Vincent Moens

NeurIPS 2026 call for papers is live now. Make your AI agents get to work. Papers won't write themselves. Tell them to hurry! /s

1

Curious about how thermal inputs are encoded across species? Our new review article, “The neuronal circuits and cellular encoding of thermosensation", is now out in Nature Reviews Neuroscience: https://lnkd.in/eutSzx7Y Together with Mikkel Vestergaard and James Poulet, we discuss how rodents, primates, and insects process thermal stimuli, highlighting both shared mechanisms and key differences. Here you can find the full-text access to a view-only version: https://rdcu.be/eThU1

70

8 Comments

I’m excited to share that our scoping review preprint has been published on medRxiv: A Scoping Review of Algorithmic Equity, Data Diversity, and Inclusive Design in the Transformer Era of Clinical NLP https://lnkd.in/eWvWM8h6 Many thanks to Abeer Badawi, Ph.D., Elham Dolatabadi, and Farah Ahmad, MBBS, MPH, PhD, for their collaboration, guidance, and support throughout this work. The rapid digitization of healthcare has positioned transformer-based NLP models as powerful tools for managing clinical text. However, their growing integration into practice raises important and unresolved questions about equity and inclusivity. This review synthesizes 56 studies (2017–2024) and examines how equity is addressed across three dimensions: ➡️ Algorithmic equity — fairness audits are largely post hoc and fragmented ➡️ Data diversity & representativeness — persistent underrepresentation creates what we define as Data Diversity Debt ➡️ Participatory design — observed in only 11% of studies, revealing a major gap beyond clinician involvement. To move beyond descriptive audits, we propose an equity-by-design roadmap to embed fairness, inclusivity, and accountability across the full lifecycle of clinical NLP systems—and to retire Data Diversity Debt. #HealthAI #ClinicalNLP #AlgorithmicEquity #DigitalHealthEquity #AIethics #ResponsibleAI #medRxiv

26

2 Comments

It’s a big day for AI evaluations 🚀 The long awaited Inspect Scout has finally been released with all its great features. It's essentially a tool for in-depth analysis for AI transcripts. To make sure that people use it to its full potential, we aggregated best practices from multiple researchers (AISI, CAISI, METR, Princeton, RAND, Cambridge) and created a practical guide with concrete examples and code implementation 🔥 We tried to cover various applications (e.g., agentic evaluations, chatbot multi-turn conversations), different scoring methodologies, validation techniques etc. We really hope this will be useful to the community! Blog post: https://lnkd.in/eE3Tiq3A Best practices paper: https://lnkd.in/e_QxDDUu Inspect Scout: https://lnkd.in/etetUBdu

42

1 Comment

We are thrilled to introduce the "Open Orthophosphate Model," an AI and ML-driven framework set to transform the prediction of orthophosphate concentrations in UK water catchments. By utilising millions of historical observations and incorporating advanced molecular embeddings, correlations, and domain expertise, this model provides valuable insights into water quality. This innovative approach aims to aid the water quality monitoring, offering scalable, cost-effective solutions in addressing environmental, regulatory, and technical challenges. Please access the model under the GitHub location: https://lnkd.in/ek6a6QSn and please provide your comments. #WaterQuality #AI #MachineLearning #EnvironmentalTech #Innovation

17

I've become interested in mechanistic interpretability recently and wanted to experiment with some of the latest innovations in the space. In a limited fashion, we can look inside a LLM and see what's going on. Using Sparse Autoencoders (SAEs) from Gemma Scope, this tool lets you peek under the hood of Google's Gemma 2 2B model and actually steer its behavior in real time. You can: - Browse and search 16K+ interpretable features by description - Visualize which features activate on any input text - Amplify or suppress specific features during generation and see how outputs change - Decompose predictions into per-feature logit contributions Built with TransformerLens, SAELens, and Gradio. Try it on Huggingface Spaces: https://lnkd.in/e88v8wym Code here: https://lnkd.in/e7dSS9WA #MechanisticInterpretability #AISafety #MachineLearning #NLP #OpenSource

6

New paper accepted to EMNLP 2025 Findings! Not in my usual speech area but linguistic variation and fairness nonetheless 😃 Large language models (LLMs) are ubiquitous in modern day natural language processing. However, previous work has shown degraded LLM performance for under-represented English dialects. We analyze the effects of typifying “standard” American English language questions as non-“standard” dialectal variants on multiple choice question answering tasks and find up to a 20% reduction in accuracy. Additionally, we investigate the grammatical basis of under-performance in non-“standard” English questions. We find that individual grammatical rules have varied effects on performance, but some are more consequential than others: three specific grammar rules (existential “it”, zero copula, and y’all) can explain the majority of performance degradation observed in multiple dialects. We call for future work to investigate bias mitigation methods focused on individual, high-impact grammatical structures.

71

3 Comments

FunctionGemma, a new 270M parameter language model from Google, designed specifically for function calling and tool execution — not general chat.   Why is FunctionGemma special? 🔹 Extremely lightweight FunctionGemma can run in ~550MB RAM, making it possible to deploy on laptops, edge devices, and even phones — all without large GPUs. 🔹 Purpose-built for function calling Unlike general LLMs, this model is trained for text-only tool calling, making it ideal for automation, agents, and API-driven workflows. 🔹 Easy local deployment The model is released in GGUF format, so it works seamlessly with llama.cpp and Ollama using quantized versions (Q4/Q5/Q8). 🔹 Fine-tuning ready from day one Thanks to Unsloth’s day-zero support, you can: Run FunctionGemma on CPU or GPU Fine-tune it locally using Colab notebooks Choose full fine-tuning or LoRA   What can you fine-tune it for? FunctionGemma is meant to be customized for specific tasks, such as: Multi-turn function-calling agents Mobile actions (calendar events, reminders, flashlight, timers) Task-specific automation and workflows Lightweight, offline AI assistants This model clearly shows a shift toward small, focused, and efficient AI — where models are trained to do one thing well and can be adapted quickly to real-world use cases.   Model (GGUF): https://lnkd.in/gxVuBhZA Docs & fine-tuning guides: https://lnkd.in/gchpAjUF Run and Deploy LLMs on your iOS or AndroidPhone: https://lnkd.in/grk72Bey

5

📢 Deep Matters: Foundations Speaker Announcement #DeepMatters ⭐️ Ronan Docherty, PhD student at Imperial College London 🎙 Talk: “Make Do and Mend: Leveraging Vision Transformers for Micrograph Segmentation” Ronan is a PhD student at Imperial College London, where he explores how modern machine learning can improve the segmentation of micrographs — a key step in materials analysis. Alongside his research, he works part-time at Polaron as a machine learning engineer, applying cutting-edge AI methods to real-world scientific problems. In his talk, Ronan will discuss how foundation models — particularly vision transformers trained on vast datasets of natural images — can be adapted for materials science. He’ll share strategies for fine-tuning these models to handle limited and out-of-distribution data, capturing intricate microstructural details essential for downstream materials characterization. By “making do and mending” with existing foundation models, Ronan demonstrates how researchers can achieve powerful results even without massive bespoke datasets — bridging the gap between modern computer vision and materials science. 📅 6 November 2025 📍 Google Cloud Startup Hub, London 🔗 https://lnkd.in/eVMsw_xk

6

Meta is investing in the future of sound. 🔊 We’ve launched a new audio lab in the UK’s Ox-Cam corridor—one of Europe’s most dynamic research hubs—to push the boundaries of spatial audio, machine learning, and next-gen sound experiences. This lab will play a key role in building the future of human-computer interaction through intelligent audio. #MetaAI #AudioInnovation #ResearchAndDevelopment

12

If you are interested in a human-centered analysis investigating the acceptability of LLM tools in code translation, take a look at this paper! This work was done in collaboration with Anna Rita Fasolino Andrea Vignali Gabriele Dario De Siano. #LLMs #HumanCenteredAI #codetranslation

18

It was wonderful to chat with Oana Stroe about the recently published MIFA guidelines, which aim to standardise how we share AI-ready datasets for bioimage analysis. But even better than anything we could say ourselves are the testimonies from our users, who shared how the guidelines are already helping their research. Read the interview below!

13

If You Are “Building #AI” in #Health⚕️, This Is the Conversation We Need to Have 🧠 I have been having many conversations lately with organizations and #startup founders who say they are “working on AI.” Everyone wants to talk about AI. Everyone wants to do AI. What I often find missing, however, is a fundamental understanding that AI is nothing without #data - and more importantly, without #gooddata. 📊 As this post by Dr. Smisha Agarwal highlights so clearly, advanced #analytics cannot compensate for weak foundations. Working in the intersection of #health x #technology, I have seen this repeatedly. #Algorithms cannot fix #missingness, #bias, or poorly defined denominators. Just as importantly, we need to think more deeply about what datasets these algorithms were trained on in the first place. 🔍 🔍What blind spots are embedded in those data? 🔍How might those #blindspots shape or distort outputs? 🔍And why does that matter for real-world decisions in #healthcare? 📉 #Predictivemodels are also too often mistaken for #causal insight, and the consequence can be particularly serious in #publichealth and #healthcare contexts. ⚠️ And too rarely do we pause to ask whether simpler analytical approaches might answer the question just as well - or better. 📉➡️📈 This is especially concerning when we consider #gender #datagaps and other structural blind spots in health data. When #women, #marginalizedpopulations, or frontline realities and nuanced #contexts are poorly captured, AI will simply scale those blind spots with confidence. ⚖️ AI can be incredibly powerful - but only when paired with: 📈 strong data systems 🗄️ a deep understanding of the underlying data 📐solid #dataanalytics and #statistical foundations 🧠careful thinking about #causality versus #prediction and #humility about what the #evidence can (and cannot) tell us 🤍 As we move rapidly toward #AI-enabled #healthsolutions, the real work remains unglamorous but essential: building high-quality data, asking the right questions, and choosing methods that serve the problem - not the trend or the technology. 🌱 Thank you Dr. Smisha Agarwal for making these critical points visible and for grounding the AI conversation in #evidence, rigor, and humility. 🙏 #AI #DigitalHealth #DataScience #HealthSystems #PublicHealth #GenderData #EvidenceBased #HealthTech #populationhealth #machinelearning The Center for Global Digital Health Innovation Johns Hopkins Bloomberg School of Public Health #OODH

21

I’m thrilled to share a new paper was recently published in the Alzheimer's Association®'s Journal, 𝘈𝘭𝘻𝘩𝘦𝘪𝘮𝘦𝘳'𝘴 𝘢𝘯𝘥 𝘋𝘦𝘮𝘦𝘯𝘵𝘪𝘢! 🧠 My previous PhD work used machine learning to explore the biology driving #Alzheimer's, this time I used #ML to validate digital tools aimed at making #dementia monitoring more accessible to patients! Collaborating with BU Alzheimer's Disease Research Center and Linus Health, we validated the Defense Automated Neurobehavioral Assessment (DANA) battery, a digital #cognitive test that can be administered to patients in their own homes, on their own smartphones! 𝘏𝘰𝘸 𝘥𝘪𝘨𝘪𝘵𝘢𝘭 𝘢𝘵 𝘩𝘰𝘮𝘦 𝘢𝘴𝘴𝘦𝘴𝘴𝘮𝘦𝘯𝘵𝘴 𝘳𝘦𝘷𝘰𝘭𝘶𝘵𝘪𝘰𝘯𝘪𝘻𝘦 𝘥𝘦𝘮𝘦𝘯𝘵𝘪𝘢 𝘤𝘢𝘳𝘦: 📲 𝐒𝐦𝐚𝐫𝐭𝐩𝐡𝐨𝐧𝐞 𝐓𝐞𝐬𝐭𝐢𝐧𝐠 = 𝐈𝐧𝐜𝐫𝐞𝐚𝐬𝐞𝐝 𝐀𝐜𝐜𝐞𝐬𝐬𝐢𝐛𝐢𝐥𝐢𝐭𝐲 ► Patients take tests at home on their own devices that they are familiar with. ► Removes travel and scheduling barriers, key obstacles for older adults with limited mobility. 📈 𝐓𝐫𝐚𝐜𝐤𝐢𝐧𝐠 𝐆𝐞𝐧𝐮𝐢𝐧𝐞 𝐂𝐨𝐠𝐧𝐢𝐭𝐢𝐯𝐞 𝐂𝐡𝐚𝐧𝐠𝐞𝐬 ► Repeat cognitive testing often causes “practice effects,” where people improve just by familiarity with test material. ► Our research validates that DANA is robust against these effects and can be used to capture true cognitive shifts over time. 🔎 𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐕𝐚𝐥𝐢𝐝𝐚𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐃𝐞𝐭𝐞𝐜𝐭𝐢𝐧𝐠 𝐂𝐨𝐠𝐧𝐢𝐭𝐢𝐯𝐞 𝐈𝐦𝐩𝐚𝐢𝐫𝐦𝐞𝐧𝐭 ► Random forest and logistic regression models trained on DANA responses achieved up to 71% accuracy at classifying cognitively impaired individuals. 🏥 𝐑𝐞𝐚𝐥 𝐖𝐨𝐫𝐥𝐝 𝐈𝐦𝐩𝐚𝐜𝐭 𝐨𝐧 𝐏𝐮𝐛𝐥𝐢𝐜 𝐇𝐞𝐚𝐥𝐭𝐡 ► Globally, there is a growing shortage of dementia specialists, coinciding with a rise in the number of patients living with dementia worldwide. ► Smart phone based digital assessment tools like DANA increase accessibility of dementia monitoring, no in-clinic visit needed! ► Remote, frequent cognitive assessments taken at home put control in the hands of patients and caregivers, enabling early detection and proactive management of brain health. It's a truly exciting time to be in the field of digital health! My PhD research has focused on leveraging cutting edge technology to both enhance our understanding of Alzheimer's disease biology and address real world health gaps in dementia care. I'm excited to keep integrating my expertise in health data science, neuroscience, and global health to improve health outcomes for neurodegenerative disease patients. 🌎 #digitalhealth #healthAI #research #machinelearning

72

6 Comments

🚨 Small number of poisoned samples can seriously compromise LLMs! I just read an eye-opening study by Anthropic, UK AISI, and The Alan Turing Institute titled “A Small Number of Samples Can Poison LLMs of Any Size.” Their core finding is astonishing: with as few as ~250 malicious documents, an attacker can successfully insert a “backdoor” into large language models (ranging from 600M to 13B parameters) — causing them to misbehave on specific trigger phrases — regardless of the total training data size or model scale. This marks a major shift from the earlier belief that an attacker must control a proportion of the dataset. Instead, this research shows that it’s the absolute number of poisoned samples that truly matters. The full technical paper — “Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples” — dives deeper into the experiments and implications. 📄 Read here: https://lnkd.in/gje8V-76 🎥 Watch a great explanation: https://lnkd.in/gMk68vj3 🙏 Thanks to Krish Naik for breaking it down so clearly!

34

2 Comments

Excited to share our new method SCONE (Scalable Contrastive Orientation for causal discovery), developed with my collaborators Khushi Desai and Sopho Kevlishvili , supervised by Elham Azizi! 🔸 SCONE tackles causal discovery under unknown soft interventions, a setting common in real-world systems such as gene regulatory responses to treatment 🔸 Introduces contrastive orientation rules across regimes, enabling recovery of causal edges that cannot be identified from any single regime alone 🔸 Scales causal discovery to large graphs by combining subset-level discovery with a global aggregation mechanism 🔸 Generalizes to out-of-distribution mechanisms, addressing a major limitation of many existing causal discovery methods 💡 Key idea: by comparing structural information across regimes, SCONE can identify edges that non-contrastive approaches fundamentally cannot recover, bridging the gap between theoretical guarantees and scalable practical methods. 📈 Experimentally, SCONE outperforms state-of-the-art methods across graph sizes and densities, including large graphs (100 nodes), even when other methods are given the true intervention targets—which SCONE does not require. 👉 Preprint: https://lnkd.in/e52n32vG 💻 Code: https://lnkd.in/eDpApMkY Huge thanks to my co-authors Khushi Desai, Sopho Kevlishvili and Elham Azizi for making this work possible! #AI #MachineLearning #CausalDiscovery #GraphLearning #Statistics #DataScience

52

2 Comments

2025 was quite the year, but thanks to some down time, I’ve been able to catch up on recent work on AI and data ethics in the context of LMIC systems. I found the RAi UK white paper on Responsible AI in LMICs interesting for how it explores AI adoption in contexts where data systems are still immature and local priorities are overridden by global frameworks. It outlines the reality many of us see in health and social systems on the ground and raises questions about who innovation is really being designed for. UNESCO also updated the Playbook on tackling gender bias and harms in AI. I liked how practical this work is. It shows how the same inputs can still produce biased outputs. In settings where women’s data is poorly captured, these biases amplify inequalities and actively make them worse. Reading these back to back brought into focus the same clear message. Ethical AI is about about better models yes, but more so about data, the systems that produce it, and the decisions that shape its use. Adding AI does not improve decision making if data is incomplete and governance is patchy. It is just the same problems just dressed up as innovation. Curious about what 2026 will bring as this space keeps changing. 🔗 RAi UK White Paper https://lnkd.in/eiKNStaJ 🔗 UNESCO Red Teaming Playbook https://lnkd.in/ekW-M6u7

22

1 Comment