ADAPT at Trinity College Dublin and IBM researcher, Yannis Belkhiter, presented two papers at a workshop hosted by the International Conference on Machine Learning (ICML 2026) in Seoul recently, highlighting advances in the efficiency and interpretability of large reasoning models.
The first paper, Step-Tagging Early-Stopping: Toward Controlling the Generation of Language Reasoning Models Through Black-Box Step Monitoring, introduced a black-box approach that reduces token generation by 20–50% with minimal impact on accuracy by monitoring reasoning steps and stopping generation when appropriate. This paper was partially supported by the ADAPT Research Ireland Centre with ADAPT Director Prof. John Kelleher as one of the co-authors.
The second paper, Local Causal Attribution of Chain-of-Thought Reasoning, proposed a black-box method for tracing how individual reasoning steps influence subsequent steps using a structural causal model.
The research was carried out in collaboration with Seshu Tirupathi, Giulio Zizzo, Prof. John D. Kelleher (ADAPT Director), Dennis Wei, Erik Miehling, and Radu Marinescu, and was supported by 6G-XCEL and the ADAPT Centre.
