While foundation models have been exploited for various expert tasks through fine-tuning, any foundation model will become outdated due to its old knowledge or limited capability. Thus the underlying foundation model should be eventually replaced by new ones, which leads to repeated cost of fine-tuning these new models. Existing work addresses this problem by inference-time tuning, i.e., modifying the output probabilities from the new foundation model with the outputs from the old foundation model and its fine-tuned model, which involves an additional overhead in inference by the latter two models. In this paper, we propose a new fine-tuning principle, Portable Reward Tuning (PRT), that reduces the inference overhead by its nature, based on the reformulation of fine-tuning as the reward maximization. Specifically, instead of fine-tuning parameters of the foundation models, PRT trains the reward model explicitly through the same loss function as in fine-tuning. During inference, the reward model can be used with any foundation model (with the same set of vocabularies or labels) through the formulation of reward maximization. Experimental results, covering both vision and language models, demonstrate that the PRT-trained model can achieve comparable accuracy to the existing work of inference-time tuning, with less inference cost.

In the realm of large-scale language models, a significant challenge arises when extrapolating sequences beyond the maximum allowable length. This is because the model's position embedding mechanisms are limited to positions encountered during training, thus preventing effective representation of positions in longer sequences. We analyzed conventional position encoding methods for long contexts and found the following characteristics. However, because it uses only a fixed scale parameter, it does not fully exploit the advantages of wavelet transforms, which capture the fine movements of non-stationary signals using multiple scales (window sizes). This limitation could explain why RoPE performs poorly in extrapolation. However, it has limitations in capturing deep dependencies because it restricts the receptive field of the model. From these insights, we propose a new position representation method that captures multiple scales (i.e., window sizes) by leveraging wavelet transforms without limiting the model's attention field. Experimental results show that this new method improves the performance of the model in both short and long contexts. In particular, our method allows extrapolation of position information without limiting the model's attention field. Several pre-trained large language models based on Transformer architecture (Vaswani et al., 2017) have demonstrated robust capabilities in various generative tasks (Devlin et al., 2019; Raffel et al., 2020; Brown et al., 2020; Touvron et al., 2023a; Jiang et al., 2023). However, limitations on the input sequence length arise due to the computational resource constraints encountered during the pre-training phase. This weakness is primarily attributed to the positional encoding's ineffectiveness in handling sequences that exceed the length of those encountered during the model's training phase (Devlin et al., 2019; Press et al., 2022). Rotary Position Embedding (RoPE) (Su et al., 2021) has become a common approach in many language models that handle long contexts, and it employs a rotation matrix to encode positional information and facilitate the processing of long sequences. To manage sequences longer than those encountered during training, various scaling strategies (Chen et al., 2023; bloc97, 2023; Peng et al., 2024; Liu et al., 2024) have been applied to RoPE, although these often require additional finetuning and incur further learning costs in addition to those of pre-training. In contrast, Attention with Linear Biases (ALiBi) (Press et al., 2022) is able to sequence length estimation beyond the limits of pre-training without requiring additional fine-tuning.

Existing Theory of Mind (ToM) benchmarks diverge from real-world scenarios in three aspects: 1) they assess a limited range of mental states such as beliefs, 2) false beliefs are not comprehensively explored, and 3) the diverse personality traits of characters are overlooked. To address these challenges, we introduce ToMATO, a new ToM benchmark formulated as multiple-choice QA over conversations. ToMATO is generated via LLM-LLM conversations featuring information asymmetry. By employing a prompting method that requires role-playing LLMs to verbalize their thoughts before each utterance, we capture both first- and second-order mental states across five categories: belief, intention, desire, emotion, and knowledge. These verbalized thoughts serve as answers to questions designed to assess the mental states of characters within conversations. Furthermore, the information asymmetry introduced by hiding thoughts from others induces the generation of false beliefs about various mental states. Assigning distinct personality traits to LLMs further diversifies both utterances and thoughts. ToMATO consists of 5.4k questions, 753 conversations, and 15 personality trait patterns. Our analysis shows that this dataset construction approach frequently generates false beliefs due to the information asymmetry between role-playing LLMs, and effectively reflects diverse personalities. We evaluate nine LLMs on ToMATO and find that even GPT-4o mini lags behind human performance, especially in understanding false beliefs, and lacks robustness to various personality traits.

Loss spikes, a phenomenon in which the loss value diverges suddenly, is a fundamental issue in the pre-training of large language models. This paper supposes that the non-uniformity of the norm of the parameters is one of the causes of loss spikes. Here, in training of neural networks, the scale of the gradients is required to be kept constant throughout the layers to avoid the vanishing and exploding gradients problem. However, to meet these requirements in the Transformer model, the norm of the model parameters must be non-uniform, and thus, parameters whose norm is smaller are more sensitive to the parameter update. To address this issue, we propose a novel technique, weight scaling as reparameterization (WeSaR). WeSaR introduces a gate parameter per parameter matrix and adjusts it to the value satisfying the requirements. Because of the gate parameter, WeSaR sets the norm of the original parameters uniformly, which results in stable training. Experimental results with the Transformer decoders consisting of 130 million, 1.3 billion, and 13 billion parameters showed that WeSaR stabilizes and accelerates training and that it outperformed compared methods including popular initialization methods.

We study the problem of completing various visual document understanding (VDU) tasks, e.g., question answering and information extraction, on real-world documents through human-written instructions. To this end, we propose InstructDoc, the first large-scale collection of 30 publicly available VDU datasets, each with diverse instructions in a unified format, which covers a wide range of 12 tasks and includes open document types/formats. Furthermore, to enhance the generalization performance on VDU tasks, we design a new instruction-based document reading and understanding model, InstructDr, that connects document images, image encoders, and large language models (LLMs) through a trainable bridging module. Experiments demonstrate that InstructDr can effectively adapt to new VDU datasets, tasks, and domains via given instructions and outperforms existing multimodal LLMs and ChatGPT without specific training.

Although named entity recognition (NER) helps us to extract domain-specific entities from text (e.g., artists in the music domain), it is costly to create a large amount of training data or a structured knowledge base to perform accurate NER in the target domain. Here, we propose self-adaptive NER, which retrieves external knowledge from unstructured text to learn the usages of entities that have not been learned well. To retrieve useful knowledge for NER, we design an effective two-stage model that retrieves unstructured knowledge using uncertain entities as queries. Our model predicts the entities in the input and then finds those of which the prediction is not confident. Then, it retrieves knowledge by using these uncertain entities as queries and concatenates the retrieved text to the original input to revise the prediction. Experiments on CrossNER datasets demonstrated that our model outperforms strong baselines by 2.35 points in F1 metric.

Automatic image editing has great demands because of its numerous applications, and the use of natural language instructions is essential to achieving flexible and intuitive editing as the user imagines. A pioneering work in text-driven image editing, StyleCLIP, finds an edit direction in the CLIP space and then edits the image by mapping the direction to the StyleGAN space. At the same time, it is difficult to tune appropriate inputs other than the original image and text instructions for image editing. In this study, we propose a method to construct the edit direction adaptively in the StyleGAN and CLIP spaces with SVM. Our model represents the edit direction as a normal vector in the CLIP space obtained by training a SVM to classify positive and negative images. The images are retrieved from a large-scale image corpus, originally used for pre-training StyleGAN, according to the CLIP similarity between the images and the text instruction. We confirmed that our model performed as well as the StyleCLIP baseline, whereas it allows simple inputs without increasing the computational time.

Visual question answering on document images that contain textual, visual, and layout information, called document VQA, has received much attention recently. Although many datasets have been proposed for developing document VQA systems, most of the existing datasets focus on understanding the content relationships within a single image and not across multiple images. In this study, we propose a new multi-image document VQA dataset, SlideVQA, containing 2.6k+ slide decks composed of 52k+ slide images and 14.5k questions about a slide deck. SlideVQA requires complex reasoning, including single-hop, multi-hop, and numerical reasoning, and also provides annotated arithmetic expressions of numerical answers for enhancing the ability of numerical reasoning. Moreover, we developed a new end-to-end document VQA model that treats evidence selection and question answering in a unified sequence-to-sequence format. Experiments on SlideVQA show that our model outperformed existing state-of-the-art QA models, but that it still has a large gap behind human performance. We believe that our dataset will facilitate research on document VQA.

One of the problems with automated audio captioning (AAC) is the indeterminacy in word selection corresponding to the audio event/scene. Since one acoustic event/scene can be described with several words, it results in a combinatorial explosion of possible captions and difficulty in training. To solve this problem, we propose a Transformer-based audio-captioning model with keyword estimation called TRACKE. It simultaneously solves the word-selection indeterminacy problem with the main task of AAC while executing the sub-task of acoustic event detection/acoustic scene classification (i.e., keyword estimation). TRACKE estimates keywords, which comprise a word set corresponding to audio events/scenes in the input audio, and generates the caption while referring to the estimated keywords to reduce word-selection indeterminacy. Experimental results on a public AAC dataset indicate that TRACKE achieved state-of-the-art performance and successfully estimated both the caption and its keywords.

We propose a new deep neural network architecture, TabNet, for table type classification. Table type is essential information for exploring the power of Web tables, and it is important to understand the semantic structures of tables in order to classify them correctly. A table is a matrix of texts, analogous to an image, which is a matrix of pixels, and each text consists of a sequence of tokens. Our hybrid architecture mirrors the structure of tables: its recurrent neural network (RNN) encodes a sequence of tokens for each cell to create a 3d table volume like image data, and its convolutional neural network (CNN) captures semantic features, e.g., the existence of rows describing properties, to classify tables. Experiments using Web tables with various structures and topics demonstrated that TabNet achieved considerable improvements over state-of-the-art methods specialized for table classification and other deep neural network architectures.