AkbasCore tests tiny model steering without retraining
AkbasCore is a tool that adds very small directional pressure to a ’s while the model is doing inference. It is designed to work without retraining the model, changing , or changing the input prompt.
In the default setup, each layer’s pressure is meant to stay at or below the bfloat16 precision floor, about 0.0078 per unit. The claim is that standard checks show no change, while other measured output-quality differences still appear.
As of June 2026, it is a pre-print and has not yet gone through peer review. The author says 76 related tests were publicly posted on Reddit before the pre-print.
Key points
- AkbasCore steers a model during inference by changing s slightly.
- It is described as working without retraining, weight changes, or prompt changes.
- The default pressure level is designed to stay under the bfloat16 precision floor of about 0.0078.
- The claim is that shows no change, but output quality still changes in measurable ways.
- It is a June 2026 pre-print and is not yet peer reviewed.