Large language models (LLMs) often leverage adapters, such as low-rank-based adapters, to achieve strong performance on downstream tasks. However, storing a separate adapter for each task significantly increases memory requirements, posing a challenge for resource-constrained environ ments such as mobile devices. Although model merging techniques can reduce storage costs, they typically result in substantial performance degradation. In this work, we introduce HydraOpt, a new model merging technique that capitalizes on the inherent similarities between the matrices of low-rank adapters. Unlike existing methods that produce a fixed trade-off between storage size and performance, HydraOpt allows us to navigate this spectrum of efficiency and performance. Our experiments show that HydraOpt significantly reduces storage size (48% reduction) compared to storing all adapters, while achieving competitive performance (0.2-1.8% drop). Furthermore, it outperforms existing merging techniques in terms of performance at the same or slightly worse storage efficiency.

Adapter parameters provide a mechanism to modify the behavior of machine learning models and have gained significant popularity in the context of large language models (LLMs) and generative AI. These parameters can be merged to support multiple tasks via a process known as task merging. However, prior work on merging in LLMs, particularly in natural language processing, has been limited to scenarios where each test example addresses only a single task. In this paper, we focus on on-device settings and study the problem of text-based compositional multi-tasking, where each test example involves the simultaneous execution of multiple tasks. For instance, generating a translated summary of a long text requires solving both translation and summarization tasks concurrently. To facilitate research in this setting, we propose a benchmark comprising four practically relevant compositional tasks. We also present an efficient method (Learnable Calibration) tailored for on-device applications, where computational resources are limited, emphasizing the need for solutions that are both resource-efficient and high-performing. Our contributions lay the groundwork for advancing the capabilities of LLMs in real-world multi-tasking scenarios, expanding their applicability to complex, resource-constrained use cases.

Large language models (LLMs) are commonly adapted for diverse downstream tasks via parameter-efficient fine-tuning techniques such as Low-Rank Adapters (LoRA). While adapters can be combined to handle multiple tasks separately, standard approaches struggle when targeting the simultaneous execution of complex tasks, such as generating a translated summary from a long conversation. To address this challenge, we propose a novel approach tailored specifically for compositional multi-tasking scenarios involving summarization and translation. Our technique involves adding a learnable projection layer on top of the combined summarization and translation adapters. This design enables effective integration while maintaining efficiency through reduced computational overhead compared to alternative strategies requiring extensive retraining or sequential processing. We demonstrate the practical viability of our method within an on-device environment by developing an Android app capable of executing compositional tasks seamlessly. Experimental results indicate our solution performs well and is fast in both cloud-based and on-device implementations, highlighting the potential benefits of adopting our framework in real-world applications demanding high-speed operation alongside resource constraints.

Dialogue State Tracking (DST) is a very complex task that requires precise understanding and information tracking of multi-domain conversations between users and dialogue systems. Many task-oriented dialogue systems use dialogue state tracking technology to infer users’ goals from the history of the conversation. Existing approaches for DST are usually conditioned on previous dialogue states. However, the dependency on previous dialogues makes it very challenging to prevent error propagation to subsequent turns of a dialogue. In this paper, we propose Neural Retrieval Augmentation to alleviate this problem by creating a Neural Index based on dialogue context. Our NRA-DST framework efficiently retrieves dialogue context from the index built using a combination of unstructured dialogue state and structured user/system utterances. We explore a simple pipeline resulting in a retrieval-guided generation approach for training a DST model. Experiments on different retrieval methods for augmentation show that neural retrieval augmentation is the best performing retrieval method for DST. Our evaluations on the large-scale MultiWOZ dataset show that our model outperforms the baseline approaches.

Dialogue State Tracking (DST) is core research in dialogue systems and has received much attention. In addition, it is necessary to define a new problem that can deal with dialogue between users as a step toward the conversational AI that extracts and recommends information from the dialogue between users. So, we introduce a new task - DST from dialogue between users about scheduling an event (DST-USERS). The DST-USERS task is much more challenging since it requires the model to understand and track dialogue states in the dialogue between users, as well as to understand who suggested the schedule and who agreed to the proposed schedule. To facilitate DST-USERS research, we develop dialogue datasets between users that plan a schedule. The annotated slot values which need to be extracted in the dialogue are date, time, and location. Previous approaches, such as Machine Reading Comprehension (MRC) and traditional DST techniques, have not achieved good results in our extensive evaluations. By adopting the knowledge-integrated learning method, we achieve exceptional results. The proposed model architecture combines gazetteer features and speaker information efficiently. Our evaluations of the dialogue datasets between users that plan a schedule show that our model outperforms the baseline model.

Since language model pretraining to learn contextualized word representations has been proposed, pretrained language models have made success in many natural language processing tasks. That is because it is helpful to use individual contextualized representations of self-attention layers as to initialize parameters for downstream tasks. Yet, unfortunately, use of pretrained language models for emotion recognition in conversations has not been studied enough. We firstly use ELECTRA which is a state-of-the-art pretrained language model and validate the performance on emotion recognition in conversations. Furthermore, we propose contextual augmentation of pretrained language models for emotion recognition in conversations, which is to consider not only previous utterances, but also conversation-related information such as speakers, speech acts and topics. We classify information based on what the information is related to, and propose position of words corresponding to the information in the entire input sequence. To validate the proposed method, we conduct experiments on the DailyDialog dataset which contains abundant annotated information of conversations. The experiments show that the proposed method achieves state-of-the-art F1 scores on the dataset and significantly improves the performance.