arXiv:2510.16206v2 Announce Type: replace Abstract: Since the rapid expansion of large language models (LLMs), people have begun to rely on them for information retrieval. While traditional search engines display ranked lists of sources shaped by search engine optimization (SEO), advertising, and personalization, LLMs typically provide a synthesized response that feels singular and authoritative. While both approaches carry risks of bias and omission, LLMs may amplify the effect by collapsing multiple perspectives into one answer, reducing users ability or inclination to compare alternatives. This concentrates power over information in a few LLM vendors whose systems effectively shape what is remembered and what is overlooked. As a result, certain narratives, individuals or groups, may be disproportionately suppressed, while others are disproportionately elevated. Over time, this creates a new threat: the gradual erasure of those with limited digital presence, and the amplification of those already prominent, reshaping collective memory. To address these concerns, this paper presents a concept of the Right To Be Remembered (RTBR) which encompasses minimizing the risk of AI-driven information omission, embracing the right of fair treatment, while ensuring that the generated content would be maximally truthful.

arXiv:2406.18851v2 Announce Type: replace-cross Abstract: Optimizing black-box functions is a fundamental problem in science and engineering. To solve this problem, many approaches learn a surrogate function that estimates the underlying objective from limited historical evaluations. Large Language Models (LLMs), with their strong pattern-matching capabilities via pretraining on vast amounts of data, stand out as a potential candidate for surrogate modeling. However, directly prompting a pretrained language model to produce predictions is not feasible in many scientific domains due to the scarcity of domain-specific data in the pretraining corpora and the challenges of articulating complex problems in natural language. In this work, we introduce LICO, a general-purpose model that extends arbitrary base LLMs for black-box optimization, with a particular application to the molecular domain. To achieve this, we equip the language model with a separate embedding layer and prediction layer, and train the model to perform in-context predictions on a diverse set of functions defined over the domain. Once trained, LICO can generalize to unseen molecule properties simply via in-context prompting. LICO performs competitively on PMO, a challenging molecular optimization benchmark comprising 23 objective functions, and achieves state-of-the-art performance on its low-budget version PMO-1K.

arXiv:2502.02904v4 Announce Type: replace-cross Abstract: Writing is a cognitively demanding activity that requires constant decision-making, heavy reliance on working memory, and frequent shifts between tasks of different goals. To build writing assistants that truly align with writers' cognition, we must capture and decode the complete thought process behind how writers transform ideas into final texts. We present ScholaWrite, the first dataset of end-to-end scholarly writing, tracing the multi-month journey from initial drafts to final manuscripts. We contribute three key advances: (1) a Chrome extension that unobtrusively records keystrokes on Overleaf, enabling the collection of realistic, in-situ writing data; (2) a novel corpus of full scholarly manuscripts, enriched with fine-grained annotations of cognitive writing intentions. The dataset includes \LaTeX-based edits from five computer science preprints, capturing nearly 62K text changes over four months; and (3) analyses and insights into the micro-dynamics of scholarly writing, highlighting gaps between human writing processes and the current capabilities of large language models (LLMs) in providing meaningful assistance. ScholaWrite underscores the value of capturing end-to-end writing data to develop future writing assistants that support, not replace, the cognitive work of scientists.

arXiv:2505.07897v3 Announce Type: replace-cross Abstract: Context lengths for models have grown rapidly, from thousands to millions of tokens in just a few years. The extreme context sizes of modern long-context models have made it difficult to construct realistic long-context benchmarks -- not only due to the cost of collecting million-context tasks but also in identifying realistic scenarios that require significant contexts. We identify code comprehension and repair as a natural testbed and challenge task for long-context models and introduce LongCodeBench (LCB), a benchmark to test LLM coding abilities in long-context scenarios. Our benchmark tests both the comprehension and repair capabilities of LCLMs in realistic and important settings by drawing from real-world GitHub issues and constructing QA (LongCodeQA) and bug fixing (LongSWE-Bench) tasks. We carefully stratify the complexity of our benchmark, enabling us to evaluate models across different scales -- ranging from Qwen2.5 14B Instruct to Google's flagship Gemini model. We find that long-context remains a weakness for all models, with performance drops such as from 29% to 3% for Claude 3.5 Sonnet, or from 70.2% to 40% for Qwen2.5. The LCB dataset is available publicly at https://huggingface.co/datasets/Steefano/LCB and the codebase to replicate the work on this paper at https://github.com/Zteefano/long-code-bench.

Background: Comprehensive preoperative education is essential for optimizing outcomes and ensuring informed consent in patients undergoing total hip arthroplasty (THA). Emerging artificial intelligence (AI) tools, such as ChatGPT, offer scalable support for patient education, but their clinical application requires rigorous evaluation to ensure accuracy, safety, and trust. Objective: This study assessed patients’ preferences and satisfaction with AI-assisted informed consent in THA, comparing traditional physician consultations to those supported by native ChatGPT and a customized version enhanced with retrieval-augmented generation (RAG). It also examined how state anxiety and general attitudes toward AI affect preferences for AI-supported consent and whether RAG integration improves ChatGPT response quality. Methods: A total of 36 patients scheduled for elective THA were assigned to one of three groups (12 each): (1) standard physician-only consultations (control), (2) physician-assisted consultations supported by native ChatGPT, and (3) supported by ChatGPT enhanced through RAG. Data collection involved standardized Likert scale questionnaires assessing patient satisfaction with the consent process, perceived informedness, anxiety levels, and attitudes toward AI. The ChatGPT responses were independently evaluated by physicians for relevance, accuracy, clarity, completeness, adherence to evidence-based guidelines, and appropriate length. Instances of hallucinations, factually incorrect or misleading outputs, were identified and rated by severity. Statistical analyses compared outcomes across groups and explored associations. Results: Patients interacting with the ChatGPT+RAG model reported significantly higher satisfaction levels with information delivery (P=.01) and perceived level of informedness (P=.01) than those using the native ChatGPT model. The mean number of patient questions in the control group was 20, compared with 39 in the native ChatGPT group (P=.06) and 52 in the ChatGPT+RAG group (P=.002). The majority of participants across all groups preferred a human clinician providing less accurate information over a more accurate AI-only assistant. These preferences were not influenced by sociodemographic variables (age, gender, and education), health literacy, state anxiety, or general attitudes toward AI. The ChatGPT+RAG model outperformed the native ChatGPT model across all evaluated response quality dimensions (all P<.01) and exhibited a significantly lower hallucination rate (5/52, 10% versus 15/39, 38%; P=.002). Conclusions: Integrating RAG with ChatGPT significantly improves the quality, clarity, and reliability of preoperative information, enhancing patient satisfaction and engagement beyond native ChatGPT. However, patients maintain a strong preference for physician-led informed consent, underscoring the role of AI chatbots as complementary tools rather than replacements. These findings support the cautious adoption of customized AI assistants to augment, not substitute, human interaction in surgical consent processes. Trial Registration:

Background: Artificial intelligence (AI) is increasingly being integrated into health care, offering a wide array of benefits. Current AI applications encompass patients’ diagnosis, treatment, data mining, and more to enhance patient care and quality of life. It is also democratizing access to expert support by providing timely and accurate disease diagnoses, better clinical management, quicker drug discovery, improved disease prevention, big data management, and health protection. Objective: The aim of the study is to document AI adoption in health care, assess participants’ perception on its usefulness in the management of health care delivery and leadership of health care systems, and identify characteristics of early adopters. Methods: We conducted a worldwide cross-sectional survey across all 6 inhabited continents using a self-administered questionnaire developed with the Qualtrics electronic data collection tool. This was piloted and reviewed to ensure completeness, accuracy, acceptability, cultural sensitivity, and relevance. Respondents were recruited by individualized email, following identification from professional associations or organizations, professional networks, and social media. Data were analyzed using SPSS (IBM Corp), with results presented as narrative, charts, and tables. Results: In total, 506 health care professionals completed the survey. While 92.3% (467/506) of respondents believed that AI has a role in patient care and health care management, only 76.5% (300/392) were willing to support AI adoption and embedding in their organization. Although top managers are mainly responsible for adoption processes, staff training remains low. AI is currently used mostly for diagnosis, patient care, and precision medicine. These uses of AI will continue in the near future, but in different ways. AI adoption was highest in Europe and lowest in Africa. Black or African American people were more likely to support AI adoption than White and Asian people. Poor knowledge of AI, fear of job loss, and resistance to change were the top barriers to AI adoption and embedding. Conclusions: AI use in health is global, but the adoption rate varies by geography and individual characteristics. AI adoption communication by executive health care management is poor, as is the level of training of health care staff. To improve AI adoption, management should improve communication with their teams, provide training on AI to their workers, and help individuals understand how AI works. Barriers such as ethical issues around data ownership and use should be addressed. African organizations should be proactive and invest in AI adoption early, so that they are not left behind in the AI revolution.