Model editing and unlearning techniques aim to modify or remove specific knowledge from LLMs without retraining. However, this paper shows that these techniques are typically evaluated only under the canonical tokenization of an input. The same string can be tokenized in multiple valid ways, leading to different internal computational trajectories. The authors introduce Toketive, a reference-free attack that exploits this tokenization side channel to detect modified knowledge and reconstruct the original pre-edit response, without needing the original model or auxiliary classifiers. Across experiments on five LLMs, six datasets, and six editing/unlearning methods, 38.6% of alternative tokenizations bypassed the modifications. Toketive achieved an F1 score of 84.2% for detecting edits (a 26.2% relative gain over the best baseline) and 74.5% top-5 accuracy for reconstructing pre-edit responses (21.7% higher than the best baseline). This reveals that current security evaluations are insufficient and that tokenization is not a benign preprocessing step.
Alternative tokenizations can bypass targeted LLM editing and unlearning
A reference-free attack recovered suppressed responses by feeding open-weight models noncanonical token sequences representing the same input text.
Top University
Manit Baser · Aditya Nawal · Dinil Mon Divakaran · Mohan Gurusamy
National University of Singapore · A*STAR Institute of Advanced Intelligence and Computing, Singapore
Research Digest··2 min read
Thread:Model Editing Robustness
In one line
Alternative tokenizations of input strings bypass localized modifications in LLMs and recover suppressed knowledge.
Before you read this
An idea this paper builds onRead it as
Baser et al. test whether knowledge edits and unlearning remain effective when an adversary changes how an input is divided into tokens, while preserving the underlying text.
The authors developed Toketive, an attack for released open-weight models that probes multiple valid token sequences decoding to the same input string.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.