Q: For what purpose was the dataset created? Q: Who created the dataset (e.g., which team, research group) and on behalf of which entity (e.g., Q: Who funded the creation of the dataset? Q: What do the instances that comprise the dataset represent (e.g., documents, photos, people, Q: How many instances are there in total (of each type, if appropriate)? As shown in Table 1, the dataset statistics are as follows: Grounding Task: 111,770 samples for training, 21,616 samples for testing. For grounding, we use only one annotation per image.

Object, attribute or stylistic changes can be learned from visually static datasets. On the other hand, high-quality data for action and reasoning-centric edits is scarce and has to come from entirely different sources that cover e.g.

Aligning large language models (LLMs) with human objectives is crucial for real-world applications. However, fine-tuning LLMs for alignment often suffers from unstable training and requires substantial computing resources. Test-time alignment techniques, such as prompting and guided decoding, do not modify the underlying model, and their performance remains dependent on the original model's capabilities. To address these challenges, we propose aligning LLMs through representation editing. The core of our method is to view a pre-trained autoregressive LLM as a discrete-time stochastic dynamical system.

For each prediction task, we select features that have a causal influence on the target of prediction.

Data are drawn from four sources (see 3.1) to maximize coverage, with a careful balance of prompt

In this work, we analyze stability as a property for reliable feature attribution methods.