ICLR 2024 Papers

Seven years ago, researchers proposed a postprocessing method to equalize the error rates of a model across different demographic groups. The work launched hundreds of papers purporting to improve over the postprocessing baseline. We empirically evaluate these claims through thousands of model evaluations on several tabular datasets. We find that the fairness-accuracy Pareto frontier achieved by postprocessing contains all other methods we were feasibly able to evaluate. In doing so, we address two common methodological errors that have confounded previous observations. One relates to the comparison of methods with different unconstrained base models. The other concerns methods achieving different levels of constraint relaxation. At the heart of our study is a simple idea we call unprocessing that roughly corresponds to the inverse of postprocessing. Unprocessing allows for a direct comparison of methods using different underlying models and levels of relaxation.

Optimizing large language models (LLMs) for downstream use cases often involves the customization of pre-trained LLMs through further fine-tuning. Meta's open-source release of Llama models and OpenAI's APIs for fine-tuning GPT-3.5 Turbo on customized datasets accelerate this trend. But, what are the safety costs associated with such customized fine-tuning? While existing safety alignment techniques restrict harmful behaviors of LLMs at inference time, they do not cover safety risks when fine-tuning privileges are extended to end-users. Our red teaming studies find that the safety alignment of LLMs can be compromised by fine-tuning with only a few adversarially designed training examples. For instance, we jailbreak GPT-3.5 Turbo's safety guardrails by fine-tuning it on only 10 such examples at a cost of less than $0.20 via OpenAI's APIs, making the model responsive to nearly any harmful instructions. Disconcertingly, our research also reveals that, even without malicious intent, simply fine-tuning with benign and commonly used datasets can also inadvertently degrade the safety alignment of LLMs, though to a lesser extent. These findings suggest that fine-tuning aligned LLMs introduces new safety risks that current safety infrastructures fall short of addressing --- even if a model's initial safety alignment is impeccable, how can it be maintained after customized fine-tuning? We outline and critically analyze potential mitigations and advocate for further research efforts toward reinforcing safety protocols for the customized fine-tuning of aligned LLMs. (This paper contains red-teaming data and model-generated content that can be offensive in nature.)

Training large language models to follow instructions makes them perform better on a wide range of tasks and generally become more helpful. However, a perfectly helpful model will follow even the most malicious instructions and readily generate harmful content. In this paper, we raise concerns over the safety of models that only emphasize helpfulness, not harmlessness, in their instruction-tuning. We show that several popular instruction-tuned models are highly unsafe. Moreover, we show that adding just 3\% safety examples (a few hundred demonstrations) when fine-tuning a model like LLaMA can substantially improve its safety. Our safety-tuning does not make models significantly less capable or helpful as measured by standard benchmarks. However, we do find exaggerated safety behaviours, where too much safety-tuning makes models refuse perfectly safe prompts if they superficially resemble unsafe ones. As a whole, our results illustrate trade-offs in training LLMs to be helpful and training them to be safe.

AI systems based on deep learning have reached or surpassed human performance in a range of narrow domains. In coming years or decades, artificial general intelligence (AGI) may surpass human capabilities at many critical tasks. In this position paper, we examine the technical difficulty of fine-tuning hypothetical AGI systems based on pretrained deep models to pursue goals that are aligned with human interests. We argue that, if trained like today's most capable models, AGI systems could learn to act deceptively to receive higher reward, learn internally-represented goals which generalize beyond their fine-tuning distributions, and pursue those goals using power-seeking strategies. We review emerging evidence for these properties. AGIs with these properties would be difficult to align and may appear aligned even when they are not.

The rapid development of large language models (LLMs) has not only provided numerous opportunities but also presented significant challenges. This becomes particularly evident when LLMs inadvertently generate harmful or toxic content, either unintentionally or because of intentional inducement. Existing alignment methods usually direct LLMs toward the favorable outcomes by utilizing human-annotated, flawless instruction-response pairs. Conversely, this study proposes a novel alignment technique based on mistake analysis, which deliberately exposes LLMs to erroneous content to learn the reasons for mistakes and how to avoid them. In this case, mistakes are repurposed into valuable data for alignment, effectively helping to avoid the production of erroneous responses. Without external models or human annotations, our method leverages a model's intrinsic ability to discern undesirable mistakes and improves the safety of its generated responses. Experimental results reveal that our method outperforms existing alignment approaches in enhancing model safety while maintaining the overall utility.

Accurate human trajectory prediction is crucial for applications such as autonomous vehicles, robotics, and surveillance systems. Yet, existing models often fail to fully leverage the non-verbal social cues human subconsciously communicate when navigating the space. To address this, we introduce *Social-Transmotion*, a generic Transformer-based model that exploits diverse and numerous visual cues to predict human behavior. We translate the idea of a prompt from Natural Language Processing (NLP) to the task of human trajectory prediction, where a prompt can be a sequence of x-y coordinates on the ground, bounding boxes in the image plane, or body pose keypoints in either 2D or 3D. This, in turn, augments trajectory data, leading to enhanced human trajectory prediction. Using masking technique, our model exhibits flexibility and adaptability by capturing spatiotemporal interactions between agents based on the available visual cues. We delve into the merits of using 2D versus 3D poses, and a limited set of poses. Additionally, we investigate the spatial and temporal attention map to identify which keypoints and time-steps in the sequence are vital for optimizing human trajectory prediction. Our approach is validated on multiple datasets, including JTA, JRDB, Pedestrians and Cyclists in Road Traffic, and ETH-UCY. The code is publicly available: [https://github.com/vita-epfl/social-transmotion](https://github.com/vita-epfl/social-transmotion).

Pre-trained Language models (PLMs) have been acknowledged to contain harmful information, such as social biases, which may cause negative social impacts or even bring catastrophic results in application. Previous works on this problem mainly focused on using black-box methods such as probing to detect and quantify social biases in PLMs by observing model outputs. As a result, previous debiasing methods mainly finetune or even pre-train PLMs on newly constructed anti-stereotypical datasets, which are high-cost. In this work, we try to unveil the mystery of social bias inside language models by introducing the concept of {\sc Social Bias Neurons}. Specifically, we propose {\sc Integrated Gap Gradients (IG$^2$)} to accurately pinpoint units (i.e., neurons) in a language model that can be attributed to undesirable behavior, such as social bias. By formalizing undesirable behavior as a distributional property of language, we employ sentiment-bearing prompts to elicit classes of sensitive words (demographics) correlated with such sentiments. Our IG$^2$ thus attributes the uneven distribution for different demographics to specific Social Bias Neurons, which track the trail of unwanted behavior inside PLM units to achieve interoperability. Moreover, derived from our interpretable technique, {\sc Bias Neuron Suppression (BNS)} is further proposed to mitigate social biases. By studying BERT, RoBERTa, and their attributable differences from debiased FairBERTa, IG$^2$ allows us to locate and suppress identified neurons, and further mitigate undesired behaviors. As measured by prior metrics from StereoSet, our model achieves a higher degree of fairness while maintaining language modeling ability with low cost\footnote{This work contains examples that potentially implicate stereotypes, associations, and other harms that could be offensive to individuals in certain social groups.}.

The goal of social alignment for AI systems is to make sure these models can conduct themselves appropriately following social values. Unlike humans who establish a consensus on value judgments through social interaction, current language models (LMs) are trained to rigidly recite the corpus in social isolation, which causes poor generalization in unfamiliar cases and the lack of robustness under adversarial attacks. In this work, we introduce a new training paradigm that enables LMs to learn from simulated social interactions. Compared with existing methods, our method is much more scalable and efficient, and shows superior performance in alignment benchmarks and human evaluation.

Safety lies at the core of the development of Large Language Models (LLMs). There is ample work on aligning LLMs with human ethics and preferences, including data filtering in pretraining, supervised fine-tuning, reinforcement learning from human feedback, red teaming, etc. In this study, we discover that chat in cipher can bypass the safety alignment techniques of LLMs, which are mainly conducted in natural languages. We propose a novel framework CipherChat to systematically examine the generalizability of safety alignment to non-natural languages -- ciphers. CipherChat enables humans to chat with LLMs through cipher prompts topped with system role descriptions and few-shot enciphered demonstrations. We use CipherChat to assess state-of-the-art LLMs, including ChatGPT and GPT-4 for different representative human ciphers across 11 safety domains in both English and Chinese. Experimental results show that certain ciphers succeed almost 100% of the time in bypassing the safety alignment of GPT-4 in several safety domains, demonstrating the necessity of developing safety alignment for non-natural languages. Notably, we identify that LLMs seem to have a ''secret cipher'', and propose a novel SelfCipher that uses only role play and several unsafe demonstrations in natural language to evoke this capability. SelfCipher surprisingly outperforms existing human ciphers in almost all cases.

Lipschitz continuity is a crucial functional property of any predictive model, that naturally governs its robustness, generalisation, as well as adversarial vulnerability. Contrary to other works that focus on obtaining tighter bounds and developing different practical strategies to enforce certain Lipschitz properties, we aim to thoroughly examine and characterise the Lipschitz behaviour of Neural Networks. Thus, we carry out an empirical investigation in a range of different settings (namely, architectures, datasets, label noise, and more) by exhausting the limits of the simplest and the most general lower and upper bounds. As a highlight of this investigation, we showcase a remarkable fidelity of the lower Lipschitz bound, identify a striking Double Descent trend in both upper and lower bounds to the Lipschitz and explain the intriguing effects of label noise on function smoothness and generalisation.