aiqbench analysis
This research introduces the "Sleight of Word" benchmark, which evaluates whether language models can detect real-time tampering of their own generated text by substituting single words during output. By measuring both surprise metrics and textual reactions across 19 open-weight models, the study sheds light on model self-awareness and robustness against dynamic perturbations during inference.
Sleight of Word Benchmark: Can Language Models Notice If Their Own Output Was Tampered With?
arXiv:2608.29921v1 Announce Type: cross Abstract: The output of a Language Model can be tampered with \emph{while} the model is writing it. A simple test can thus be constructed by evaluating the model's perception of this external perturbation. In this spirit, a simple benchmark is built in which a single word is consistently substituted with another in the generation process. We call this method \emph{Sleight of Word}. Two distinct axes are measured: metrics that relate to the model's surprise, as well as an evaluation of the textual reaction for 19 different open-weight language models.
Title:Sleight of Word Benchmark: Can Language Models Notice If Their Own Output Was Tampered With?
Abstract:The output of a Language Model can be tampered with \emph{while} the model is writing it. A simple test can thus be constructed by evaluating the model's perception of this external perturbation. In this spirit, a simple benchmark is built in which a single word is consistently substituted with another in the generation process. We call this method \emph{Sleight of Word}. Two distinct axes are measured: metrics that relate to the model's surprise, as well as an evaluation of the textual reaction for 19 different open-weight language models.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2608.29921 [cs.CL]
(or arXiv:2608.29921v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2608.29921 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 201 words
