https://gaoithee.github.io/saracandussio.github.io/tags/interpretability/InterpretabilityHugo Blox Builder (https://hugoblox.com)en-usFri, 24 Apr 2026 12:05:00 +0000https://gaoithee.github.io/saracandussio.github.io/media/icon_hu7729264130191091259.pnghttps://gaoithee.github.io/saracandussio.github.io/tags/interpretability/https://gaoithee.github.io/saracandussio.github.io/event/clcg-linguistics-lunch/Fri, 24 Apr 2026 12:05:00 +0000https://gaoithee.github.io/saracandussio.github.io/event/clcg-linguistics-lunch/<p>Invited talk at the <a href="https://www.rug.nl/research/clcg/colloquia_discussiongroups_linguisticevents/linguistics_lunch/" target="_blank" rel="noopener">CLCG Linguistics Lunch</a> at the University of Groningen.</p> <p>Chain-of-Thought prompting asks models to reason step by step — but most of what they write is filler. This talk presents work on identifying the semantic minima of AI reasoning: the tiny subset of tokens that actually carry predictive weight, detectable in real time from the model&rsquo;s internal states. Erasing up to 95% of the output leaves a sparse set of words that still perfectly predicts the correct answer.</p>