Microsoft Research´s latest update spotlights its significant contributions to the upcoming CHI 2025 and ICLR 2025 conferences, highlighting advancements across causal reasoning with large language models (LLMs), tools for augmenting human cognition, and robust language model security. At CHI 2025 in Yokohama, researchers will present over 30 sessions, including papers and workshops on how Artificial Intelligence enables ´tools for thought´—novel systems designed to enhance human knowledge work. The Tools for Thought initiative demonstrates prototype systems aimed at supporting diverse cognitive tasks, and invites the community to help define the evolving interplay between human reasoning and generative Artificial Intelligence.
Multiple new research works are introduced, such as a comprehensive study selected for ICLR 2025 that explores how LLMs generate and validate causal arguments. This foundational research proposes frameworks for leveraging LLMs in impactful domains like medicine, science, and policy by bridging common sense and domain knowledge with formal causal methodologies. In the area of language model alignment and safety, Microsoft presents ADV-LLM, a novel iterative self-tuning approach for generating adversarial suffixes to stress-test LLM security. ADV-LLM achieves near-perfect attack success rates on open-source and even closed-source models (such as GPT-3.5 and GPT-4), supporting future alignment research and safer LLM deployment by facilitating large-scale safety dataset generation.
The ChatBench study re-examines standard LLM benchmarks, contrasting ´AI-alone´ performance with human-AI collaborations. Analyzing more than 144,000 answers and thousands of user-LLM interactions across multiple domains, the research shows that accuracy in automated settings does not guarantee improved outcomes in real-world, collaborative contexts. This insight is operationalized via user simulators for scalable, interactive evaluation of future Artificial Intelligence agents. On the technical frontier, Distill-MOS reveals a highly compressed speech quality assessment model—over 100 times smaller than leading models—enabling non-intrusive, real-time evaluation in resource-limited environments.
Healthcare innovation also features prominently. On the NEJM Catalyst podcast, Microsoft Health SVP Jim Weinstein and Intermountain Health´s Dan Liljenquist discuss efforts to address rural US healthcare challenges through advanced technology, telemedicine, and cybersecurity. Related podcasts further explore how Artificial Intelligence empowers patients and transforms digital health business models, as well as the growing impact of pre-trained models and ambient clinical intelligence on biomedical research and healthcare delivery. Through these initiatives, Microsoft Research demonstrates its ongoing commitment to advancing the societal and practical impact of Artificial Intelligence across technical, human, and healthcare domains.
