arXiv:2511.11597v1 Announce Type: new Abstract: Evaluating how Large Language Models (LLMs) handle complex, specialized knowledge remains a critical challenge. We address this through the lens of climate change by introducing CLINB, a benchmark that assesses models on open-ended, grounded, multimodal question answering tasks with clear requirements for knowledge quality and evidential support. CLINB relies on a dataset of real users' questions and evaluation rubrics curated by leading climate scientists. We implement and validate a model-based evaluation process and evaluate several frontier models. Our findings reveal a critical dichotomy. Frontier models demonstrate remarkable knowledge synthesis capabilities, often exhibiting PhD-level understanding and presentation quality. They outperform "hybrid" answers curated by domain experts assisted by weaker models. However, this performance is countered by failures in grounding. The quality of evidence varies, with substantial hallucination rates for references and images. We argue that bridging this gap between knowledge synthesis and verifiable attribution is essential for the deployment of AI in scientific workflows and that reliable, interpretable benchmarks like CLINB are needed to progress towards building trustworthy AI systems.

arXiv:2412.15925v1 Announce Type: cross Abstract: Problem: Pancreas radiological imaging is challenging due to the small size, blurred boundaries, and variability of shape and position of the organ among patients. Goal: In this work we present MiniGPT-Pancreas, a Multimodal Large Language Model (MLLM), as an interactive chatbot to support clinicians in pancreas cancer diagnosis by integrating visual and textual information. Methods: MiniGPT-v2, a general-purpose MLLM, was fine-tuned in a cascaded way for pancreas detection, tumor classification, and tumor detection with multimodal prompts combining questions and computed tomography scans from the National Institute of Health (NIH), and Medical Segmentation Decathlon (MSD) datasets. The AbdomenCT-1k dataset was used to detect the liver, spleen, kidney, and pancreas. Results: MiniGPT-Pancreas achieved an Intersection over Union (IoU) of 0.595 and 0.550 for the detection of pancreas on NIH and MSD datasets, respectively. For the pancreas cancer classification task on the MSD dataset, accuracy, precision, and recall were 0.876, 0.874, and 0.878, respectively. When evaluating MiniGPT-Pancreas on the AbdomenCT-1k dataset for multi-organ detection, the IoU was 0.8399 for the liver, 0.722 for the kidney, 0.705 for the spleen, and 0.497 for the pancreas. For the pancreas tumor detection task, the IoU score was 0.168 on the MSD dataset. Conclusions: MiniGPT-Pancreas represents a promising solution to support clinicians in the classification of pancreas images with pancreas tumors. Future research is needed to improve the score on the detection task, especially for pancreas tumors.

arXiv:2511.11590v1 Announce Type: cross Abstract: Artificial intelligence (AI) is increasingly embedded in NHS workflows, but its probabilistic and adaptive behaviour conflicts with the deterministic assumptions underpinning existing clinical-safety standards. DCB0129 and DCB0160 provide strong governance for conventional software yet do not define how AI-specific transparency, interpretability, or model drift should be evidenced within Safety Cases, Hazard Logs, or post-market monitoring. This paper proposes an Explainability-Enabled Clinical Safety Framework (ECSF) that integrates explainability into the DCB0129/0160 lifecycle, enabling Clinical Safety Officers to use interpretability outputs as structured safety evidence without altering compliance pathways. A cross-regulatory synthesis mapped DCB clauses to principles from Good Machine Learning Practice, the NHS AI Assurance and T.E.S.T. frameworks, and the EU AI Act. The resulting matrix links regulatory clauses, principles, ECSF checkpoints, and suitable explainability outputs. ECSF introduces five checkpoints: global transparency for hazard identification, case-level interpretability for verification, clinician usability for evaluation, traceable decision pathways for risk control, and longitudinal interpretability monitoring for post-market surveillance. Techniques such as SHAP, LIME, Integrated Gradients, saliency mapping, and attention visualisation are mapped to corresponding DCB artefacts. ECSF reframes explainability as a core element of clinical-safety assurance, bridging deterministic risk governance with the probabilistic behaviour of AI and supporting alignment with GMLP, the EU AI Act, and NHS AI Assurance principles.

arXiv:2511.13373v1 Announce Type: cross Abstract: Large Language Models (LLMs) face significant challenges in distributed healthcare, including consolidating specialized domain knowledge across institutions while maintaining privacy, reducing computational overhead, and preventing catastrophic forgetting during model updates.This paper presents a systematic evaluation of six parameter-space merging techniques applied to two architecturally compatible medical LLMs derived from the Mistral-7B base model. We introduce a novel hierarchical method that combines selective Optimal Transport (OT) alignment for attention layers with cosine similarity-weighted interpolation, designed to address permutation variance while minimizing computational overhead for edge deployment scenarios. Our study evaluates Task Arithmetic, Linear Averaging, DARE-TIES, DELLA, Breadcrumbs, and our Hierarchical approach across five medical benchmarks. Results demonstrate that architecturally compatible models benefit significantly from simple averaging methods, with Task Arithmetic achieving 45.80% accuracy on MedQA, outperforming complex pruning-based approaches. These findings offer critical insights for the deployment of distributed medical AI in resource-constrained IoT environments, where computational efficiency and model compatibility are paramount. Our work establishes that for architecturally compatible models, simple averaging provides a robust and computationally efficient baseline for knowledge consolidation, offering a pragmatic path forward for scalable medical AI systems.

arXiv:2511.13525v1 Announce Type: cross Abstract: Fairness in artificial intelligence (AI) has become a growing concern due to discriminatory outcomes in AI-based decision-making systems. While various methods have been proposed to mitigate bias, most rely on complete demographic information, an assumption often impractical due to legal constraints and the risk of reinforcing discrimination. This survey examines fairness in AI when demographics are incomplete, addressing the gap between traditional approaches and real-world challenges. We introduce a novel taxonomy of fairness notions in this setting, clarifying their relationships and distinctions. Additionally, we summarize existing techniques that promote fairness beyond complete demographics and highlight open research questions to encourage further progress in the field.

npj Digital Medicine, Published online: 17 November 2025; doi:10.1038/s41746-025-02025-y

A large language model-based approach to quantifying the effects of social determinants in liver transplant decisions

Background: Sensor-based digital health technologies (sDHTs) are increasingly used to support scientific and clinical decision-making. The digital measures (DMs) they generate offer significant potential to accelerate the drug development timeline, decrease clinical trial costs, and improve access to care. However, choosing appropriate statistical methodology when conducting analytical validation (AV) of a DM is complicated, particularly for novel DMs, for which appropriate, established reference measures may not exist. More understanding of, and a standardization of approach to, AV in these scenarios is needed. Objective: In a prior simulation study, three statistical methods were tested for their ability to estimate a simulated relationship between a sDHT-derived DM and several clinical outcome assessment (COA) reference measures. The aim of this work was to assess the feasibility of these methods’ implementation in real data, and to examine the impact of AV study design factors on the relationships estimated. Methods: Four real-world datasets, captured using sDHTs, were used to prepare hypothetical AV studies that represented a range of scenarios with respect to three key study design properties: temporal coherence, construct coherence and data completeness. For each hypothetical study, two-factor, correlated-factors confirmatory factor analysis (CFA) models, and a combination of simple and multiple linear regression models, were built using the DM and reference measure data. The factor correlation, R2 and adjusted R2 statistics were calculated, in addition to the Pearson Correlation Coefficients (PCCs). Results: The majority of CFA models exhibited an acceptable fit according to the majority of the fit statistics employed, and each model was able to estimate a factor correlation. For each model, these correlations were greater than or equal to the corresponding PCC in magnitude. Correlations were strongest in the hypothetical studies with strong temporal and construct coherence. Conclusions: The performance of the measures shown in this work supports the feasibility of the selected statistical methods when implemented in real-world data. Our findings in particular support the use of CFA to assess the relationship between a novel digital measure and a COA reference measure. The observed impact of AV study design factors on the relationships estimated allowed the authors to determine practical recommendations to aid in appropriate study design for AV of novel digital measures. By utilizing a standardized methodology for evaluating novel digital measures, sDHT developers, biostatisticians and clinical researchers will be able to navigate the complex validation landscape more easily, with more certainty, and with more tools at their disposal, expediting the pathway to validation and regulatory review.

arXiv:2511.08609v1 Announce Type: cross Abstract: The energy transition is a key theme of the last decades to determine a future of eco-sustainability, and an area of such importance cannot disregard digitization, innovation and the new technological tools available. This is the context in which the Generative Artificial Intelligence models described in this paper are positioned, developed by Engineering Ingegneria Informatica SpA in order to automate the plant structures acquisition of SNAM energy infrastructure, a leading gas transportation company in Italy and Europe. The digitization of a gas plant consists in registering all its relevant information through the interpretation of the related documentation. The aim of this work is therefore to design an effective solution based on Artificial Intelligence techniques to automate the extraction of the information necessary for the digitization of a plant, in order to streamline the daily work of MGM users. The solution received the P&ID of the plant as input, each one in pdf format, and uses OCR, Vision LLM, Object Detection, Relational Reasoning and optimization algorithms to return an output consisting of two sets of information: a structured overview of the relevant design data and the hierarchical framework of the plant. To achieve convincing results, we extend a state-of-the-art model for Scene Graph Generation introducing a brand new Transformer architecture with the aim of deepening the analysis of the complex relations between the plant's components. The synergistic use of the listed AI-based technologies allowed to overcome many obstacles arising from the high variety of data, due to the lack of standardization. An accuracy of 91\% has been achieved in the extraction of textual information relating to design data. Regarding the plants topology, 93\% of components are correctly identified and the hierarchical structure is extracted with an accuracy around 80\%.

arXiv:2511.08721v1 Announce Type: cross Abstract: In behavioral sciences, experiments such as the ultimatum game are conducted to assess preferences for fairness or self-interest of study participants. In the dictator game, a simplified version of the ultimatum game where only one of two players makes a single decision, the dictator unilaterally decides how to split a fixed sum of money between themselves and the other player. Although recent studies have explored behavioral patterns of AI agents based on Large Language Models (LLMs) instructed to adopt different personas, we question the robustness of these results. In particular, many of these studies overlook the role of the system prompt - the underlying instructions that shape the model's behavior - and do not account for how sensitive results can be to slight changes in prompts. However, a robust baseline is essential when studying highly complex behavioral aspects of LLMs. To overcome previous limitations, we propose the LLM agent behavior study (LLM-ABS) framework to (i) explore how different system prompts influence model behavior, (ii) get more reliable insights into agent preferences by using neutral prompt variations, and (iii) analyze linguistic features in responses to open-ended instructions by LLM agents to better understand the reasoning behind their behavior. We found that agents often exhibit a strong preference for fairness, as well as a significant impact of the system prompt on their behavior. From a linguistic perspective, we identify that models express their responses differently. Although prompt sensitivity remains a persistent challenge, our proposed framework demonstrates a robust foundation for LLM agent behavior studies. Our code artifacts are available at https://github.com/andreaseinwiller/LLM-ABS.

arXiv:2510.00229v4 Announce Type: replace Abstract: The deployment of Large Language Models (LLMs) as agentic orchestrators has revolutionized task automation, but the need for privacy-preserving, cost-effective solutions demands on-device inference capabilities. However, local LLMs consistently underperform compared to frontier models in tool calling scenarios, struggling with both tool selection from large tool sets and accurate argument generation for complex parameter structures. We introduce a methodology that disaggregates a tool-calling task into two distinct subtasks: tool selection and argument generation. We propose "decoupled fine-tuning", a novel post-training approach that employs LoRA fine-tuning to create dedicated LoRA adapters for tool selection and tool-specific argument generation using separate loss masking for each of the subtasks. Furthermore, we present AgentFlux, an inference framework that leverages the LoRA adapters created using decoupled fine-tuning to perform efficient agent orchestration with the help of local models on end-user devices. AgentFlux decomposes the tool-call generation step into tool selection and argument generation, and dynamically loads the corresponding LoRA adapters to generate tool calls. Additionally, AgentFlux implements hierarchical orchestration to restrict the number of tools required for tool selection. Our experiments on the MCP-Bench benchmark demonstrate that the Qwen-2.5-7B model trained using decoupled fine-tuning improves the tool calling accuracy of the base model by 46%, and outperforms other local reasoning, non-reasoning and fine-tuned models of similar size in all cases, and models that are 2x larger, in most cases.

arXiv:2410.15281v4 Announce Type: replace-cross Abstract: With the broader adoption and highly successful development of Large Language Models (LLMs), there has been growing interest and demand for applying LLMs to autonomous driving technology. Driven by their natural language understanding and reasoning capabilities, LLMs have the potential to enhance various aspects of autonomous driving systems, from perception and scene understanding to interactive decision-making. In this paper, we first introduce the novel concept of designing Large Language Models for Autonomous Driving (LLM4AD), followed by a review of existing LLM4AD studies. Then, we propose a comprehensive benchmark for evaluating the instruction-following and reasoning abilities of LLM4AD systems, which includes LaMPilot-Bench, CARLA Leaderboard 1.0 Benchmark in simulation and NuPlanQA for multi-view visual question answering. Furthermore, we conduct extensive real-world experiments on autonomous vehicle platforms, examining both on-cloud and on-edge LLM deployment for personalized decision-making and motion control. Next, we explore the future trends of integrating language diffusion models into autonomous driving, exemplified by the proposed ViLaD (Vision-Language Diffusion) framework. Finally, we discuss the main challenges of LLM4AD, including latency, deployment, security and privacy, safety, trust and transparency, and personalization.

arXiv:2508.01159v2 Announce Type: replace-cross Abstract: This study evaluates the capacity of large language models (LLMs) to generate structured clinical consultation templates for electronic consultation. Using 145 expert-crafted templates developed and routinely used by Stanford's eConsult team, we assess frontier models -- including o3, GPT-4o, Kimi K2, Claude 4 Sonnet, Llama 3 70B, and Gemini 2.5 Pro -- for their ability to produce clinically coherent, concise, and prioritized clinical question schemas. Through a multi-agent pipeline combining prompt optimization, semantic autograding, and prioritization analysis, we show that while models like o3 achieve high comprehensiveness (up to 92.2\%), they consistently generate excessively long templates and fail to correctly prioritize the most clinically important questions under length constraints. Performance varies across specialties, with significant degradation in narrative-driven fields such as psychiatry and pain medicine. Our findings demonstrate that LLMs can enhance structured clinical information exchange between physicians, while highlighting the need for more robust evaluation methods that capture a model's ability to prioritize clinically salient information within the time constraints of real-world physician communication.

Stereo-seq V2 facilitates single-cell-resolution spatial RNA mapping in FFPE samples through random primer capture, uncovering ncRNAs, host-pathogen transcriptome profiling, and spatial immune repertoires in situ.

npj Digital Medicine, Published online: 12 November 2025; doi:10.1038/s41746-025-02013-2

Equipping mathematical models for hospital dynamics using information theory

Last week, we hosted EmTech MIT, MIT Technology Review’s annual flagship conference in Cambridge, Massachusetts. Over the course of three days of main-stage sessions, I learned about innovations in AI, biotech, and robotics. 

But as you might imagine, some of this climate reporter’s favorite moments came in the climate sessions. I was listening especially closely to my colleague James Temple’s discussion with Lucia Tian, head of advanced energy technologies at Google. 

They spoke about the tech giant’s growing energy demand and what sort of technologies the company is looking to to help meet it. In case you weren’t able to join us, let’s dig into that session and consider how the company is thinking about energy in the face of AI’s rapid rise. 

I’ve been closely following Google’s work in energy this year. Like the rest of the tech industry, the company is seeing ballooning electricity demand in its data centers. That could get in the way of a major goal that Google has been talking about for years. 

See, back in 2020, the company announced an ambitious target: by 2030, it aimed to run on carbon-free energy 24-7. Basically, that means Google would purchase enough renewable energy on the grids where it operates to meet its entire electricity demand, and the purchases would match up so the electricity would have to be generated when the company was actually using energy. (For more on the nuances of Big Tech’s renewable-energy pledges, check out James’s piece from last year.)

Google’s is an ambitious goal, and on stage, Tian said that the company is still aiming for it but acknowledged that it’s looking tough with the rise of AI. 

“It was always a moonshot,” she said. “It’s something very, very hard to achieve, and it’s only harder in the face of this growth. But our perspective is, if we don’t move in that direction, we’ll never get there.”

Google’s total electricity demand more than doubled from 2020 to 2024, according to its latest Environmental Report. As for that goal of 24-7 carbon-free energy? The company is basically treading water. While it was at 67% for its data centers in 2020, last year it came in at 66%. 

Not going backwards is something of an accomplishment, given the rapid growth in electricity demand. But it still leaves the company some distance away from its finish line.

To close the gap, Google has been signing what feels like constant deals in the energy space. Two recent announcements that Tian talked about on stage were a project involving carbon capture and storage at a natural-gas plant in Illinois and plans to reopen a shuttered nuclear power plant in Iowa. 

Let’s start with carbon capture. Google signed an agreement to purchase most of the electricity from a new natural-gas plant, which will capture and store about 90% of its carbon dioxide emissions. 

That announcement was controversial, with critics arguing that carbon capture keeps fossil-fuel infrastructure online longer and still releases greenhouse gases and other pollutants into the atmosphere. 

One question that James raised on stage: Why build a new natural-gas plant rather than add equipment to an already existing facility? Tacking on equipment to an operational plant would mean cutting emissions from the status quo, rather than adding entirely new fossil-fuel infrastructure. 

The company did consider many existing plants, Tian said. But, as she put it, “Retrofits aren’t going to make sense everywhere.” Space can be limited at existing plants, for example, and many may not have the right geology to store carbon dioxide underground. 

“We wanted to lead with a project that could prove this technology at scale,” Tian said. This site has an operational Class VI well, the type used for permanent sequestration, she added, and it also doesn’t require a big pipeline buildout. 

Tian also touched on the company’s recent announcement that it’s collaborating with NextEra Energy to reopen Duane Arnold Energy Center, a nuclear power plant in Iowa. The company will purchase electricity from that plant, which is scheduled to reopen in 2029. 

As I covered in a story earlier this year, Duane Arnold was basically the final option in the US for companies looking to reopen shuttered nuclear power plants. “Just a few years back, we were still closing down nuclear plants in this country,” Tian said on stage. 

While each reopening will look a little different, Tian highlighted the groups working to restart the Palisades plant in Michigan, which was the first reopening to be announced, last spring. “They’re the real heroes of the story,” she said.

I’m always interested to get a peek behind the curtain at how Big Tech is thinking about energy. I’m skeptical but certainly interested to see how Google’s, and the rest of the industry’s, goals shape up over the next few years. 

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here.

npj Digital Medicine, Published online: 12 November 2025; doi:10.1038/s41746-025-02103-1

Reply to “When do large language models cross the line: “reasoning” red teaming in healthcare”

arXiv:2511.07669v1 Announce Type: new Abstract: Current large language models (LLMs) excel in verifiable domains where outputs can be checked before action but prove less reliable for high-stakes strategic decisions with uncertain outcomes. This gap, driven by mutually reinforcing cognitive biases in both humans and artificial intelligence (AI) systems, threatens the defensibility of valuations and sustainability of investments in the sector. This report describes a framework emerging from systematic qualitative assessment across 7 frontier-grade LLMs and 3 market-facing venture vignettes under time pressure. Detailed prompting specifying decision partnership and explicitly instructing avoidance of sycophancy, confabulation, solution drift, and nihilism achieved initial partnership state but failed to maintain it under operational pressure. Sustaining protective partnership state required an emergent 7-stage calibration sequence, built upon a 4-stage initialization process, within a 5-layer protection architecture enabling bias self-monitoring, human-AI adversarial challenge, partnership state verification, performance degradation detection, and stakeholder protection. Three discoveries resulted: partnership state is achievable through ordered calibration but requires emergent maintenance protocols; reliability degrades when architectural drift and context exhaustion align; and dissolution discipline prevents costly pursuit of fundamentally wrong directions. Cross-model validation revealed systematic performance differences across LLM architectures. This approach demonstrates that human-AI teams can achieve cognitive partnership capable of preventing avoidable regret in high-stakes decisions, addressing return-on-investment expectations that depend on AI systems supporting consequential decision-making without introducing preventable cognitive traps when verification arrives too late.

arXiv:2509.22242v2 Announce Type: replace Abstract: Clinical dataset labels are rarely certain as annotators disagree and confidence is not uniform across cases. Typical aggregation procedures, such as majority voting, obscure this variability. In simple experiments on medical imaging benchmarks, accounting for the confidence in binary labels significantly impacts model rankings. We therefore argue that machine-learning evaluations should explicitly account for annotation uncertainty using probabilistic metrics that directly operate on distributions. These metrics can be applied independently of the annotations' generating process, whether modeled by simple counting, subjective confidence ratings, or probabilistic response models. They are also computationally lightweight, as closed-form expressions have linear-time implementations once examples are sorted by model score. We thus urge the community to release raw annotations for datasets and to adopt uncertainty-aware evaluation so that performance estimates may better reflect clinical data.

arXiv:2511.05385v1 Announce Type: cross Abstract: Retrieval-Augmented Generation (RAG) utilizes external knowledge to augment Large Language Models' (LLMs) reliability. For flexibility, agentic RAG employs autonomous, multi-round retrieval and reasoning to resolve queries. Although recent agentic RAG has improved via reinforcement learning, they often incur substantial token overhead from search and reasoning processes. This trade-off prioritizes accuracy over efficiency. To address this issue, this work proposes TeaRAG, a token-efficient agentic RAG framework capable of compressing both retrieval content and reasoning steps. 1) First, the retrieved content is compressed by augmenting chunk-based semantic retrieval with a graph retrieval using concise triplets. A knowledge association graph is then built from semantic similarity and co-occurrence. Finally, Personalized PageRank is leveraged to highlight key knowledge within this graph, reducing the number of tokens per retrieval. 2) Besides, to reduce reasoning steps, Iterative Process-aware Direct Preference Optimization (IP-DPO) is proposed. Specifically, our reward function evaluates the knowledge sufficiency by a knowledge matching mechanism, while penalizing excessive reasoning steps. This design can produce high-quality preference-pair datasets, supporting iterative DPO to improve reasoning conciseness. Across six datasets, TeaRAG improves the average Exact Match by 4% and 2% while reducing output tokens by 61% and 59% on Llama3-8B-Instruct and Qwen2.5-14B-Instruct, respectively. Code is available at https://github.com/Applied-Machine-Learning-Lab/TeaRAG.