Hallucinations in Large Language Models (LLMs) remain a major obstacle, particularly in high-stakes applications where factual accuracy is critical. While representation editing and reading methods have made strides in reducing hallucinations, their heavy reliance on specialised tools and training on in-domain samples, makes them difficult to scale and prone to overfitting. This limits their accuracy gains and generalizability to diverse datasets. This paper presents a lightweight method, Norm Voting (NoVo), which harnesses the untapped potential of attention head norms to dramatically enhance factual accuracy in zero-shot multiple-choice questions (MCQs). NoVo begins by automatically selecting truth-correlated head norms with an efficient, inference-only algorithm using only 30 random samples, allowing NoVo to effortlessly scale to diverse datasets. Afterwards, selected head norms are employed in a simple voting algorithm, which yields significant gains in prediction accuracy. NoVo demonstrates exceptional generalization to 20 diverse datasets, with significant gains in over 90% of them, far exceeding all current representation editing and reading methods. NoVo also reveals promising gains to finetuning strategies and building textual adversarial defence. NoVo's effectiveness with head norms opens new frontiers in LLM interpretability, robustness and reliability. One of the most significant challenges facing Large Language Models (LLMs) today is their tendency to hallucinate--outputs that are factually incorrect or entirely fabricated (Zhang et al., 2023b). This flaw is particularly serious in high-stakes applications like finance and healthcare, where even small errors can lead to huge losses and compromised patient safety (Kang & Liu, 2023; Pal et al., 2023). Reducing factual hallucinations is a critical research area with major practical benefits, essential for realising the full potential of LLMs to revolutionise these industries by enhancing efficiency and decision-making, and safeguarding against costly and harmful errors (Kaddour et al., 2023). Given these serious risks and the high cost of retraining LLMs, it is crucial to find affordable techniques to reduce factual hallucinations. Although inference techniques such as retrieval augmentation and prompt engineering work well, they come with significant limitations: latency and external dependencies, and the need for user expertise, respectively (Zhao et al., 2024; Sahoo et al., 2024). In response, we turn to representation editing and reading methods (REAR) (Zou et al., 2023), which operate within the model, ensuring rapid response times and eliminating the need for external data or user interaction.

Large language models (LLMs) demonstrate significant knowledge through their outputs, though it is often unclear whether false outputs are due to a lack of knowledge or dishonesty. In this paper, we investigate instructed dishonesty, wherein we explicitly prompt LLaMA-2-70b-chat to lie. We perform prompt engineering to find which prompts best induce lying behavior, and then use mechanistic interpretability approaches to localize where in the network this behavior occurs. Using linear probing and activation patching, we localize five layers that appear especially important for lying. We then find just 46 attention heads within these layers that enable us to causally intervene such that the lying model instead answers honestly. We show that these interventions work robustly across many prompts and dataset splits. Overall, our work contributes a greater understanding of dishonesty in LLMs so that we may hope to prevent it.

Large language models(LLMs) have sparked a new wave of exciting AI applications. Hosting these models at scale requires significant memory resources. One crucial memory bottleneck for the deployment stems from the context window. It is commonly recognized that model weights are memory hungry; however, the size of key-value embedding stored during the generation process (KV cache) can easily surpass the model size. The enormous size of the KV cache puts constraints on the inference batch size, which is crucial for high throughput inference workload. Inspired by an interesting observation of the attention scores, we hypothesize the persistence of importance: only pivotal tokens, which had a substantial influence at one step, will significantly influence future generations. Based on our empirical verification and theoretical analysis around this hypothesis, we propose Scissorhands, a system that maintains the memory usage of the KV cache at a fixed budget without finetuning the model. In essence, Scissorhands manages the KV cache by storing the pivotal tokens with a higher probability. We validate that Scissorhands reduces the inference memory usage of the KV cache by up to 5X without compromising model quality. We further demonstrate that Scissorhands can be combined with 4-bit quantization, traditionally used to compress model weights, to achieve up to 20X compression.

Generative Transformer-based models have achieved remarkable proficiency on solving diverse problems. However, their generalization ability is not fully understood and not always satisfying. Researchers take basic mathematical tasks like n-digit addition or multiplication as important perspectives for investigating their generalization behaviors. Curiously, it is observed that when training on n-digit operations (e.g., additions) in which both input operands are n-digit in length, models generalize successfully on unseen n-digit inputs (in-distribution (ID) generalization), but fail miserably and mysteriously on longer, unseen cases (out-of-distribution (OOD) generalization). Studies try to bridge this gap with workarounds such as modifying position embedding, fine-tuning, and priming with more extensive or instructive data. However, without addressing the essential mechanism, there is hardly any guarantee regarding the robustness of these solutions. We bring this unexplained performance drop into attention and ask whether it is purely from random errors. Here we turn to the mechanistic line of research which has notable successes in model interpretability. We discover that the strong ID generalization stems from structured representations, while behind the unsatisfying OOD performance, the models still exhibit clear learned algebraic structures. Specifically, these models map unseen OOD inputs to outputs with equivalence relations in the ID domain. These highlight the potential of the models to carry useful information for improved generalization.

Artificial neural networks are exceptionally good in learning to detect correlations within data that are associated with specified outcomes. However to deepen knowledge and support further research, researchers have to be able to explain predicted outcomes within the data's domain. Furthermore, domain experts like Healthcare Providers need these explanations to assess whether a predicted outcome can be trusted in high stakes scenarios and to help them incorporating a model into their own routine. In this paper we introduce Convolutional Motif Kernel Networks, a neural network architecture that incorporates learning a feature representation within a subspace of the reproducing kernel Hilbert space of the motif kernel function. The resulting model has state-of-the-art performance and enables researchers and domain experts to directly interpret and verify prediction outcomes without the need for a post hoc explainability method.