https://awarenesssoftwaregroup.com/research/ Open research on how AI models actually learn. One finding per entry, in plain English first and in full technical detail underneath, including the approaches that turned out not to work. en Tue, 29 Sep 2026 14:01:43 +0000 https://awarenesssoftwaregroup.com/research/a-delay-by-most-measures/ https://awarenesssoftwaregroup.com/research/a-delay-by-most-measures/ The easy-sum warm-up that looked like a trap mostly delayed learning: given longer, most runs learned the hard sum, several thousand steps later than Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-higher-ceiling-not-a-head-start/ https://awarenesssoftwaregroup.com/research/a-higher-ceiling-not-a-head-start/ Given more training, models that started on a random mix of easier tasks reached 94 percent on the hard task while models trained on it alone stalled at Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-learning-rate-warm-up-does-the-job/ https://awarenesssoftwaregroup.com/research/a-learning-rate-warm-up-does-the-job/ Starting training with small steps and ramping up did as well as our best easy-data warm-up, and better than the easy-task mix. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-sweet-spot-not-the-nearest/ https://awarenesssoftwaregroup.com/research/a-sweet-spot-not-the-nearest/ Before a hard memory task, the best single warm-up was a version easier by a clear margin. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-task-battery-with-designed-headroom/ https://awarenesssoftwaregroup.com/research/a-task-battery-with-designed-headroom/ Our quality checks all passed because the benchmark scored 99% and nothing could fail. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/against-the-best-plain-run-not-established/ https://awarenesssoftwaregroup.com/research/against-the-best-plain-run-not-established/ On the second task, starting with easier sums solved the hard sum on 11 of 16 runs against 6 for the best ordinary recipe, but the accuracy difference is Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/at-width-96-it-holds/ https://awarenesssoftwaregroup.com/research/at-width-96-it-holds/ In a model twice as wide, on a hard sum matched to it, the easy-sum start reached the solution in about 46 percent fewer steps, and every run of both Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/at-width-96-the-gap-holds-steady/ https://awarenesssoftwaregroup.com/research/at-width-96-the-gap-holds-steady/ In a model twice as wide, trained 2.5 times longer, the easy-task warm-up stayed about five points ahead but did not pull further away, unlike in the Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/differences-help-too/ https://awarenesssoftwaregroup.com/research/differences-help-too/ Warming up on easy subtraction problems sped up learning a hard addition problem almost as much as easy addition did. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/easier-data-adds-nothing-to-a-rate-warm-up/ https://awarenesssoftwaregroup.com/research/easier-data-adds-nothing-to-a-rate-warm-up/ Combined with a standard learning-rate warm-up, our best easy-data warm-up added nothing. The best result of the whole study needed no special data at all. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/forty-percent-fewer-steps/ https://awarenesssoftwaregroup.com/research/forty-percent-fewer-steps/ On 32 fresh runs, starting with a short mix of easier sums reached the solution to a hard sum in about 40 percent fewer training steps than the best Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/it-must-be-an-early-phase/ https://awarenesssoftwaregroup.com/research/it-must-be-an-early-phase/ Easy sums helped a small model learn a hard sum as a short phase at the start. Mixed in for the whole run, the same sums left every run unable to learn it. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/it-replicates-on-a-second-target/ https://awarenesssoftwaregroup.com/research/it-replicates-on-a-second-target/ On a different hard sum, the easy-sum mix again beat the best ordinary recipe on fresh runs, solving it on 25 of 32 runs against 2. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/kept-in-they-get-in-the-way/ https://awarenesssoftwaregroup.com/research/kept-in-they-get-in-the-way/ Even with the hard sum three quarters of the training data, keeping easy sums in the mix stopped every run learning it, worse than no easy sums at all. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/no-other-claim-ran-hot/ https://awarenesssoftwaregroup.com/research/no-other-claim-ran-hot/ We checked every standing speed-up result in the archive for the mistake that undid our warm-up results. Only the warm-up results themselves had made it. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/on-a-transformer-not-established/ https://awarenesssoftwaregroup.com/research/on-a-transformer-not-established/ On a small transformer, the easy-sum start got more runs to the target, but most runs of both recipes had not got there by the end, so the result is not Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/on-fresh-seeds-the-mixture-wins/ https://awarenesssoftwaregroup.com/research/on-fresh-seeds-the-mixture-wins/ On the adding task, starting with a mix of easier sums beat the best ordinary recipe, learning-rate warm-up included, on 32 new runs: 91 percent against Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/one-nearby-version-does-most-of-it/ https://awarenesssoftwaregroup.com/research/one-nearby-version-does-most-of-it/ Warming up on a single easier version close to the hard task gave most of the benefit, on every run. Mixing several easier versions added at most a little. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/order-can-make-the-warm-up-harmful/ https://awarenesssoftwaregroup.com/research/order-can-make-the-warm-up-harmful/ The same three easier sums helped a small model when shuffled and stopped it learning the hard sum at all when given from hardest to easiest. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/order-did-not-help-mixing-might/ https://awarenesssoftwaregroup.com/research/order-did-not-help-mixing-might/ Working up from easy to hard did not speed learning at matched compute; a random mix of easier tasks early ended clearly higher, which we are now testing Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/plain-gets-there-eventually/ https://awarenesssoftwaregroup.com/research/plain-gets-there-eventually/ Given three times as long, ordinary training learned the harder sum on 10 of 16 runs. The easy-sum mix got there in about a third of the steps. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/predicting-the-cost-from-the-task/ https://awarenesssoftwaregroup.com/research/predicting-the-cost-from-the-task/ What it costs to take something away from a model turns out to be set by the structure of the exercise, not by the model. Change the exercise and it moves. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/questions/ https://awarenesssoftwaregroup.com/research/questions/ The 10 questions our public AI training research is organised around, each with its current answer and the published results behind it. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/re-tuned-for-decay-plain-still-trails/ https://awarenesssoftwaregroup.com/research/re-tuned-for-decay-plain-still-trails/ Given its best settings for a decaying learning rate, ordinary training reached 49 percent on the hard sum; the easy-sum mix reached 90. Tue, 29 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/slower-than-tuned-not-hotter/ https://awarenesssoftwaregroup.com/research/slower-than-tuned-not-hotter/ A check on one earlier result: its task learns fastest at twice the learning rate it was run at, so it did not have the too-large-rate problem. Tue, 29 Sep 2026 00:00:00 +0000