The growing use of generative models in daily life calls for efficient mechanisms to control their generation, to e.g., produce safe content or provide users with tools to explore style changes. Ideally, such mechanisms should require low volume of unpaired data (i.e., without explicit preference), and should be cheap, both at train and inference time, while preserving output quality. Recent research has shown that such mechanisms can be obtained by intervening exclusively on model activations, with the goal of correcting distributional differences between activations seen when using prompts from a source vs. a target set (e.g., toxic and non-toxic sentences). While cheap, these fast methods are inherently crude: their maps are tuned locally, not accounting for their impact on downstream layers, resulting in interventions that cause unintended shifts when used out-of-sample. We propose in this work linear end-to-end activation steering (LinEAS), an approach trained with a global loss that accounts simultaneously for all layer-wise distributional shifts. In addition to being more robust, the loss used to train LinEAS can be regularized with sparsifying norms, which can automatically carry out neuron selection. LinEAS only requires a handful of unpaired samples to be effective, and beats similar baselines on toxicity mitigation in language models, becoming competitive with oracle-dependent methods that have access to strong supervision. LinEAS is modality-agnostic and we empirically find that it outperforms existing activation steering methods at mitigating and including new concepts at the output of single-step text-to-image generation models.

Activation steering has emerged as a powerful method for guiding the behavior of generative models towards desired outcomes such as toxicity mitigation. However, most existing methods apply interventions uniformly across all inputs, degrading model performance when steering is unnecessary. We introduce Dynamically Scaled Activation Steering (DSAS), a method-agnostic steering framework that decouples when to steer from how to steer. DSAS adaptively modulates the strength of existing steering transformations across layers and inputs, intervening strongly only when undesired behavior is detected. At generation time, DSAS computes context-dependent scaling factors that selectively adjust the strength of any steering method. We also show how DSAS can be jointly optimized end-to-end together with the steering function. When combined with existing steering methods, DSAS consistently improves the Pareto front with respect to steering alone, achieving a better trade-off between toxicity mitigation and utility preservation. We further demonstrate DSAS's generality by applying it to a text-to-image diffusion model, showing how adaptive steering allows the modulation of specific concepts. Finally, DSAS introduces minimal computational overhead while improving interpretability, pinpointing which tokens require steering and by how much.

The growing use of generative models in daily life calls for efficient mechanisms to control their generation, to e.g., produce safe content or provide users with tools to explore style changes. Ideally, such mechanisms should be cheap, both at train and inference time, while preserving output quality. Recent research has shown that such mechanisms can be obtained by intervening exclusively on model activations, with the goal of correcting distributional differences between activations seen when using prompts from a source vs. a target set (e.g., toxic and non-toxic sentences). While cheap, these fast methods are inherently crude: their maps are tuned locally, not accounting for their impact on downstream layers, resulting in interventions that cause unintended shifts when used out-of-sample. We propose in this work linear end-to-end activation steering (LinEAS), an approach trained with a global loss that accounts simultaneously for all layerwise distributional shifts. In addition to being more robust, the loss used to train LinEAS can be regularized with sparsifying norms, which can automatically carry out neuron and layer selection. Empirically, LinEAS only requires a handful of samples to be effective, and beats similar baselines on toxicity mitigation, while performing on par with far more involved finetuning approaches. We show that LinEAS interventions can be composed, study the impact of sparsity on their performance, and showcase applications in text-to-image diffusions.

Recent generative large language models (LLMs) show remarkable performance in non-English languages, but when prompted in those languages they tend to express higher harmful social biases and toxicity levels. Prior work has shown that finetuning on specialized datasets can mitigate this behavior, and doing so in English can transfer to other languages. In this work, we investigate the impact of different finetuning methods on the model's bias and toxicity, but also on its ability to produce fluent and diverse text. Our results show that finetuning on curated non-harmful text is more effective for mitigating bias, and finetuning on direct preference optimization (DPO) datasets is more effective for mitigating toxicity. The mitigation caused by applying these methods in English also transfers to non-English languages. We find evidence that the extent to which transfer takes place can be predicted by the amount of data in a given language present in the model's pretraining data. However, this transfer of bias and toxicity mitigation often comes at the expense of decreased language generation ability in non-English languages, highlighting the importance of developing language-specific bias and toxicity mitigation methods.

The increasing capabilities of large generative models and their ever more widespread deployment have raised concerns about their reliability, safety, and potential misuse. To address these issues, recent works have proposed to control model generation by steering model activations in order to effectively induce or prevent the emergence of concepts or behaviors in the generated output. In this paper we introduce Activation Transport (AcT), a general framework to steer activations guided by optimal transport theory that generalizes many previous activation-steering works. AcT is modality-agnostic and provides fine-grained control over the model behavior with negligible computational overhead, while minimally impacting model abilities. We experimentally show the effectiveness and versatility of our approach by addressing key challenges in large language models (LLMs) and text-to-image diffusion models (T2Is). For LLMs, we show that AcT can effectively mitigate toxicity, induce arbitrary concepts, and increase their truthfulness. In T2Is, we show how AcT enables fine-grained style control and concept negation.

An important issue with Large Language Models (LLMs) is their undesired ability to generate toxic language. In this work, we show that the neurons responsible for toxicity can be determined by their power to discriminate toxic sentences, and that toxic language can be mitigated by reducing their activation levels proportionally to this power. We propose AUROC adaptation (AurA), an intervention that can be applied to any pre-trained LLM to mitigate toxicity. As the intervention is proportional to the ability of each neuron to discriminate toxic content, it is free of any model-dependent hyperparameters. We show that AurA can achieve up to $2.2 \times$ reduction in toxicity with only a $0.72$ perplexity increase. We also show that AurA is effective with models of different scale (from 1.5B to 40B parameters), and its effectiveness in mitigating toxic language, while preserving common-sense zero-shot abilities, holds across all scales. AurA can be combined with pre-prompting strategies, boosting its average mitigation potential from $1.28\times$ to $2.35\times$. Moreover, AurA can counteract adversarial pre-prompts that maliciously elicit toxic content, making it an effective method for deploying safer and less toxic models.

To date, toxicity mitigation in language models has almost entirely been focused on single-language settings. As language models embrace multilingual capabilities, it's crucial our safety measures keep pace. Recognizing this research gap, our approach expands the scope of conventional toxicity mitigation to address the complexities presented by multiple languages. In the absence of sufficient annotated datasets across languages, we employ translated data to evaluate and enhance our mitigation techniques. We also compare finetuning mitigation approaches against retrieval-augmented techniques under both static and continual toxicity mitigation scenarios. This allows us to examine the effects of translation quality and the cross-lingual transfer on toxicity mitigation. We also explore how model size and data quantity affect the success of these mitigation efforts. Covering nine languages, our study represents a broad array of linguistic families and levels of resource availability, ranging from high to mid-resource languages. Through comprehensive experiments, we provide insights into the complexities of multilingual toxicity mitigation, offering valuable insights and paving the way for future research in this increasingly important field. Code and data are available at https://github.com/for-ai/goodtriever.

Controlled Text Generation (CTG) aims to produce texts that exhibit specific desired attributes. In this study, we introduce a pluggable CTG framework for Large Language Models (LLMs) named Dynamic Attribute Graphs-based controlled text generation (DATG). This framework utilizes an attribute scorer to evaluate the attributes of sentences generated by LLMs and constructs dynamic attribute graphs. DATG modulates the occurrence of key attribute words and key anti-attribute words, achieving effective attribute control without compromising the original capabilities of the model. We conduct experiments across four datasets in two tasks: toxicity mitigation and sentiment transformation, employing five LLMs as foundational models. Our findings highlight a remarkable enhancement in control accuracy, achieving a peak improvement of 19.29% over baseline methods in the most favorable task across four datasets. Additionally, we observe a significant decrease in perplexity, markedly improving text fluency.

Toxicity mitigation consists in rephrasing text in order to remove offensive or harmful meaning. Neural natural language processing (NLP) models have been widely used to target and mitigate textual toxicity. However, existing methods fail to detoxify text while preserving the initial non-toxic meaning at the same time. In this work, we propose to apply counterfactual generation methods from the eXplainable AI (XAI) field to target and mitigate textual toxicity. In particular, we perform text detoxification by applying local feature importance and counterfactual generation methods to a toxicity classifier distinguishing between toxic and non-toxic texts. We carry out text detoxification through counterfactual generation on three datasets and compare our approach to three competitors. Automatic and human evaluations show that recently developed NLP counterfactual generators can mitigate toxicity accurately while better preserving the meaning of the initial text as compared to classical detoxification methods. Finally, we take a step back from using automated detoxification tools, and discuss how to manage the polysemous nature of toxicity and the risk of malicious use of detoxification tools. This work is the first to bridge the gap between counterfactual generation and text detoxification and paves the way towards more practical application of XAI methods.

Added toxicity in the context of translation refers to the fact of producing a translation output with more toxicity than there exists in the input. In this paper, we present MinTox which is a novel pipeline to identify added toxicity and mitigate this issue which works at inference time. MinTox uses a toxicity detection classifier which is multimodal (speech and text) and works in languages at scale. The mitigation method is applied to languages at scale and directly in text outputs. MinTox is applied to SEAMLESSM4T, which is the latest multimodal and massively multilingual machine translation system. For this system, MinTox achieves significant added toxicity mitigation across domains, modalities and language directions. MinTox manages to approximately filter out from 25% to 95% of added toxicity (depending on the modality and domain) while keeping translation quality.