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 Sat, 26 Sep 2026 06:57:35 +0000 https://awarenesssoftwaregroup.com/research/a-fixed-budget-thinks-better/ https://awarenesssoftwaregroup.com/research/a-fixed-budget-thinks-better/ Penalising a model for thinking cut its thinking by 38% at almost no cost. Always letting it think for three steps beat that, and we found out why. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-free-check-saturates-at-three/ https://awarenesssoftwaregroup.com/research/a-free-check-saturates-at-three/ A one-pass check that tells a helpful trained model from a harmful one gets all of its value from the first three candidates. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-rule-from-1990/ https://awarenesssoftwaregroup.com/research/a-rule-from-1990/ We finally asked what a pre-neural predictor scores on the task most of our results were measured on. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-short-model-is-a-lottery/ https://awarenesssoftwaregroup.com/research/a-short-model-is-a-lottery/ Two of our own experiments reached opposite conclusions from one trained model each. Sat, 26 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. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-window-that-parks/ https://awarenesssoftwaregroup.com/research/a-window-that-parks/ We wrote down the practice plan an adaptive teacher discovered, as a two-setting rule. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/being-choosy-about-data-costs-more-than-it-saves/ https://awarenesssoftwaregroup.com/research/being-choosy-about-data-costs-more-than-it-saves/ Inspect four batches, train only on the most informative. It sounds efficient. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/calibrate-after-learning/ https://awarenesssoftwaregroup.com/research/calibrate-after-learning/ Training a thinking model's early answers blocked learning when done from the start. Done after it learned, it helped every run. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/does-the-early-signal-generalise/ https://awarenesssoftwaregroup.com/research/does-the-early-signal-generalise/ Two findings that looked independent turn out to have identical boundaries. The early-warning gap appears exactly where the internal reorganisation does. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/eight-runs-is-not-enough/ https://awarenesssoftwaregroup.com/research/eight-runs-is-not-enough/ A free experiment produced two statistically significant results pointing in opposite directions. Our first analysis reported one of them as a discovery. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/how-you-cut-a-model-matters/ https://awarenesssoftwaregroup.com/research/how-you-cut-a-model-matters/ Delete three quarters of a model's internal connections two different ways, keeping exactly the same number, and one way costs twice as much as the other. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/is-transition-timing-controllable/ https://awarenesssoftwaregroup.com/research/is-transition-timing-controllable/ A nudge of one part in a hundred million delays learning by 21 steps. A nudge ten million times larger delays it by 24. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/it-already-knew-the-order/ https://awarenesssoftwaregroup.com/research/it-already-knew-the-order/ We found the order in which a model learns three related skills, then tried teaching them in that order. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/it-sees-the-damage-not-the-repair/ https://awarenesssoftwaregroup.com/research/it-sees-the-damage-not-the-repair/ Rounding a model's weights to fewer bits damages it. Measuring its internal shape spots the damage, but a repaired model works again with a changed shape. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/learning-it-again-is-not-like-learning-it/ https://awarenesssoftwaregroup.com/research/learning-it-again-is-not-like-learning-it/ We changed the task underneath a trained model and watched it learn again. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/looking-is-not-using/ https://awarenesssoftwaregroup.com/research/looking-is-not-using/ One model finds the right word as fast as the other and answers correctly less than half as often. Finding and using turn out to be separate abilities. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/no-sample-small-enough/ https://awarenesssoftwaregroup.com/research/no-sample-small-enough/ Maybe a clever way to prevent forgetting beats simple replay when old examples are scarce. It does not: with few examples both do equally badly. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/one-jump-or-several/ https://awarenesssoftwaregroup.com/research/one-jump-or-several/ Averaging accuracy across positions hid four separate events. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/one-number-for-every-size/ https://awarenesssoftwaregroup.com/research/one-number-for-every-size/ We published one figure for what our early-warning measurement costs, then quoted it in eight results. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/predicting-it-without-watching/ https://awarenesssoftwaregroup.com/research/predicting-it-without-watching/ Our only working early-warning instrument costs 55% extra compute to buy about 23 steps of notice. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/protected-and-adaptable/ https://awarenesssoftwaregroup.com/research/protected-and-adaptable/ A new method teaches a model a new skill only through the directions its old skill leaves unused. It works. Replaying a few old examples works better. Sat, 26 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. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/representation-geometry-is-not-load-bearing/ https://awarenesssoftwaregroup.com/research/representation-geometry-is-not-load-bearing/ The white paper closing our first research programme. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/revisiting-rescues-the-easy-tasks/ https://awarenesssoftwaregroup.com/research/revisiting-rescues-the-easy-tasks/ Adding one rule to a hand-written practice plan, go back to whatever slipped, closes half the gap to an adaptive teacher. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/right-order-right-answers-no-room/ https://awarenesssoftwaregroup.com/research/right-order-right-answers-no-room/ Trained to answer at every step, a thinking model finally stops in the right places with the right answers. Sat, 26 Sep 2026 00:00:00 +0000