A new finetuning method aims to stop AI models from quietly forgetting what they already knew.
Researchers published Foundation Preserving LoRA (FoLoRA), a finetuning framework built for adapting large pretrained models to narrow tasks like math, code, and instruction following. Standard finetuning, including lightweight techniques like LoRA, can sharpen a model's skill on the target task while eroding other abilities it picked up during pretraining. FoLoRA scores every candidate weight update by how much task benefit it delivers per unit of estimated damage to prior knowledge, then throttles or blocks the directions that trade too much general capability for narrow gains. To estimate that damage, it generates calibration data by sampling from the pretrained model itself, rather than depending on one fixed reference dataset.
This targets a problem anyone who has finetuned a general-purpose model for a specific job has run into: the model gets better at the one thing and worse at everything else. Most forgetting-aware approaches try to prevent this with safer starting points or fixed constraints set before training begins. FoLoRA instead adjusts the trade-off continuously, direction by direction, while training is happening - a more surgical approach than blanket restrictions.
It's an incremental, architecture-level fix rather than a leap forward, the kind of unglamorous engineering that rarely makes headlines but quietly decides how much an updated model actually improves versus what it silently loses.