LLM Safety | Yiqiao Jin CS PhD @ Georgia Techhttps://ahren09.github.io/tags/llm-safety/LLM SafetyHugo Blox Builder (https://hugoblox.com)en-usWed, 01 Jul 2026 00:00:00 +0000https://ahren09.github.io/media/icon_hu_eee6347cbdb2cc3f.pngLLM Safetyhttps://ahren09.github.io/tags/llm-safety/Reasoning Is Not All You Need: Examining LLMs for Multi-Turn Mental Health Conversationshttps://ahren09.github.io/publication/acl26_mental_health/Wed, 01 Jul 2026 00:00:00 +0000https://ahren09.github.io/publication/acl26_mental_health/<h2 id="abstract">Abstract</h2> <p>Mental health support over multi-turn conversations stresses LLM reasoning, empathy, and safety in distinct ways. We systematically examine the limits of reasoning-focused LLMs in multi-turn mental health conversations, isolating failure modes that pure reasoning cannot address.</p> <h2 id="links">Links</h2> <ul> <li> </li> </ul>MedHalu: Hallucinations in Responses to Healthcare Queries by Large Language Modelshttps://ahren09.github.io/publication/icwsm26_medhalu/Mon, 01 Jun 2026 00:00:00 +0000https://ahren09.github.io/publication/icwsm26_medhalu/<h2 id="abstract">Abstract</h2> <p>Large language models are increasingly used for consumer healthcare queries, but their responses can contain subtle hallucinations with serious implications for patient safety. We introduce MedHalu, a benchmark for studying hallucinations in LLM responses to healthcare queries, with fine-grained annotations of hallucination types and severity.</p> <h2 id="links">Links</h2> <ul> <li> </li> </ul>Sysformer: Safeguarding Frozen Large Language Models with Adaptive System Promptshttps://ahren09.github.io/publication/iclr26_sysformer/Thu, 23 Apr 2026 00:00:00 +0000https://ahren09.github.io/publication/iclr26_sysformer/<h2 id="abstract">Abstract</h2> <p>Aligning frozen large language models without modifying their weights is a key challenge for safe and adaptive deployment. We introduce Sysformer, a system that learns adaptive system prompts to safeguard frozen LLMs across diverse risk scenarios. Sysformer treats the system prompt as a learnable, query-conditioned intervention, enabling fine-grained safety control without parameter updates and improving robustness across multiple LLM families.</p> <h2 id="links">Links</h2> <ul> <li> </li> <li> </li> </ul>UniGuard: Towards Universal Safety Guardrails for Jailbreak Attacks on Multimodal Large Language Modelshttps://ahren09.github.io/publication/aaai25_uniguard/Mon, 03 Mar 2025 00:00:00 +0000https://ahren09.github.io/publication/aaai25_uniguard/<h2 id="abstract">Abstract</h2> <p>Multimodal LLMs are vulnerable to jailbreak attacks that exploit cross-modal interactions. We introduce UniGuard, a universal safety guardrail framework that defends multimodal LLMs against jailbreak attacks across image and text channels.</p> <h2 id="links">Links</h2> <ul> <li> </li> </ul>PrivacyMind: Large Language Models Can Be Contextual Privacy Protection Learnershttps://ahren09.github.io/publication/emnlp24_privacymind/Tue, 12 Nov 2024 00:00:00 +0000https://ahren09.github.io/publication/emnlp24_privacymind/<h2 id="abstract">Abstract</h2> <p>Deploying LLMs on private text data raises serious contextual privacy concerns. We introduce PrivacyMind, which teaches LLMs to be contextual privacy protection learners — recognizing sensitive content in context and adapting their outputs accordingly.</p> <h2 id="links">Links</h2> <ul> <li> </li> </ul>