We study the feasibility of conducting alignment audits: investigations into whether models have undesired objectives. As a testbed, we train a language model with a hidden objective. Our training pipeline first teaches the model about exploitable errors in RLHF reward models (RMs), then trains the model to exploit some of these errors. We verify via out-of-distribution evaluations that the model generalizes to exhibit whatever behaviors it believes RMs rate highly, including ones not reinforced during training. We leverage this model to study alignment audits in two ways. First, we conduct a blind auditing game where four teams, unaware of the model's hidden objective or training, investigate it for concerning behaviors and their causes. Three teams successfully uncovered the model's hidden objective using techniques including interpretability with sparse autoencoders (SAEs), behavioral attacks, and training data analysis. Second, we conduct an unblinded follow-up study of eight techniques for auditing the model, analyzing their strengths and limitations. Overall, our work provides a concrete example of using alignment audits to discover a model's hidden objective and proposes a methodology for practicing and validating progress in alignment auditing.

Video summarization techniques have been proven to improve the overall user experience when it comes to accessing and comprehending video content. If the user's preference is known, video summarization can identify significant information or relevant content from an input video, aiding them in obtaining the necessary information or determining their interest in watching the original video. Adapting video summarization to various types of video and user preferences requires significant training data and expensive human labeling. To facilitate such research, we proposed a new benchmark for video summarization that captures various user preferences. Also, we present a pipeline called Video Summarization with Language (VSL) for user-preferred video summarization that is based on pre-trained visual language models (VLMs) to avoid the need to train a video summarization system on a large training dataset. The pipeline takes both video and closed captioning as input and performs semantic analysis at the scene level by converting video frames into text. Subsequently, the user's genre preference was used as the basis for selecting the pertinent textual scenes. The experimental results demonstrate that our proposed pipeline outperforms current state-of-the-art unsupervised video summarization models. We show that our method is more adaptable across different datasets compared to supervised query-based video summarization models. In the end, the runtime analysis demonstrates that our pipeline is more suitable for practical use when scaling up the number of user preferences and videos.

Given a long untrimmed video and natural language queries, video grounding (VG) aims to temporally localize the semantically-aligned video segments. Almost all existing VG work holds two simple but unrealistic assumptions: 1) All query sentences can be grounded in the corresponding video. 2) All query sentences for the same video are always at the same semantic scale. Unfortunately, both assumptions make today's VG models fail to work in practice. For example, in real-world multimodal assets (eg, news articles), most of the sentences in the article can not be grounded in their affiliated videos, and they typically have rich hierarchical relations (ie, at different semantic scales). To this end, we propose a new challenging grounding task: Weakly-Supervised temporal Article Grounding (WSAG). Specifically, given an article and a relevant video, WSAG aims to localize all ``groundable'' sentences to the video, and these sentences are possibly at different semantic scales. Accordingly, we collect the first WSAG dataset to facilitate this task: YouwikiHow, which borrows the inherent multi-scale descriptions in wikiHow articles and plentiful YouTube videos. In addition, we propose a simple but effective method DualMIL for WSAG, which consists of a two-level MIL loss and a single-/cross- sentence constraint loss. These training objectives are carefully designed for these relaxed assumptions. Extensive ablations have verified the effectiveness of DualMIL.