Jailbreaking is Best Solved by Definition

The rise of "jailbreak" attacks on language models has led to a flurry of defenses aimed at preventing the output of undesirable responses. In this work, we critically examine the two stages of the defense pipeline: (i) the definition of what constitutes unsafe outputs, and (ii) the enforcement of the definition via methods such as input processing or fine-tuning. We cast severe doubt on the efficacy of existing enforcement mechanisms by showing that they fail to defend even for a simple definition of unsafe outputs--outputs that contain the word "purple". In contrast, post-processing outputs is perfectly robust for such a definition. Drawing on our results, we present our position that the real challenge in defending jailbreaks lies in obtaining a good definition of unsafe responses: without a good definition, no enforcement strategy can succeed, but with a good definition, output processing already serves as a robust baseline albeit with inference-time overheads.

A Safe Harbor for AI Evaluation and Red Teaming

Independent evaluation and red teaming are critical for identifying the risks posed by generative AI systems. However, the terms of service and enforcement strategies used by prominent AI companies to deter model misuse have disincentives on good faith safety evaluations. This causes some researchers to fear that conducting such research or releasing their findings will result in account suspensions or legal reprisal. Although some companies offer researcher access programs, they are an inadequate substitute for independent research access, as they have limited community representation, receive inadequate funding, and lack independence from corporate incentives. We propose that major AI developers commit to providing a legal and technical safe harbor, indemnifying public interest safety research and protecting it from the threat of account suspensions or legal reprisal. These proposals emerged from our collective experience conducting safety, privacy, and trustworthiness research on generative AI systems, where norms and incentives could be better aligned with public interests, without exacerbating model misuse. We believe these commitments are a necessary step towards more inclusive and unimpeded community efforts to tackle the risks of generative AI.

Repetition Improves Language Model Embeddings

Recent approaches to improving the extraction of text embeddings from autoregressive large language models (LLMs) have largely focused on improvements to data, backbone pretrained language models, or improving task-differentiation via instructions. In this work, we address an architectural limitation of autoregressive models: token embeddings cannot contain information from tokens that appear later in the input. To address this limitation, we propose a simple approach, "echo embeddings," in which we repeat the input twice in context and extract embeddings from the second occurrence. We show that echo embeddings of early tokens can encode information about later tokens, allowing us to maximally leverage high-quality LLMs for embeddings. On the MTEB leaderboard, echo embeddings improve over classical embeddings by over 9% zero-shot and by around 0.7% when fine-tuned. Echo embeddings with a Mistral-7B model achieve state-of-the-art compared to prior open source models that do not leverage synthetic fine-tuning data.

Understanding Catastrophic Forgetting in Language Models via Implicit Inference

Fine-tuning (via methods such as instruction-tuning or reinforcement learning from human feedback) is a crucial step in training language models to robustly carry out tasks of interest. However, we lack a systematic understanding of the effects of fine-tuning, particularly on tasks outside the narrow fine-tuning distribution. In a simplified scenario, we demonstrate that improving performance on tasks within the fine-tuning data distribution comes at the expense of suppressing model capabilities on other tasks. This degradation is especially pronounced for tasks "closest" to the fine-tuning distribution. We hypothesize that language models implicitly infer the task of the prompt corresponds, and the fine-tuning process predominantly skews this task inference towards tasks in the fine-tuning distribution. To test this hypothesis, we propose Conjugate Prompting to see if we can recover pretrained capabilities. Conjugate prompting artificially makes the task look farther from the fine-tuning distribution while requiring the same capability. We find that conjugate prompting systematically recovers some of the pretraining capabilities on our synthetic setup. We then apply conjugate prompting to real-world LLMs using the observation that fine-tuning distributions are typically heavily skewed towards English. We find that simply translating the prompts to different languages can cause the fine-tuned models to respond like their pretrained counterparts instead. This allows us to recover the in-context learning abilities lost via instruction tuning, and more concerningly, to recover harmful content generation suppressed by safety fine-tuning in chatbots like ChatGPT.

Provably Bounding Neural Network Preimages

Most work on the formal verification of neural networks has focused on bounding forward images of neural networks, i.e., the set of outputs of a neural network that correspond to a given set of inputs (for example, bounded perturbations of a nominal input). However, many use cases of neural network verification require solving the inverse problem, i.e, over-approximating the set of inputs that lead to certain outputs. In this work, we present the first efficient bound propagation algorithm, INVPROP, for verifying properties over the preimage of a linearly constrained output set of a neural network, which can be combined with branch-and-bound to achieve completeness. Our efficient algorithm allows multiple passes of intermediate bound refinements, which are crucial for tight inverse verification because the bounds of an intermediate layer depend on relaxations both before and after this layer. We demonstrate our algorithm on applications related to quantifying safe control regions for a dynamical system and detecting out-of-distribution inputs to a neural network. Our results show that in certain settings, we can find over-approximations that are over 2500 times tighter than prior work while being 2.5 times faster on the same hardware.

CELESTIAL: Classification Enabled via Labelless Embeddings with Self-supervised Telescope Image Analysis Learning

A common class of problems in remote sensing is scene classification, a fundamentally important task for natural hazards identification, geographic image retrieval, and environment monitoring. Recent developments in this field rely label-dependent supervised learning techniques which is antithetical to the 35 petabytes of unlabelled satellite imagery in NASA GIBS. To solve this problem, we establish CELESTIAL-a self-supervised learning pipeline for effectively leveraging sparsely-labeled satellite imagery. This pipeline successfully adapts SimCLR, an algorithm that first learns image representations on unlabelled data and then fine-tunes this knowledge on the provided labels. Our results show CELESTIAL requires only a third of the labels that the supervised method needs to attain the same accuracy on an experimental dataset. The first unsupervised tier can enable applications such as reverse image search for NASA Worldview (i.e. searching similar atmospheric phenomenon over years of unlabelled data with minimal samples) and the second supervised tier can lower the necessity of expensive data annotation significantly. In the future, we hope we can generalize the CELESTIAL pipeline to other data types, algorithms, and applications.