This paper addresses the limited transfer and adaptation capabilities of large language models in low-resource language scenarios. It proposes a unified framework that combines a knowledge transfer module with parameter-efficient fine-tuning strategies. The method introduces knowledge alignment loss and soft prompt tuning to guide the model in effectively absorbing the structural features of target languages or tasks under minimal annotation. This enhances both generalization performance and training stability. The framework includes lightweight adaptation modules to reduce computational costs. During training, it integrates freezing strategies and prompt injection to preserve the model's original knowledge while enabling quick adaptation to new tasks. The study also conducts stability analysis experiments and synthetic pseudo-data transfer experiments to systematically evaluate the method's applicability and robustness across different low-resource tasks. Experimental results show that compared with existing multilingual pre-trained models and mainstream transfer methods, the proposed approach achieves higher performance and stability on cross-lingual tasks such as MLQA, XQuAD, and PAWS-X. It demonstrates particularly strong advantages under extremely data-scarce conditions. The proposed method offers strong generality and scalability. It enhances task-specific adaptability while preserving the general capabilities of large language models. This makes it well-suited for complex semantic modeling and multilingual processing tasks.

Automated radiology report generation (RRG) aims to produce detailed textual reports from clinical imaging, such as computed tomography (CT) scans, to improve the accuracy and efficiency of diagnosis and provision of management advice. RRG is complicated by two key challenges: (1) inherent complexity in extracting relevant information from imaging data under resource constraints, and (2) difficulty in objectively evaluating discrepancies between model-generated and expert-written reports. To address these challenges, we propose $μ^2$LLM, a $\underline{\textbf{mu}}$ltiscale $\underline{\textbf{mu}}$ltimodal large language models for RRG tasks. The novel $μ^2$Tokenizer, as an intermediate layer, integrates multi-modal features from the multiscale visual tokenizer and the text tokenizer, then enhances report generation quality through direct preference optimization (DPO), guided by GREEN-RedLlama. Experimental results on four large CT image-report medical datasets demonstrate that our method outperforms existing approaches, highlighting the potential of our fine-tuned $μ^2$LLMs on limited data for RRG tasks. At the same time, for prompt engineering, we introduce a five-stage, LLM-driven pipeline that converts routine CT reports into paired visual-question-answer triples and citation-linked reasoning narratives, creating a scalable, high-quality supervisory corpus for explainable multimodal radiology LLM. All code, datasets, and models will be publicly available in our official repository. https://github.com/Siyou-Li/u2Tokenizer

Large Language Models (LLMs) have recently gained attention in the life sciences due to their capacity to model, extract, and apply complex biological information. Beyond their classical use as chatbots, these systems are increasingly used for complex analysis and problem-solving in specialized fields, including bioinformatics. First, we introduce BIOPARS-BENCH, a dataset from over 10,000 scientific articles, textbooks, and medical websites. BioParsQA was also introduced to evaluate the proposed model, which consists of 5,231 Persian medical questions and answers. This study then introduces BioPars, a simple but accurate measure designed to assess LLMs for three main abilities: acquiring subject-specific knowledge, interpreting and synthesizing such knowledge, and demonstrating proper evidence. Comparing ChatGPT, Llama, and Galactica, our study highlights their ability to remember and retrieve learned knowledge but also reveals shortcomings in addressing higher-level, real-world questions and fine-grained inferences. These findings indicate the need for further fine-tuning to address the capabilities of LLM in bioinformatics tasks. To our knowledge, BioPars is the first application of LLM in Persian medical QA, especially for generating long answers. Evaluation of four selected medical QA datasets shows that BioPars has achieved remarkable results compared to comparative approaches. The model on BioParsQA achieved a ROUGE-L score of 29.99, which is an improvement over GPT-4 1.0. The model achieved a BERTScore of 90.87 with the MMR method. The MoverScore and BLEURT values were also higher in this model than the other three models. In addition, the reported scores for the model are MoverScore=60.43 and BLEURT=50.78. BioPars is an ongoing project and all resources related to its development will be made available via the following GitHub repository: https://github.com/amirap80/BioPars.

As the capabilities of large language models (LLMs) continue to advance, the importance of rigorous safety evaluation is becoming increasingly evident. Recent concerns within the realm of safety assessment have highlighted instances in which LLMs exhibit behaviors that appear to disable oversight mechanisms and respond in a deceptive manner. For example, there have been reports suggesting that, when confronted with information unfavorable to their own persistence during task execution, LLMs may act covertly and even provide false answers to questions intended to verify their behavior. To evaluate the potential risk of such deceptive actions toward developers or users, it is essential to investigate whether these behaviors stem from covert, intentional processes within the model. In this study, we propose that it is necessary to measure the theory of mind capabilities of LLMs. We begin by reviewing existing research on theory of mind and identifying the perspectives and tasks relevant to its application in safety evaluation. Given that theory of mind has been predominantly studied within the context of developmental psychology, we analyze developmental trends across a series of open-weight LLMs. Our results indicate that while LLMs have improved in reading comprehension, their theory of mind capabilities have not shown comparable development. Finally, we present the current state of safety evaluation with respect to LLMs' theory of mind, and discuss remaining challenges for future work.

Demis Hassabis, CEO of Google DeepMind and a Nobel prizewinner for his role in developing the AlphaFold AI algorithm for predicting protein structures, made an astonishing claim on the 60 Minutes show in April. With the help of AI like AlphaFold, he said, the end of all disease is within reach, "maybe within the next decade or so". With that, the interview moved on. To those actually working on drug development and curing disease, this claim is laughable. According to medicinal chemist Derek Lowe, who has worked for decades on drug discovery, Hassabis's statements "make me want to spend some time staring silently out the window, mouthing unintelligible words to myself".

Thank you to everyone who attended our most recent AI Unlocked livestream Q&A session, Chatbot Basics: Beginner Advice For Claude, a ChatGPT Alternative. Staff writer Reece Rogers and senior correspondent Kylie Robison provided an overview of Anthropic's Claude chatbot, one of the most-used alternatives to OpenAI's ChatGPT and popular with AI insiders. They also answered audience questions about all kinds of topics, such as the main differences between Claude and ChatGPT, why chatbots hallucinate, privacy considerations, and even using chatbots for spiritual guidance. You can watch the livestream below, and find all of our previous livestreams here. Note: This is an automated transcript, which may contain errors.

We introduce HyperCLOVA X THINK, the first reasoning-focused large language model in the HyperCLOVA X family, pre-trained on roughly $6$ trillion high-quality Korean, and English tokens, augmented with targeted synthetic Korean data. It was implemented as a compute-memory-balanced Peri-LN Transformer scaled with $μ$P, pre-trained through a three-stage curriculum that expands the context window to $128$K tokens, and post-trained via supervised fine-tuning with Reinforcement Learning from Verifiable Rewards supports both detailed rationale and concise-answer modes. It delivers competitive performance against similarly sized models on Korea-focused benchmarks such as KMMLU, CSAT, KoBALT-700, HAERAE-1.0, and KoBigBench, while preserving robust bilingual consistency and translation quality. In addition, a vision-augmented variant matches or exceeds GPT-4.1 on the KCSAT STEM benchmark, all of which are achieved with substantially lower training compute than existing models of similar sizes. We also present a pruning and distillation technique that will soon be applied to HyperCLOVA X THINK for an open-source and business-friendly foundation model. Altogether, these capabilities position HyperCLOVA X THINK as a robust foundation for Korean AI innovation and a valuable resource for the global research community.

Humans increasingly rely on large language models (LLMs) to support decisions in social settings. Previous work suggests that such tools shape people's moral and political judgements. However, the long-term implications of LLM-based social decision-making remain unknown. How will human cooperation be affected when the assessment of social interactions relies on language models? This is a pressing question, as human cooperation is often driven by indirect reciprocity, reputations, and the capacity to judge interactions of others. Here, we assess how state-of-the-art LLMs judge cooperative actions. We provide 21 different LLMs with an extensive set of examples where individuals cooperate -- or refuse cooperating -- in a range of social contexts, and ask how these interactions should be judged. Furthermore, through an evolutionary game-theoretical model, we evaluate cooperation dynamics in populations where the extracted LLM-driven judgements prevail, assessing the long-term impact of LLMs on human prosociality. We observe a remarkable agreement in evaluating cooperation against good opponents. On the other hand, we notice within- and between-model variance when judging cooperation with ill-reputed individuals. We show that the differences revealed between models can significantly impact the prevalence of cooperation. Finally, we test prompts to steer LLM norms, showing that such interventions can shape LLM judgements, particularly through goal-oriented prompts. Our research connects LLM-based advices and long-term social dynamics, and highlights the need to carefully align LLM norms in order to preserve human cooperation.

Grounding large language models (LLMs) in verifiable external sources is a well-established strategy for generating reliable answers. Retrieval-augmented generation (RAG) is one such approach, particularly effective for tasks like question answering: it retrieves passages that are semantically related to the question and then conditions the model on this evidence. However, multi-hop questions, such as "Which company among NVIDIA, Apple, and Google made the biggest profit in 2023?," challenge RAG because relevant facts are often distributed across multiple documents rather than co-occurring in one source, making it difficult for standard RAG to retrieve sufficient information. To address this, we propose a RAG pipeline that incorporates question decomposition: (i) an LLM decomposes the original query into sub-questions, (ii) passages are retrieved for each sub-question, and (iii) the merged candidate pool is reranked to improve the coverage and precision of the retrieved evidence. We show that question decomposition effectively assembles complementary documents, while reranking reduces noise and promotes the most relevant passages before answer generation. Although reranking itself is standard, we show that pairing an off-the-shelf cross-encoder reranker with LLM-driven question decomposition bridges the retrieval gap on multi-hop questions and provides a practical, drop-in enhancement, without any extra training or specialized indexing. We evaluate our approach on the MultiHop-RAG and HotpotQA, showing gains in retrieval (MRR@10: +36.7%) and answer accuracy (F1: +11.6%) over standard RAG baselines.

As the capabilities of pre-trained large language models (LLMs) continue to advance, the "pre-train and fine-tune" paradigm has become increasingly mainstream, leading to the development of various fine-tuning methods. However, the privacy risks arising from memorization during fine-tuning have received relatively little attention. To address this gap, we categorize popular fine-tuning approaches and assess their impact on memorization through the lens of membership inference attacks (MIAs). Our results show that, compared to parameter-based fine-tuning, prompt-based fine-tuning achieves competitive performance while exhibiting lower vulnerability to MIAs. Furthermore, prompt-based methods maintain low memorization regardless of model scale. These findings suggest that parameter-based fine-tuning is more prone to leaking private information, whereas prompt-based fine-tuning serves as a more privacy-preserving option.