Research on scaling large language models (LLMs) has primarily focused on model parameters and training data size, overlooking the role of vocabulary size. % Intuitively, larger vocabularies enable more efficient tokenization by representing sentences with fewer tokens, but they also increase the risk of under-fitting representations for rare tokens. We investigate how vocabulary size impacts LLM scaling laws by training models ranging from 33M to 3B parameters on up to 500B characters with various vocabulary configurations. We propose three complementary approaches for predicting the compute-optimal vocabulary size: IsoFLOPs analysis, derivative estimation, and parametric fit of the loss function. Our approaches converge on the same result that the optimal vocabulary size depends on the available compute budget and that larger models deserve larger vocabularies. However, most LLMs use too small vocabulary sizes. For example, we predict that the optimal vocabulary size of Llama2-70B should have been at least 216K, 7 times larger than its vocabulary of 32K. We validate our predictions empirically by training models with 3B parameters across different FLOPs budgets. Adopting our predicted optimal vocabulary size consistently improves downstream performance over commonly used vocabulary sizes. By increasing the vocabulary size from the conventional 32K to 43K, we improve performance on ARC-Challenge from 29.1 to 32.0 with the same 2.3e21 FLOPs. Our work emphasizes the necessity of jointly considering model parameters and vocabulary size for efficient scaling.

Large neural models are often compressed before deployment. Model compression is necessary for many practical reasons, such as inference latency, memory footprint, and energy consumption. Compressed models are assumed to be miniature versions of corresponding large neural models. However, we question this belief in our work. We compare compressed models with corresponding large neural models using four model characteristics: prediction errors, data representation, data distribution, and vulnerability to adversarial attack. We perform experiments using the BERT-large model and its five compressed versions. For all four model characteristics, compressed models significantly differ from the BERT-large model. Even among compressed models, they differ from each other on all four model characteristics. Apart from the expected loss in model performance, there are major side effects of using compressed models to replace large neural models.

Recommender Systems (RS) play an integral role in enhancing user experiences by providing personalized item suggestions. This survey reviews the progress in RS inclusively from 2017 to 2024, effectively connecting theoretical advances with practical applications. We explore the development from traditional RS techniques like content-based and collaborative filtering to advanced methods involving deep learning, graph-based models, reinforcement learning, and large language models. We also discuss specialized systems such as context-aware, review-based, and fairness-aware RS. The primary goal of this survey is to bridge theory with practice. It addresses challenges across various sectors, including e-commerce, healthcare, and finance, emphasizing the need for scalable, real-time, and trustworthy solutions. Through this survey, we promote stronger partnerships between academic research and industry practices. The insights offered by this survey aim to guide industry professionals in optimizing RS deployment and to inspire future research directions, especially in addressing emerging technological and societal trends

One way to increase confidence in the outputs of Large Language Models (LLMs) is to support them with reasoning that is clear and easy to check -- a property we call legibility. We study legibility in the context of solving grade-school math problems and show that optimizing chain-of-thought solutions only for answer correctness can make them less legible. To mitigate the loss in legibility, we propose a training algorithm inspired by Prover-Verifier Game from Anil et al. (2021). Our algorithm iteratively trains small verifiers to predict solution correctness, "helpful" provers to produce correct solutions that the verifier accepts, and "sneaky" provers to produce incorrect solutions that fool the verifier. We find that the helpful prover's accuracy and the verifier's robustness to adversarial attacks increase over the course of training. Furthermore, we show that legibility training transfers to time-constrained humans tasked with verifying solution correctness. Over course of LLM training human accuracy increases when checking the helpful prover's solutions, and decreases when checking the sneaky prover's solutions. Hence, training for checkability by small verifiers is a plausible technique for increasing output legibility. Our results suggest legibility training against small verifiers as a practical avenue for increasing legibility of large LLMs to humans, and thus could help with alignment of superhuman models.

Studying adversarial attacks on artificial intelligence (AI) systems helps discover model shortcomings, enabling the construction of a more robust system. Most existing adversarial attack methods only concentrate on single-task single-model or single-task cross-model scenarios, overlooking the multi-task characteristic of artificial intelligence systems. As a result, most of the existing attacks do not pose a practical threat to a comprehensive and collaborative AI system. However, implementing cross-task attacks is highly demanding and challenging due to the difficulty in obtaining the real labels of different tasks for the same picture and harmonizing the loss functions across different tasks. To address this issue, we propose a self-supervised Cross-Task Attack framework (CTA), which utilizes co-attention and anti-attention maps to generate cross-task adversarial perturbation. Specifically, the co-attention map reflects the area to which different visual task models pay attention, while the anti-attention map reflects the area that different visual task models neglect. CTA generates cross-task perturbations by shifting the attention area of samples away from the co-attention map and closer to the anti-attention map. We conduct extensive experiments on multiple vision tasks and the experimental results confirm the effectiveness of the proposed design for adversarial attacks.

Deploying deep learning-based imaging tools across various clinical sites poses significant challenges due to inherent domain shifts and regulatory hurdles associated with site-specific fine-tuning. For histopathology, stain normalization techniques can mitigate discrepancies, but they often fall short of eliminating inter-site variations. Therefore, we present Data Alchemy, an explainable stain normalization method combined with test time data calibration via a template learning framework to overcome barriers in cross-site analysis. Data Alchemy handles shifts inherent to multi-site data and minimizes them without needing to change the weights of the normalization or classifier networks. Our approach extends to unseen sites in various clinical settings where data domain discrepancies are unknown. Extensive experiments highlight the efficacy of our framework in tumor classification in hematoxylin and eosin-stained patches. Our explainable normalization method boosts classification tasks' area under the precision-recall curve (AUPR) by 0.165, 0.545 to 0.710. Additionally, Data Alchemy further reduces the multisite classification domain gap, by improving the 0.710 AUPR an additional 0.142, elevating classification performance further to 0.852, from 0.545. Our Data Alchemy framework can popularize precision medicine with minimal operational overhead by allowing for the seamless integration of pre-trained deep learning-based clinical tools across multiple sites.

Context. Risk analysis assesses potential risks in specific scenarios. Risk analysis principles are context-less; the same methodology can be applied to a risk connected to health and information technology security. Risk analysis requires a vast knowledge of national and international regulations and standards and is time and effort-intensive. A large language model can quickly summarize information in less time than a human and can be fine-tuned to specific tasks. Aim. Our empirical study aims to investigate the effectiveness of Retrieval-Augmented Generation and fine-tuned LLM in risk analysis. To our knowledge, no prior study has explored its capabilities in risk analysis. Method. We manually curated 193 unique scenarios leading to 1283 representative samples from over 50 mission-critical analyses archived by the industrial context team in the last five years. We compared the base GPT-3.5 and GPT-4 models versus their Retrieval-Augmented Generation and fine-tuned counterparts. We employ two human experts as competitors of the models and three other human experts to review the models and the former human experts' analysis. The reviewers analyzed 5,000 scenario analyses. Results and Conclusions. Human experts demonstrated higher accuracy, but LLMs are quicker and more actionable. Moreover, our findings show that RAG-assisted LLMs have the lowest hallucination rates, effectively uncovering hidden risks and complementing human expertise. Thus, the choice of model depends on specific needs, with FTMs for accuracy, RAG for hidden risks discovery, and base models for comprehensiveness and actionability. Therefore, experts can leverage LLMs as an effective complementing companion in risk analysis within a condensed timeframe. They can also save costs by averting unnecessary expenses associated with implementing unwarranted countermeasures.

Empirical risk minimization often fails to provide robustness against adversarial attacks in test data, causing poor out-of-sample performance. Adversarially robust optimization (ARO) has thus emerged as the de facto standard for obtaining models that hedge against such attacks. However, while these models are robust against adversarial attacks, they tend to suffer severely from overfitting. To address this issue for logistic regression, we study the Wasserstein distributionally robust (DR) counterpart of ARO and show that this problem admits a tractable reformulation. Furthermore, we develop a framework to reduce the conservatism of this problem by utilizing an auxiliary dataset (e.g., synthetic, external, or out-of-domain data), whenever available, with instances independently sampled from a nonidentical but related ground truth. In particular, we intersect the ambiguity set of the DR problem with another Wasserstein ambiguity set that is built using the auxiliary dataset. We analyze the properties of the underlying optimization problem, develop efficient solution algorithms, and demonstrate that the proposed method consistently outperforms benchmark approaches on real-world datasets.

The increasing use of large language models (LLMs) trained by third parties raises significant security concerns. In particular, malicious actors can introduce backdoors through poisoning attacks to generate undesirable outputs. While such attacks have been extensively studied in image domains and classification tasks, they remain underexplored for natural language generation (NLG) tasks. To address this gap, we conduct an investigation of various poisoning techniques targeting the LLM's fine-tuning phase via prefix-tuning, a Parameter Efficient Fine-Tuning (PEFT) method. We assess their effectiveness across two generative tasks: text summarization and text completion; and we also introduce new metrics to quantify the success and stealthiness of such NLG poisoning attacks. Through our experiments, we find that the prefix-tuning hyperparameters and trigger designs are the most crucial factors to influence attack success and stealthiness. Moreover, we demonstrate that existing popular defenses are ineffective against our poisoning attacks. Our study presents the first systematic approach to understanding poisoning attacks targeting NLG tasks during fine-tuning via PEFT across a wide range of triggers and attack settings. We hope our findings will aid the AI security community in developing effective defenses against such threats.

The NLP community has recently shown a growing interest in leveraging Large Language Models (LLMs) for knowledge-intensive tasks, viewing LLMs as potential knowledge bases (KBs). However, the reliability and extent to which LLMs can function as KBs remain underexplored. While previous studies suggest LLMs can encode knowledge within their parameters, the amount of parametric knowledge alone is not sufficient to evaluate their effectiveness as KBs. This study defines criteria that a reliable LLM-as-KB should meet, focusing on factuality and consistency, and covering both seen and unseen knowledge. We develop several metrics based on these criteria and use them to evaluate 26 popular LLMs, while providing a comprehensive analysis of the effects of model size, instruction tuning, and in-context learning (ICL). Our results paint a worrying picture. Even a high-performant model like GPT-3.5-turbo is not factual or consistent, and strategies like ICL and fine-tuning are unsuccessful at making LLMs better KBs.