In recent years, Multimodal Large Language Models (MLLM) have achieved notable advancements, demonstrating the feasibility of developing an intelligent biomedical assistant. However, current biomedical MLLMs predominantly focus on image-level understanding and restrict interactions to textual commands, thus limiting their capability boundaries and the flexibility of usage. In this paper, we introduce a novel end-to-end multimodal large language model for the biomedical domain, named MedPLIB, which possesses pixel-level understanding. Excitingly, it supports visual question answering (VQA), arbitrary pixel-level prompts (points, bounding boxes, and free-form shapes), and pixel-level grounding. We propose a novel Mixture-of-Experts (MoE) multi-stage training strategy, which divides MoE into separate training phases for a visual-language expert model and a pixel-grounding expert model, followed by fine-tuning using MoE. This strategy effectively coordinates multitask learning while maintaining the computational cost at inference equivalent to that of a single expert model. To advance the research of biomedical MLLMs, we introduce the Medical Complex Vision Question Answering Dataset (MeCoVQA), which comprises an array of 8 modalities for complex medical imaging question answering and image region understanding. Experimental results indicate that MedPLIB has achieved state-of-the-art outcomes across multiple medical visual language tasks. More importantly, in zero-shot evaluations for the pixel grounding task, MedPLIB leads the best small and large models by margins of 19.7 and 15.6 respectively on the mDice metric. The codes, data, and model checkpoints will be made publicly available at https://github.com/ShawnHuang497/MedPLIB.

Detecting user frustration in modern-day task-oriented dialog (TOD) systems is imperative for maintaining overall user satisfaction, engagement, and retention. However, most recent research is focused on sentiment and emotion detection in academic settings, thus failing to fully encapsulate implications of real-world user data. To mitigate this gap, in this work, we focus on user frustration in a deployed TOD system, assessing the feasibility of out-of-the-box solutions for user frustration detection. Specifically, we compare the performance of our deployed keyword-based approach, open-source approaches to sentiment analysis, dialog breakdown detection methods, and emerging in-context learning LLM-based detection. Our analysis highlights the limitations of open-source methods for real-world frustration detection, while demonstrating the superior performance of the LLM-based approach, achieving a 16\% relative improvement in F1 score on an internal benchmark. Finally, we analyze advantages and limitations of our methods and provide an insight into user frustration detection task for industry practitioners.

The advancements in generative AI inevitably raise concerns about their risks and safety implications, which, in return, catalyzes significant progress in AI safety. However, as this field continues to evolve, a critical question arises: are our current efforts on AI safety aligned with the advancements of AI as well as the long-term goal of human civilization? This paper presents a blueprint for an advanced human society and leverages this vision to guide current AI safety efforts. It outlines a future where the Internet of Everything becomes reality, and creates a roadmap of significant technological advancements towards this envisioned future. For each stage of the advancements, this paper forecasts potential AI safety issues that humanity may face. By projecting current efforts against this blueprint, this paper examines the alignment between the current efforts and the long-term needs, and highlights unique challenges and missions that demand increasing attention from AI safety practitioners in the 2020s. This vision paper aims to offer a broader perspective on AI safety, emphasizing that our current efforts should not only address immediate concerns but also anticipate potential risks in the expanding AI landscape, thereby promoting a safe and sustainable future of AI and human civilization.

Expressive zero-shot voice conversion (VC) is a critical and challenging task that aims to transform the source timbre into an arbitrary unseen speaker while preserving the original content and expressive qualities. Despite recent progress in zero-shot VC, there remains considerable potential for improvements in speaker similarity and speech naturalness. Moreover, existing zero-shot VC systems struggle to fully reproduce paralinguistic information in highly expressive speech, such as breathing, crying, and emotional nuances, limiting their practical applicability. To address these issues, we propose Takin-VC, a novel expressive zero-shot VC framework via adaptive hybrid content encoding and memory-augmented context-aware timbre modeling. Specifically, we introduce an innovative hybrid content encoder that incorporates an adaptive fusion module, capable of effectively integrating quantized features of the pre-trained WavLM and HybridFormer in an implicit manner, so as to extract precise linguistic features while enriching paralinguistic elements. For timbre modeling, we propose advanced memory-augmented and context-aware modules to generate high-quality target timbre features and fused representations that seamlessly align source content with target timbre. To enhance real-time performance, we advocate a conditional flow matching model to reconstruct the Mel-spectrogram of the source speech. Experimental results show that our Takin-VC consistently surpasses state-of-the-art VC systems, achieving notable improvements in terms of speech naturalness, speech expressiveness, and speaker similarity, while offering enhanced inference speed.

Tens of thousands of Californians are turning to a crowdsourced, nonprofit app called Watch Duty for critical, up-to-the-moment disaster updates as deadly fires continue to rage through the state. The app, which uses a mixture of official government and volunteer data to track wildfires, surpassed OpenAI's ChatGPT and Meta's Threads as the most downloaded app on the Apple App Store on Wednesday. Social media users have encouraged residents in affected areas to download the app in order to track the fire's rapid movements and stay aware of possible evacuation orders. Apps like Watch Duty, which have seen a surge in interest in recent years, may become even more important as climate change-related natural disasters intensify in scope and scale. It gives you updates on fires nearby, evacuation notices, and even will show you where an evacuation center is if you need to evacuate!

Artificial intelligence companies have run out of data for training their models and have "exhausted" the sum of human knowledge, Elon Musk has said. The world's richest person suggested technology firms would have to turn to "synthetic" data – or material created by AI models – to build and fine-tune new systems, a process already taking place with the fast-developing technology. "The cumulative sum of human knowledge has been exhausted in AI training. That happened basically last year," said Musk in an interview livestreamed on his social media platform, X. AI models such as the GPT-4o model powering the ChatGPT chatbot are "trained" on a vast array of data taken from the internet, where they in effect learn to spot patterns in that information – allowing them to predict, for instance, the next word in a sentence. Musk said the "only way" to counter the lack of source material for training new models was to move to synthetic data created by AI.

The human cell is a miserable thing to study. Tens of trillions of them exist in the body, forming an enormous and intricate network that governs every disease and metabolic process. Each cell in that circuit is itself the product of an equally dense and complex interplay among genes, proteins, and other bits of profoundly small biological machinery. Our understanding of this world is hazy and constantly in flux. As recently as a few years ago, scientists thought there were only a few hundred distinct cell types, but new technologies have revealed thousands (and that's just the start).

With the rapid development of generative artificial intelligence, particularly large language models, a number of sub-fields of deep learning have made significant progress and are now very useful in everyday applications. For example, well-known financial institutions simulate a wide range of scenarios for various models created by their research teams using reinforcement learning, both before production and after regular operations. In this work, we propose a backdoor attack that focuses solely on data poisoning. This particular backdoor attack is classified as an attack without prior consideration or trigger, and we name it FinanceLLMsBackRL. Our aim is to examine the potential effects of large language models that use reinforcement learning systems for text production or speech recognition, finance, physics, or the ecosystem of contemporary artificial intelligence models.

This paper introduces a novel approach to creating adaptive language agents by integrating active inference with large language models (LLMs). While LLMs demonstrate remarkable capabilities, their reliance on static prompts limits adaptation to new information and changing environments. We address this by implementing an active inference framework that acts as a cognitive layer above an LLM-based agent, dynamically adjusting prompts and search strategies through principled information-seeking behavior. Our framework models the environment using three state factors (prompt, search, and information states) with seven observation modalities capturing quality metrics. By framing the agent's learning through the free energy principle, we enable systematic exploration of prompt combinations and search strategies. Experimental results demonstrate the effectiveness of this approach, with the agent developing accurate models of environment dynamics evidenced by emergent structure in observation matrices. Action selection patterns reveal sophisticated exploration-exploitation behavior, transitioning from initial information-gathering to targeted prompt testing. The integration of thermodynamic principles with language model capabilities provides a principled framework for creating robust, adaptable agents, extending active inference beyond traditional low-dimensional control problems to high-dimensional, language-driven environments.

The rapid advancement of Large Language Models (LLMs), particularly those trained on multilingual corpora, has intensified the need for a deeper understanding of their performance across a diverse range of languages and model sizes. Our research addresses this critical need by studying the performance and scaling behavior of multilingual LLMs in text classification and machine translation tasks across 204 languages. We systematically examine both seen and unseen languages across three model families of varying sizes in zero-shot and few-shot settings. Our findings show significant differences in scaling behavior between zero-shot and two-shot scenarios, with striking disparities in performance between seen and unseen languages. Model scale has little effect on zero-shot performance, which remains mostly flat. However, in two-shot settings, larger models show clear linear improvements in multilingual text classification. For translation tasks, however, only the instruction-tuned model showed clear benefits from scaling. Our analysis also suggests that overall resource levels, not just the proportions of pretraining languages, are better predictors of model performance, shedding light on what drives multilingual LLM effectiveness.