Weight Editing
An editing method that directly modifies model parameter weights. Its effects are persistent but irreversible, and it can readily cause cascading degradation, so careful validation is required.
Timeline (last 90 days)
In the QUEN 3.5 0.8b case, Astra broke a reasoning loop by editing an extremely small proportion of the weights; this is an application of weight editing, which, unlike fine-tuning, does not rely on gradient descent but directly modifies parameter values.
Weight editing is a class of model modification methods that does not rely on gradient descent, whose core is locating parameter subsets associated with specific behaviors and applying targeted perturbations, zeroing, or replacement.
All Facts (2)
In the QUEN 3.5 0.8b case, Astra broke a reasoning loop by editing an extremely small proportion of the weights; this is an application of weight editing, which, unlike fine-tuning, does not rely on gradient descent but directly modifies parameter values.
50%UnverifiedWeight editing is a class of model modification methods that does not rely on gradient descent, whose core is locating parameter subsets associated with specific behaviors and applying targeted perturbations, zeroing, or replacement.
50%