Language models pretrained on text-only corpora often struggle with tasks that require auditory commonsense knowledge. Previous work addresses this problem by augmenting the language model to retrieve knowledge from external audio databases. This approach has several limitations, such as the potential lack of relevant audio in databases and the high costs associated with constructing and querying the databases. To address these issues, we propose Imagine to Hear, a novel approach that dynamically generates auditory knowledge using generative models. Our framework detects multiple audio-related textual spans from the given prompt and generates corresponding auditory knowledge. We develop several mechanisms to efficiently process multiple auditory knowledge, including a CLAP-based rejection sampler and a language-audio fusion module. Our experiments show that our method achieves state-of-the-art performance on AuditoryBench without relying on external databases, highlighting the effectiveness of our generation-based approach.

Weight space learning is an emerging paradigm in the deep learning community. The primary goal of weight space learning is to extract informative features from a set of parameters using specially designed neural networks, often referred to as metanetworks. However, it remains unclear how these metanetworks learn solely from parameters. To address this, we take the first step toward understanding representations of metanetworks, specifically graph metanetworks (GMNs), which achieve state-of-the-art results in this field, using centered kernel alignment (CKA). Through various experiments, we reveal that GMNs and general neural networks (e.g., multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs)) differ in terms of their representation space. The field of weight space learning has gained a lot of attention these days.

Depth pruning aims to reduce the inference cost of a large language model without any hardware-specific complications, by simply removing several less important transformer blocks. However, our empirical findings suggest that the importance of a transformer block may be highly task-dependent -- a block that is crucial for a task can be removed without degrading the accuracy on another task. Based on this observation, we develop a dynamic depth pruning algorithm, coined PuDDing (Prompt-routed Dynamic Depth Pruning), which determines which blocks to omit from the model based on the input prompt. PuDDing operates by training a lightweight router to predict the best omission set among a set of options, where this option set has also been constructed in a data-driven manner. Empirical results on commonsense reasoning benchmarks demonstrate that PuDDing effectively accelerates the inference language models, and achieves better on-task performance than static depth pruning baselines.

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 3,000 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Neural fields are an emerging paradigm that represent data as continuous functions parameterized by neural networks. Despite many advantages, neural fields often have a high training cost, which prevents a broader adoption. In this paper, we focus on a popular family of neural fields, called sinusoidal neural fields (SNFs), and study how it should be initialized to maximize the training speed. We find that the standard initialization scheme for SNFs--designed based on the signal propagation principle--is suboptimal. In particular, we show that by simply multiplying each weight (except for the last layer) by a constant, we can accelerate SNF training by 10 . This method, coined weight scaling, consistently provides a significant speedup over various data domains, allowing the SNFs to train faster than more recently proposed architectures. To understand why the weight scaling works well, we conduct extensive theoretical and empirical analyses which reveal that the weight scaling not only resolves the spectral bias quite effectively but also enjoys a well-conditioned optimization trajectory. Neural field (NF) is a special family of neural networks designed to represent a single datum (Xie et al., 2022). Precisely, NFs parametrize each datum with the weights of a neural net which is trained to fit the mapping from spatiotemporal coordinates to corresponding signal values. For neural fields, the training speed is of vital importance. Representing each datum as an NF requires tedious training of a neural network, which can take up to several hours of GPU-based training (Mildenhall et al., 2020). This computational burden becomes a critical obstacle toward adoption to tasks that involve large-scale data, such as NF-based inference or generation (Ma et al., 2024; Papa et al., 2024). To address this issue, many research efforts have been taken to accelerate NF training, including fast-trainable NF architectures (Sitzmann et al., 2020b; Müller et al., 2022), meta-learning (Sitzmann et al., 2020a; Chen & Wang, 2022), and data transformations (Seo et al., 2024).

How can sLLMs efficiently utilize the supervision of LLMs to improve their generative quality? This question has been well studied in scenarios where there is no restriction on the number of LLM supervisions one can use, giving birth to many decoding algorithms that utilize supervision without further training. However, it is still unclear what is an effective strategy under the limited supervision scenario, where we assume that no more than a few tokens can be generated by LLMs. To this end, we develop an algorithm to effectively aggregate the sLLM and LLM predictions on initial tokens so that the generated tokens can more accurately condition the subsequent token generation by sLLM only. Critically, we find that it is essential to adaptively overtrust or disregard the LLM prediction based on the confidence of the sLLM. Through our experiments on a wide range of models and datasets, we demonstrate that our method provides a consistent improvement over conventional decoding strategies.

This work suggests fundamentally rethinking the current practice of pruning large language models (LLMs). The way it is done is by divide and conquer: split the model into submodels, sequentially prune them, and reconstruct predictions of the dense counterparts on small calibration data one at a time; the final model is obtained simply by putting the resulting sparse submodels together. While this approach enables pruning under memory constraints, it generates high reconstruction errors. In this work, we first present an array of reconstruction techniques that can significantly reduce this error by more than $90\%$. Unwittingly, however, we discover that minimizing reconstruction error is not always ideal and can overfit the given calibration data, resulting in rather increased language perplexity and poor performance at downstream tasks. We find out that a strategy of self-generating calibration data can mitigate this trade-off between reconstruction and generalization, suggesting new directions in the presence of both benefits and pitfalls of reconstruction for pruning LLMs.

Despite recent advances in LLM quantization, activation quantization remains to be challenging due to the activation outliers. Conventional remedies, e.g., mixing precisions for different channels, introduce extra overhead and reduce the speedup. In this work, we develop a simple yet effective strategy to facilitate per-tensor activation quantization by preventing the generation of problematic tokens. Precisely, we propose a method to find a set of key-value cache, coined CushionCache, which mitigates outliers in subsequent tokens when inserted as a prefix. CushionCache works in two steps: First, we greedily search for a prompt token sequence that minimizes the maximum activation values in subsequent tokens. Then, we further tune the token cache to regularize the activations of subsequent tokens to be more quantization-friendly. The proposed method successfully addresses activation outliers of LLMs, providing a substantial performance boost for per-tensor activation quantization methods. We thoroughly evaluate our method over a wide range of models and benchmarks and find that it significantly surpasses the established baseline of per-tensor W8A8 quantization and can be seamlessly integrated with the recent activation quantization method.

We propose a communication-efficient collaborative inference framework in the domain of edge inference, focusing on the efficient use of vision transformer (ViT) models. The partitioning strategy of conventional collaborative inference fails to reduce communication cost because of the inherent architecture of ViTs maintaining consistent layer dimensions across the entire transformer encoder. Therefore, instead of employing the partitioning strategy, our framework utilizes a lightweight ViT model on the edge device, with the server deploying a complicated ViT model. To enhance communication efficiency and achieve the classification accuracy of the server model, we propose two strategies: 1) attention-aware patch selection and 2) entropy-aware image transmission. Attention-aware patch selection leverages the attention scores generated by the edge device's transformer encoder to identify and select the image patches critical for classification. This strategy enables the edge device to transmit only the essential patches to the server, significantly improving communication efficiency. Entropy-aware image transmission uses min-entropy as a metric to accurately determine whether to depend on the lightweight model on the edge device or to request the inference from the server model. In our framework, the lightweight ViT model on the edge device acts as a semantic encoder, efficiently identifying and selecting the crucial image information required for the classification task. Our experiments demonstrate that the proposed collaborative inference framework can reduce communication overhead by 68% with only a minimal loss in accuracy compared to the server model on the ImageNet dataset.

Recent advances in text-guided image compression have shown great potential to enhance the perceptual quality of reconstructed images. These methods, however, tend to have significantly degraded pixel-wise fidelity, limiting their practicality. To fill this gap, we develop a new text-guided image compression algorithm that achieves both high perceptual and pixel-wise fidelity. In particular, we propose a compression framework that leverages text information mainly by text-adaptive encoding and training with joint image-text loss. By doing so, we avoid decoding based on text-guided generative models -- known for high generative diversity -- and effectively utilize the semantic information of text at a global level. Experimental results on various datasets show that our method can achieve high pixel-level and perceptual quality, with either human- or machine-generated captions. In particular, our method outperforms all baselines in terms of LPIPS, with some room for even more improvements when we use more carefully generated captions.