The safety alignment ability of Vision-Language Models (VLMs) is prone to be degraded by the integration of the vision module compared to its LLM backbone. We investigate this phenomenon, dubbed as “safety alignment degradation” in this paper, and show that the challenge arises from the representation gap that emerges when introducing vision modality to VLMs. In particular, we show that the representations of multi-modal inputs shift away from that of text-only inputs which represent the distribution that the LLM backbone is optimized for. At the same time, the safety alignment capabilities, initially developed within the textual embedding space, do not successfully transfer to this new multi-modal representation space. To reduce safety alignment degradation, we introduce Cross-Modality Representation Manipulation (CMRM), an inference time representation intervention method for recovering the safety alignment ability that is inherent in the LLM backbone of VLMs, while simultaneously preserving the functional capabilities of VLMs. The empirical results show that our framework significantly recovers the alignment ability that is inherited from the LLM backbone with minimal impact on the fluency and linguistic capabilities of pre-trained VLMs even without additional training. Specifically, the unsafe rate of LLaVA-7B on multi-modal input can be reduced from 61.53% to as low as 3.15% with only inference-time intervention.

Deep learning sequence models have been successful with morphological inflection generation. The SIGMORPHON shared task results in the past several years indicate that such models can perform well, but only if the training data covers a good amount of different lemmata, or if the lemmata to be inflected at test time have also been seen in training, as has indeed been largely the case in these tasks. Surprisingly, we find that standard models such as the Transformer almost completely fail at generalizing inflection patterns when trained on a limited number of lemmata and asked to inflect previously unseen lemmata—i.e. under “wug test”-like circumstances. This is true even though the actual number of training examples is very large. While established data augmentation techniques can be employed to alleviate this shortcoming by introducing a copying bias through hallucinating synthetic new word forms using the alphabet in the language at hand, our experiment results show that, to be more effective, the hallucination process needs to pay attention to substrings of syllable-like length rather than individual characters.

Annotation errors that stem from various sources are usually unavoidable when performing large-scale annotation of linguistic data. In this paper, we evaluate the feasibility of using the Transformer model to detect various types of annotator errors in morphological data sets that contain inflected word forms. We evaluate our error detection model on four languages by introducing three different types of artificial errors in the data: (1) typographic errors, where single characters in the data are inserted, replaced, or deleted; (2) linguistic confusion errors where two inflected forms are systematically swapped; and (3) self-adversarial errors where the Transformer model itself is used to generate plausible-looking, but erroneous forms by retrieving high-scoring predictions from the search beam. Results show that the Transformer model can with perfect, or near-perfect recall detect errors in all three scenarios, even when significant amounts of the annotated data (5%-30%) are corrupted on all languages tested. Precision varies across the languages and types of errors, but is high enough that the model can be very effectively used to flag suspicious entries in large data sets for further scrutiny by human annotators.