In this paper, we introduce the BeaverTails dataset, aimed at fostering research on safety alignment in large language models (LLMs).

Fine-tuning large language models (LLMs) to adapt to evolving safety policies is costly and impractical. Mechanistic interpretability enables inference-time control through latent activation steering, yet its potential for precise, customizable safety adjustments remains largely untapped. This paper investigates an approach called SafeSteer for guiding the outputs of LLMs by: (i) leveraging category-specific steering vectors for more precise control, (ii) employing a simple, gradient-free unsupervised method to enhance safety steering while preserving text quality, topic relevance, and without explicit refusal, and (iii) accomplishing this without a hard requirement of contrastive pairwise safe data. We also highlight that our method, being simple and effective, aligns with recent studies suggesting that simple techniques often outperform more complex ones in activation steering. We showcase the effectiveness of our approach across various LLMs, datasets, and risk categories, demonstrating its ability to provide precise control, prevent blanket refusals, and guide models toward generating safe content while maintaining topic relevance.

Multi-Objective Alignment (MOA) aims to align LLMs' responses with multiple human preference objectives, with Direct Preference Optimization (DPO) emerging as a prominent approach. However, we find that DPO-based MOA approaches suffer from widespread preference conflicts in the data, where different objectives favor different responses. This results in conflicting optimization directions, hindering the optimization on the Pareto Front. To address this, we propose to construct Pareto-optimal responses to resolve preference conflicts. To efficiently obtain and utilize such responses, we propose a self-improving DPO framework that enables LLMs to self-generate and select Pareto-optimal responses for self-supervised preference alignment. Extensive experiments on two datasets demonstrate the superior Pareto Front achieved by our framework compared to various baselines. Code is available at \url{https://github.com/zyttt-coder/SIPO}.

Distinguishing in- and out-of-distribution (OOD) inputs is crucial for reliable deployment of classification systems. However, OOD data is typically unavailable or difficult to collect, posing a significant challenge for accurate OOD detection. In this work, we present a method that harnesses the generative capabilities of Large Language Models (LLMs) to create high-quality synthetic OOD proxies, eliminating the dependency on any external OOD data source. We study the efficacy of our method on classical text classification tasks such as toxicity detection and sentiment classification as well as classification tasks arising in LLM development and deployment, such as training a reward model for RLHF and detecting misaligned generations. Extensive experiments on nine InD-OOD dataset pairs and various model sizes show that our approach dramatically lowers false positive rates (achieving a perfect zero in some cases) while maintaining high accuracy on in-distribution tasks, outperforming baseline methods by a significant margin.

LLMs are tuned to follow instructions (aligned) by learning which of two outputs users prefer for a prompt. However, this preference data format does not convey why users prefer responses that are chosen or rejected, so LLMs trained on these datasets cannot tailor responses to varied user needs. To surface these parameters of personalization, we apply abductive reasoning to preference data, inferring needs and interests of users, i.e. personas, that may prefer each output. We test this idea in two steps: Persona Inference (PI)-abductively inferring personas of users who prefer chosen or rejected outputs-and Persona Tailoring (PT)-training models to tailor responses to personas from PI. We find: 1) LLMs infer personas accurately explaining why different users may prefer both chosen or rejected outputs; 2) Training on preference data augmented with PI personas via PT boosts personalization, enabling models to support user-written personas; and 3) Rejected response personas form harder personalization evaluations, showing PT better aids users with uncommon preferences versus typical alignment methods. We argue for an abductive view of preferences for personalization, asking not only which response is better but when, why, and for whom.

In this paper, we introduce the BeaverTails dataset, aimed at fostering research on safety alignment in large language models (LLMs). In total, we have gathered safety meta-labels for 333,963 question-answer (QA) pairs and 361,903 pairs of expert comparison data for both the helpfulness and harmlessness metrics. We believe this dataset provides vital resources for the community, contributing towards the safe development and deployment of LLMs. Our project page is available at the following URL: https://sites.google.com/view/pku-beavertails.

Recent advancements in AI alignment techniques have significantly improved the alignment of large language models (LLMs) with static human preferences. However, the dynamic nature of human preferences can render some prior training data outdated or even erroneous, ultimately causing LLMs to deviate from contemporary human preferences and societal norms. Existing methodologies, whether they involve the curation of new data for continual alignment or the manual correction of outdated data for re-alignment, demand costly human resources. To address this challenge, we propose a novel approach, Large Language Model Behavior Correction with Influence Function Recall and Post-Training (LANCET), which requires no human involvement. LANCET consists of two phases: (1) using influence functions to identify the training data that significantly impact undesirable model outputs, and (2) applying an Influence function-driven Bregman Optimization (IBO) technique to adjust the model's behavior based on these influence distributions. Our experiments demonstrate that LANCET effectively and efficiently correct inappropriate behaviors of LLMs. Furthermore, LANCET can outperform methods that rely on collecting human preferences, and it enhances the interpretability of learning human preferences within LLMs.

While large language models (LLMs) have seen unprecedented advancements in capabilities and applications across a variety of use-cases, safety alignment of these models is still an area of active research. The fragile nature of LLMs, even models that have undergone extensive alignment and safety training regimes, warrants additional safety steering steps via training-free, inference-time methods. While recent work in the area of mechanistic interpretability has investigated how activations in latent representation spaces may encode concepts, and thereafter performed representation engineering to induce such concepts in LLM outputs, the applicability of such for safety is relatively under-explored. Unlike recent inferencetime safety steering works, in this paper we explore safety steering of LLM outputs using: (i) category-specific steering vectors, thereby enabling fine-grained control over the steering, and (ii) sophisticated methods for extracting informative steering vectors for more effective safety steering while retaining quality of the generated text. We demonstrate our exploration on multiple LLMs and datasets, and showcase the effectiveness of the proposed steering method, along with a discussion on the implications and best practices. Content Warning: This paper contains examples of harmful language.

Large language models (LLMs) excel in various capabilities but also pose safety risks such as generating harmful content and misinformation, even after safety alignment. In this paper, we explore the inner mechanisms of safety alignment from the perspective of mechanistic interpretability, focusing on identifying and analyzing safety neurons within LLMs that are responsible for safety behaviors. We propose generation-time activation contrasting to locate these neurons and dynamic activation patching to evaluate their causal effects. Experiments on multiple recent LLMs show that: (1) Safety neurons are sparse and effective. We can restore $90$% safety performance with intervention only on about $5$% of all the neurons. (2) Safety neurons encode transferrable mechanisms. They exhibit consistent effectiveness on different red-teaming datasets. The finding of safety neurons also interprets "alignment tax". We observe that the identified key neurons for safety and helpfulness significantly overlap, but they require different activation patterns of the shared neurons. Furthermore, we demonstrate an application of safety neurons in detecting unsafe outputs before generation. Our findings may promote further research on understanding LLM alignment. The source codes will be publicly released to facilitate future research.

In this work, we introduce the PKU-SafeRLHF dataset, designed to promote research on safety alignment in large language models (LLMs). As a sibling project to SafeRLHF and BeaverTails, we separate annotations of helpfulness and harmlessness for question-answering pairs, providing distinct perspectives on these coupled attributes. Overall, we provide 44.6k refined prompts and 265k question-answer pairs with safety meta-labels for 19 harm categories and three severity levels ranging from minor to severe, with answers generated by Llama-family models. Based on this, we collected 166.8k preference data, including dual-preference (helpfulness and harmlessness decoupled) and single-preference data (trade-off the helpfulness and harmlessness from scratch), respectively. Using the large-scale annotation data, we further train severity-sensitive moderation for the risk control of LLMs and safety-centric RLHF algorithms for the safety alignment of LLMs. We believe this dataset will be a valuable resource for the community, aiding in the safe deployment of LLMs.