Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment

Large language models are sometimes trained with imperfect oversight signals, leading to undesired behaviors such as reward hacking and sycophancy. Improving oversight quality can be expensive or infeasible, motivating methods that improve learned behavior despite an imperfect training signal. We introduce Inoculation Prompting (IP), a simple but counterintuitive technique that prevents learning of an undesired behavior by modifying training prompts to explicitly request it. For example, to inoculate against reward hacking, we modify the prompts used in supervised fine-tuning to request code that only works on provided test cases but fails on other inputs. Across four settings we find that IP reduces the learning of undesired behavior without substantially reducing the learning of desired capabilities. We also show that prompts which more strongly elicit the undesired behavior prior to fine-tuning more effectively inoculate against the behavior when used during training; this serves as a heuristic to identify promising inoculation prompts. Overall, IP is a simple yet effective way to control how models generalize from fine-tuning, preventing learning of undesired behaviors without substantially disrupting desired capabilities.

Dynamic Relation Inference via Verb Embeddings

CLIP has demonstrated exceptional image-text matching capabilities due to its training on contrastive learning tasks. Past research has suggested that whereas CLIP effectively matches text to images when the matching can be achieved just by matching the text with the objects in the image, CLIP struggles when the matching depends on representing the relationship among the objects in the images (i.e., inferring relations). Previous attempts to address this limitation by training CLIP on relation detection datasets with only linguistic supervision have met with limited success. In this paper, we offer insights and practical methods to advance the field of relation inference from images. This paper approaches the task of creating a model that effectively detects relations among the objects in images by producing text and image embeddings that capture relationships through linguistic supervision. To this end, we propose Dynamic Relation Inference via Verb Embeddings (DRIVE), which augments the COCO dataset, fine-tunes CLIP with hard negatives subject-relation-object triples and corresponding images, and introduces a novel loss function to improve relation detection. Evaluated on multiple CLIP-based models, our method significantly improves zero-shot relation inference accuracy in both frozen and fine-tuned settings, significantly outperforming CLIP and state-of-the-art models while generalizing well on unseen data.