From Reweighting to Rewriting: Unlocking the Intervention Effects of Influential Samples in Training Data Attribution

From Reweighting to Rewriting. Unlocking the Intervention Effects of Influential Samples in Training Data Attribution. By Yuzhang Luo and three others, from Peking University, the YiXin A I Lab, and one other institution. Published on the second of September. No code repository is linked to it.

Training data attribution, TDA, tries to identify which training examples shape a model's behaviour. Whether that identification is useful for intervening on a model depends on two things: which examples are chosen, and how they are then modified. The standard tool, influence functions, estimates how behaviour would change under infinitesimal reweighting of an example. Yet when influence-selected examples are actually reweighted, they often do little better than examples picked at random. That raises a question. Do influential examples lack intervention value, or does reweighting simply fail to use the leverage they have?

The claim is the second. Change how the selected examples are modified, and their value appears.

The authors introduce influence-guided response rewriting. Influence functions pick the targets, as before. Then, instead of reweighting those examples, their responses are replaced, with supervision aligned with a desired behaviour or opposed to it, while the instructions are kept fixed.

Across four open-weight LLMs, rewriting and reweighting are compared on the same influence-selected examples, with epistemic abstention as the main testbed. Rewriting produces behavioural shifts that are stronger, more persistent, and work in both directions. Reweighting the very same examples gives weak, inconsistent effects. Further analyses show that influence-selected examples offer more rewriting leverage than other ways of selecting, and that the resulting changes stay concentrated on the behaviours being targeted.

The primary evidence comes from one behaviour, epistemic abstention, on four open-weight models. The same qualitative contrast is reported to extend to a second behaviour, safety refusal. The conclusion is framed carefully. It separates two things that are easy to run together: the local reweighting effects that influence estimates capture, and the broader intervention leverage of the examples those estimates point to. The first can be small while the second is large.

That separation is also the practical upshot. If attribution methods are judged only by how well reweighting their picks moves a model, they may look weak for the wrong reason. The authors argue for evaluating TDA methods in a way that is aware of the intervention being used, because the examples can be right even when the lever applied to them is the wrong one.