metrics from the same knob pointing opposite ways, August 4, 2026.

"Third axis" is the paper's own framing; we do not endorse it. "Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation" (arXiv 2607.27372) self-reports 4.1× compute efficiency and 6.2× sample efficiency. The shape of the numbers former Meta researcher Armen Aghajanyan reported is unusually clean: as the number of candidates generated per step rises, "the objective worked as intended: selected velocity MSE improved 0.0896→0.0763→0.0703 as K rose. But inference MSE (actual rollouts on held-out data) worsened 0.1437→0.1631→0.1835" — one knob, two metrics, opposite directions, both monotonic (the post, August 4, 2026). Held-out data is data the model never saw in training, reserved specifically to test it; improving on the training objective while degrading on held-out data is the textbook shape of the optimized metric coming apart from the property you actually want. ⚠️ That reversal is reported by a single source, not reproduced by us, and we have not confirmed whether his experimental setup is comparable to the paper's — which is exactly the failure mode this item is criticizing, so it is a lead worth chasing, not an established refutation; we have not read the full paper, and the dispute lies entirely in the experiments section. Three questions worth taking away (none of it requires understanding the technical detail): where was this metric measured, on the training distribution or on held-out data? Was anything removed from the denominator? And is this new, or old work under a new name? Efficiency claims are becoming inputs to valuation models and procurement decisions, and these three questions are the shortest path back to something checkable.