Current LLM-based services typically require users to submit raw text regardless of its sensitivity.While intuitive, such practice introduces substantial privacy risks, as unauthorized access may expose personal, medical, or legal information.Although prior defenses strived to mitigate these risks, they often incur substantial computational overhead and degrade model performance.To overcome this privacy-efficiency trade-off, we introduce Privacy-Preserving Fine-Tuning (PPFT), a novel training pipeline that eliminates the need for transmitting raw prompt text while maintaining a favorable balance between privacy preservation and model utility for both clients and service providers.Our approach operates in two stages: first, we train a client-side encoder together with a server-side projection module and LLM, enabling the server to condition on k-pooled prompt embeddings instead of raw text; second, we fine-tune the projection module and LLM on private, domain-specific data using noise-injected embeddings, allowing effective adaptation without exposing plain text prompts and requiring access to the decoder's internal parameters.Extensive experiments on domainspecific and general benchmarks demonstrate that PPFT achieves a striking balance between privacy and utility, maintaining competitive performance with minimal degradation compared to noise-free upper bounds.

Conversational agents have traditionally been developed for either task-oriented dialogue (TOD) or open-ended chitchat, with limited progress in unifying the two.Yet, real-world conversations naturally involve fluid transitions between these modes.To address this gap, we introduce TACT (TOD-And-Chitchat Transition), a dataset designed for transitionaware dialogue modeling that incorporates structurally diverse and integrated mode flows.TACT supports both user-and agent-driven mode switches, enabling robust modeling of complex conversational dynamics.To evaluate an agent's ability to initiate and recover from mode transitions, we propose two new metrics-Switch and Recovery.Models trained on TACT outperform baselines in both intent detection and mode transition handling.Moreover, applying Direct Preference Optimization (DPO) to TACT-trained models yields additional gains, achieving 75.74% joint mode-intent accuracy and a 70.1% win rate against GPT-4O in human evaluation.These results demonstrate that pairing structurally diverse data with DPO enhances response quality and transition control, paving the way for more proactive and transition-aware conversational agents.

Task-oriented dialogue (TOD) systems are commonly designed with the presumption that each utterance represents a single intent. However, this assumption may not accurately reflect real-world situations, where users frequently express multiple intents within a single utterance. While there is an emerging interest in multi-intent detection (MID), existing in-domain datasets such as MixATIS and MixSNIPS have limitations in their formulation. To address these issues, we present BlendX, a suite of refined datasets featuring more diverse patterns than their predecessors, elevating both its complexity and diversity. For dataset construction, we utilize both rule-based heuristics as well as a generative tool—OpenAI’s ChatGPT—which is augmented with a similarity-driven strategy for utterance selection. To ensure the quality of the proposed datasets, we also introduce three novel metrics that assess the statistical properties of an utterance related to word count, conjunction use, and pronoun usage. Extensive experiments on BlendX reveal that state-of-the-art MID models struggle with the challenges posed by the new datasets, highlighting the need to reexamine the current state of the MID field. The dataset is available at https://github.com/HYU-NLP/BlendX.

How can we leverage inherent frequency features of stock signals for effective portfolio optimization? Portfolio optimization in the domain of finance revolves around strategically allocating assets to maximize returns. Recent advancements highlight the efficacy of deep learning and reinforcement learning (RL) in capturing temporal asset patterns for portfolio optimization. However, previous methodologies focusing on time-domain often fail to detect sudden market shifts and abrupt events because their models are overly tailored to prevalent patterns, resulting in significant losses.

MOTIVATION: With the vast improvements in sequencing technologies and increased number of protocols, sequencing is being used to answer complex biological problems. Subsequently, analysis pipelines have become more time consuming and complicated, usually requiring highly extensive prevalidation steps. Here, we present SeqWho, a program designed to assess heuristically the quality of sequencing files and reliably classify the organism and protocol type by using Random Forest classifiers trained on biases native in k-mer frequencies and repeat sequence identities. RESULTS: Using one of our primary models, we show that our method accurately and rapidly classifies human and mouse sequences from nine different sequencing libraries by species, library and both together, 98.32%, 97.86% and 96.38% of the time, respectively. Ultimately, we demonstrate that SeqWho is a powerful method for reliably validating the quality and identity of the sequencing files used in any pipeline. AVAILABILITY AND IMPLEMENTATION: https://github.com/DaehwanKimLab/seqwho. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Algorithms for classifying chromosomes, like convolutional deep neural networks (CNNs), show promise to augment cytogeneticists' workflows; however, a critical limitation is their inability to accurately classify various structural chromosomal abnormalities. In hematopathology, recurrent structural cytogenetic abnormalities herald diagnostic, prognostic and therapeutic implications, but are laborious for expert cytogeneticists to identify. Non-recurrent cytogenetic abnormalities also occur frequently cancerous cells. Here, we demonstrate the feasibility of using CNNs to accurately classify many recurrent cytogenetic abnormalities while being able to reliably detect non-recurrent, spurious abnormal chromosomes, as well as provide insights into dataset assembly, model selection and training methodology that improve overall generalizability and performance for chromosome classification. RESULTS: Our top-performing model achieved a mean weighted F1 score of 96.86% on the validation set and 94.03% on the test set. Gradient class activation maps indicated that our model learned biologically meaningful feature maps, reinforcing the clinical utility of our proposed approach. Altogether, this work: proposes a new dataset framework for training chromosome classifiers for use in a clinical environment, reveals that residual CNNs and cyclical learning rates confer superior performance, and demonstrates the feasibility of using this approach to automatically screen for many recurrent cytogenetic abnormalities while adeptly classifying non-recurrent abnormal chromosomes. AVAILABILITY AND IMPLEMENTATION: Software is freely available at https://github.com/DaehwanKimLab/Chromosome-ReAd. The data underlying this article cannot be shared publicly due to it being protected patient information. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

This study proposes a visualization that can assist computer scientists and data scientists to make decisions by exploring technology trends. While it is important for them to understand the technology trends in the rapidly changing computer science and data science fields, it takes considerable time and knowledge to acquire good information about these trends. Particularly, data/computer scientists with little experience in the field find it difficult to obtain information on such trends. Therefore, we propose a visualization system that can easily and quickly explore the technology trends in computer and data science. This study aims to identify the key technologies and developers in a specific field, and other technologies deeply related to specific technologies, and explore the changes in popularity of technologies, languages, and libraries over time. This study includes two case studies to obtain information using the proposed visualization. We demonstrate our system with GitHub repositories data.

Performance analysis is essential for improving classification models. However, existing performance analysis tools do not provide actionable insights such as the cause of misclassification. Machine learning practitioners face difficulties such as prioritizing model, looking over confusion between classes. In addition, existing performance analysis tools that provide feature-level analysis are difficult to apply to image classification problems. This study has been proposed to solve these difficulties. In this paper, we present an interactive visual analytics system for diagnosing the performance of multiclass classification models. Our system is able to compare multiple models, find weaknesses, and obtain actionable insights for improving models. Our visualization consists of three views for analyzing performance at the class, confusion, and instance levels. We demonstrate our system using MNIST handwritten digits data.