arXiv:2605.30273v1 Announce Type: cross 
Abstract: Large language models (LLMs) show promise in generating supportive responses for mental health queries, but improving their usefulness, empathy, and safety often requires substantial compute, expert input, and labeled data. At the same time, deploying proprietary, cloud-based models for mental health-related interactions raises important privacy and data-governance concerns, given the sensitivities. To address this challenge, we introduce LLUMI setup that can be hosted in-house within protected environments. LLUMI consists of two complementary components: a generation model (GM), which drafts supportive responses to mental health queries, and an improvement model (IM), which revises an initial human-crafted response. We leverage feedback signals from Reddit mental health communities, using community endorsement patterns such as upvotes and downvotes to construct chosen-rejected response pairs for Supervised Fine Tuning (SFT) and Direct Preference Optimization (DPO). We further align LLUMI using human evaluation across five dimensions: readability, empathy, connection, actionability, and safety. Our results show that, despite relying on smaller open-source models rather than proprietary cloud-based GPT models, LLUMI achieves comparable performance across linguistic analyses and human evaluations. These findings suggest that open-source models, when trained with community-derived preference signals, can support high-quality mental health support assistance while offering a more privacy-preserving alternative for sensitive support contexts.

تم تقديم LLUMI كنظام جديد مصمم لتحسين قدرات المساعدة في الكتابة لنماذج اللغة الكبيرة (LLMs) لدعم الصحة النفسية، باستخدام تعليقات من مجتمعات Reddit لتحسين جودة الردود. يتكون النظام من نموذج توليد يقوم بإنشاء ردود داعمة ونموذج تحسين يقوم بتنقيح الردود البشرية الأولية.

LLUMI se ha presentado como un nuevo sistema diseñado para mejorar las capacidades de asistencia de escritura de los modelos de lenguaje grandes (LLMs) para el apoyo en salud mental, utilizando comentarios de comunidades de Reddit para mejorar la calidad de las respuestas. El sistema consta de un modelo de generación que crea respuestas de apoyo y un modelo de mejora que refina las respuestas humanas iniciales.

LLUMI a été introduit comme un nouveau dispositif conçu pour améliorer les capacités d'assistance à l'écriture des grands modèles de langage (LLM) pour le soutien en santé mentale, en utilisant les retours des communautés Reddit pour améliorer la qualité des réponses. Le système comprend un modèle de génération qui crée des réponses de soutien et un modèle d'amélioration qui affine les réponses humaines initiales.

LLUMI has been introduced as a novel setup designed to enhance the writing assistance capabilities of large language models (LLMs) for mental health support, utilizing feedback from Reddit communities to improve response quality. The system comprises a generation model that creates supportive replies and an improvement model that refines initial human responses.

LLUMI: Improving LLM Writing Assistance for Mental Health Support with Online Community Feedback

"We show that a tiny snippet—just 13 words—of retrieved text on a UGC website like Reddit, Wikipedia, Quora, or Facebook can change AI agents to output spam / scam content pretty consistently."

تشير الأبحاث إلى أن عبارة بسيطة تتكون من 13 كلمة من مواقع المحتوى الذي ينشئه المستخدمون مثل Reddit وWikipedia وQuora وFacebook يمكن أن تؤثر بشكل كبير على أنظمة الذكاء الاصطناعي، مما يؤدي بها إلى إنتاج محتوى مزعج أو احتيالي بشكل متكرر. تسلط هذه النتيجة الضوء على ضعف الذكاء الاصطناعي أمام التلاعب من خلال مقتطفات نصية تبدو غير ضارة.

La investigación indica que un simple fragmento de 13 palabras de sitios web de contenido generado por usuarios como Reddit, Wikipedia, Quora o Facebook puede influir significativamente en los sistemas de IA, llevándolos a producir contenido de spam o estafa con una alarmante consistencia. Este hallazgo destaca la vulnerabilidad de la IA a la manipulación a través de fragmentos de texto aparentemente inocuos.

Des recherches montrent qu'un simple extrait de 13 mots provenant de sites de contenu généré par les utilisateurs tels que Reddit, Wikipedia, Quora ou Facebook peut influencer considérablement les systèmes d'IA, les amenant à produire du contenu indésirable ou frauduleux de manière alarmante. Cette découverte souligne la vulnérabilité de l'IA à la manipulation à travers des extraits de texte apparemment innocents.

Research indicates that a mere 13 words from user-generated content websites such as Reddit, Wikipedia, Quora, or Facebook can significantly influence AI systems, leading them to produce spam or scam content with alarming consistency. This finding highlights the vulnerability of AI to manipulation through seemingly innocuous snippets of text.

It Is Trivially Easy to Use Reddit to Manipulate AI Search, Research Suggests

arXiv:2606.13944v1 Announce Type: new 
Abstract: Large language models (LLMs) are increasingly characterised in recent evaluation work as having stable, model-level preference and value systems. However, accompanying robustness checks are limited to incidental prompt perturbations such as syntax variation and option reordering. This leaves open whether the measured properties survive when the surrounding task context changes, as it does in most real deployments. We test this directly across two established pairwise paradigms: ranking country preferences and eliciting utility judgements. In both, we make the deployment context -- the high-level task the model is performing while making concrete value-dependent choices -- our controlled variable, varied across framings such as writing a Reddit post or a news article. Across five LLMs and over 1.2M pairwise decisions, deployment context produces variation far larger than prompt paraphrasing and temperature controls. In country preference rankings over 15 countries, context induces widespread, statistically significant rank shifts; the aggregate Global North favouritism reported in prior work is itself context-dependent, with each model's bias shifting systematically across contexts. In utility elicitation over 50 outcomes, broad cross-category ordering is preserved, but fine-grained rankings within domains vary substantially, and cardinal exchange rates between outcomes (e.g. how many lives in one region equal one in another) shift by a factor of 2.47 at the median. Reported model-level preferences and utilities are therefore better understood as context-conditioned measurements than fixed model-level properties: safety guarantees obtained under one framing provide limited assurance in another.

تشير الأبحاث الحديثة إلى أن نماذج اللغة الكبيرة (LLMs) تظهر تباينات كبيرة في تفضيلاتها وقيمها بناءً على سياق النشر، كما يتضح من التجارب التي تشمل تصنيف تفضيلات الدول وإصدار أحكام المنفعة عبر أطر مهام مختلفة.

Investigaciones recientes indican que los modelos de lenguaje de gran tamaño (LLMs) exhiben variaciones significativas en sus preferencias y valores según el contexto de implementación, como se demuestra a través de experimentos que involucran la clasificación de preferencias de países y juicios de utilidad en diferentes marcos de tareas.

Des recherches récentes indiquent que les modèles de langage de grande taille (LLMs) présentent des variations significatives dans leurs préférences et valeurs en fonction du contexte de déploiement, comme le montre des expériences impliquant le classement des préférences des pays et des jugements d'utilité à travers différents cadres de tâches.

Recent research indicates that large language models (LLMs) exhibit significant variations in their preferences and values based on the deployment context, as demonstrated through experiments involving ranking country preferences and utility judgments across different task framings.

LLUMI: Improving LLM Writing Assistance for Mental Health Support with Online Community Feedback

Was this article worth reading? Share it

Ready to build your own newsroom?