Large language models (LLMs) frequently make errors when handling even simple numerical problems, such as comparing two small numbers. A natural hypothesis is that these errors stem from how LLMs represent numbers, and specifically, whether their representations of numbers capture their numeric values. We tackle this question from the observation that LLM errors on numerical tasks are often distributed across \textit{the digits} of the answer rather than normally around \textit{its numeric value}. Through a series of probing experiments and causal interventions, we show that LLMs internally represent numbers with individual circular representations per-digit in base 10. This digit-wise representation, as opposed to a value representation, sheds light on the error patterns of models on tasks involving numerical reasoning and could serve as a basis for future studies on analyzing numerical mechanisms in LLMs.

While Large language model (LLM)-based programming assistants such as CoPilot and ChatGPT can help improve the productivity of professional software developers, they can also facilitate cheating in introductory computer programming courses. Assuming instructors have limited control over the industrial-strength models, this paper investigates the baseline performance of 5 widely used LLMs on a collection of introductory programming problems, examines adversarial perturbations to degrade their performance, and describes the results of a user study aimed at understanding the efficacy of such perturbations in hindering actual code generation for introductory programming assignments. The user study suggests that i) perturbations combinedly reduced the average correctness score by 77%, ii) the drop in correctness caused by these perturbations was affected based on their detectability.

Traditional Retrieval-Augmented Generation (RAG) methods are limited by their reliance on a fixed number of retrieved documents, often resulting in incomplete or noisy information that undermines task performance. Although recent adaptive approaches alleviated these problems, their application in intricate and real-world multimodal tasks remains limited. To address these, we propose a new approach called Self-adaptive Multimodal Retrieval-Augmented Generation (SAM-RAG), tailored specifically for multimodal contexts. SAM-RAG not only dynamically filters relevant documents based on the input query, including image captions when needed, but also verifies the quality of both the retrieved documents and the output. Extensive experimental results show that SAM-RAG surpasses existing state-of-the-art methods in both retrieval accuracy and response generation. By further ablation experiments and effectiveness analysis, SAM-RAG maintains high recall quality while improving overall task performance in multimodal RAG task. Our codes are available at https://github.com/SAM-RAG/SAM_RAG.

Bias in machine learning models has been a chronic problem, especially as these models influence decision-making in human society. In generative AI, such as Large Language Models, the impact of bias is even more profound compared to the classification models. LLMs produce realistic and human-like content that users may unconsciously trust, which could perpetuate harmful stereotypes to the uncontrolled public. It becomes particularly concerning when utilized in journalism or education. While prior studies have explored and quantified bias in individual AI models, no work has yet compared bias similarity across different LLMs. To fill this gap, we take a comprehensive look at ten open- and closed-source LLMs from four model families, assessing the extent of biases through output distribution. Using two datasets-one containing 4k questions and another with one million questions for each of the four bias dimensions -- we measure functional similarity to understand how biases manifest across models. Our findings reveal that 1) fine-tuning does not significantly alter output distributions, which would limit its ability to mitigate bias, 2) LLMs within the same family tree do not produce similar output distributions, implying that addressing bias in one model could have limited implications for others in the same family, and 3) there is a possible risk of training data information leakage, raising concerns about privacy and data security. Our analysis provides insight into LLM behavior and highlights potential risks in real-world deployment.

We consider the problem of detecting, isolating and classifying elephant calls in continuously recorded audio. Such automatic call characterisation can assist conservation efforts and inform environmental management strategies. In contrast to previous work in which call detection was performed at a segment level, we perform call detection at a frame level which implicitly also allows call endpointing, the isolation of a call in a longer recording. For experimentation, we employ two annotated datasets, one containing Asian and the other African elephant vocalisations. We evaluate several shallow and deep classifier models, and show that the current best performance can be improved by using an audio spectrogram transformer (AST), a neural architecture which has not been used for this purpose before, and which we have configured in a novel sequence-to-sequence manner. We also show that using transfer learning by pre-training leads to further improvements both in terms of computational complexity and performance. Finally, we consider sub-call classification using an accepted taxonomy of call types, a task which has not previously been considered. We show that also in this case the transformer architectures provide the best performance. Our best classifiers achieve an average precision (AP) of 0.962 for framewise binary call classification, and an area under the receiver operating characteristic (AUC) of 0.957 and 0.979 for call classification with 5 classes and sub-call classification with 7 classes respectively. All of these represent either new benchmarks (sub-call classifications) or improvements on previously best systems. We conclude that a fully-automated elephant call detection and subcall classification system is within reach. Such a system would provide valuable information on the behaviour and state of elephant herds for the purposes of conservation and management.

Recent advances in knowledge distillation (KD) have enabled smaller student models to approach the performance of larger teacher models. However, popular methods such as supervised KD and on-policy KD, are adversely impacted by the knowledge gaps between teacher-student in practical scenarios. Supervised KD suffers from a distribution mismatch between training with a static dataset and inference over final student-generated outputs. Conversely, on-policy KD, which uses student-generated samples for training, can suffer from low-quality training examples with which teacher models are not familiar, resulting in inaccurate teacher feedback. To address these limitations, we introduce Speculative Knowledge Distillation (SKD), a novel approach that leverages cooperation between student and teacher models to generate high-quality training data on-the-fly while aligning with the student's inference-time distribution. In SKD, the student proposes tokens, and the teacher replaces poorly ranked ones based on its own distribution, transferring high-quality knowledge adaptively. We evaluate SKD on various text generation tasks, including translation, summarization, math, and instruction following, and show that SKD consistently outperforms existing KD methods across different domains, data sizes, and model initialization strategies. Figure 1: SKD outperforms supervised and on-policy KD for our tested tasks: Assamese-to-English translation, dialogue summarization, and arithmetic reasoning. Supervised KD is trained on ground-truth outputs, while on-policy KD uses self-generated data. All models use greedy decoding for evaluation. Work done as a student researcher at Google Cloud AI Research. Left: SKD addresses the limitations of on-policy knowledge distillation (KD) by filtering out low-quality student samples and replacing them with teacher generated tokens. However, the substantial inference-time costs and memory footprint associated with LLMs present significant challenges for practical deployment (Agarwal et al., 2024). Therefore, compressing LLMs while maintaining their performance is crucial for real-time practical applications. Knowledge Distillation (KD) (Hinton et al., 2015) is a widely used method to compress LLMs by transferring knowledge from a larger teacher model to a smaller student model. Traditional KD approaches, such as supervised KD (Sanh et al., 2020) and SeqKD (Kim & Rush, 2016b), rely on a static dataset of outputs to train the student model. However, this fixed dataset can lead to a distribution mismatch between the training data and the student's generated samples at inference time, hindering the student's learning.

Their vast parameters (Kaplan et al., 2020) and extensive training data grant them strong capabilities, but they may still generate outputs that conflict with human values, such as helpless or harmful content. Therefore, AI alignment research has emerged with the goal of fine-tuning LLMs to make them align with human values. One of the most popular alignment methods is RLHF(Reinforcement Learning from Human Feedback) framework (Stiennon et al., 2020; Ziegler et al., 2019; Ouyang et al., 2022), which initially apply supervised fine-tuning to the base model to follow human instructions. Subsequently, a reward model is trained from the human preference data, then optimizing the LLM via PPO algorithm (Schulman et al., 2017) to align with huamn preferences. RLHF requires at least three large models for training, making the process quite complex, and the PPO algorithm itself is highly sophisticated and challenging to parameter-tuning. This drives researchers to explore simpler and more straightforward methods to align language models with human preferences. To simplify alignment, (Rafailov et al., 2023) introduced Direct Preference Optimization (DPO), which provides a closed-form alignment solution and directly uses human preferences for alignment without a separate reward model. Other approaches, like RRHF (Yuan et al., 2023a) and PRO (Song et al., 2024), use SFT-like loss based on multi-ranking datasets to provide richer supervision for alignment.

Humanitarian organizations can enhance their effectiveness by analyzing data to discover trends, gather aggregated insights, manage their security risks, support decision-making, and inform advocacy and funding proposals. However, data about violent incidents with direct impact and relevance for humanitarian aid operations is not readily available. An automatic data collection and NLP-backed classification framework aligned with humanitarian perspectives can help bridge this gap. In this paper, we present HumVI - a dataset comprising news articles in three languages (English, French, Arabic) containing instances of different types of violent incidents categorized by the humanitarian sector they impact, e.g., aid security, education, food security, health, and protection. Reliable labels were obtained for the dataset by partnering with a data-backed humanitarian organization, Insecurity Insight. We provide multiple benchmarks for the dataset, employing various deep learning architectures and techniques, including data augmentation and mask loss, to address different task-related challenges, e.g., domain expansion. The dataset is publicly available at https://github.com/dataminr-ai/humvi-dataset.

Low Rank Adaptation (LoRA) is a popular Parameter Efficient Fine Tuning (PEFT) method that effectively adapts large pre-trained models for downstream tasks. LoRA parameterizes model updates using low-rank matrices at each layer, significantly reducing the number of trainable parameters and, consequently, resource requirements during fine-tuning. However, the lower bound on the number of trainable parameters remains high due to the use of the low-rank matrix model. In this paper, we address this limitation by proposing a novel approach that employs a low rank tensor parametrization for model updates. The proposed low rank tensor model can significantly reduce the number of trainable parameters, while also allowing for finer-grained control over adapter size. Our experiments on Natural Language Understanding, Instruction Tuning, Preference Optimization and Protein Folding benchmarks demonstrate that our method is both efficient and effective for fine-tuning large language models, achieving a substantial reduction in the number of parameters while maintaining comparable performance.

Large Language Models (LLMs) rely on the contextual information embedded in examples/demonstrations to perform in-context learning (ICL). To mitigate the risk of LLMs potentially leaking private information contained in examples in the prompt, we introduce a novel data-adaptive differentially private algorithm called AdaDPSyn to generate synthetic examples from the private dataset and then use these synthetic examples to perform ICL. The objective of AdaDPSyn is to adaptively adjust the noise level in the data synthesis mechanism according to the inherent statistical properties of the data, thereby preserving high ICL accuracy while maintaining formal differential privacy guarantees. A key innovation in AdaDPSyn is the Precision-Focused Iterative Radius Reduction technique, which dynamically refines the aggregation radius - the scope of data grouping for noise addition - based on patterns observed in data clustering, thereby minimizing the amount of additive noise. We conduct extensive experiments on standard benchmarks and compare AdaDPSyn with DP few-shot generation algorithm (Tang et al., 2023). The experiments demonstrate that AdaDPSyn not only outperforms DP few-shot generation, but also maintains high accuracy levels close to those of non-private baselines, providing an effective solution for ICL with privacy protection.