The Claw Machine Task: Modelling Confidence in Tool Use
- MetaTool
- Aug 13
- 1 min read
METATOOL researcher Polina Arbuzova (Humboldt-Universität zu Berlin) took part in the 2026 Cognitive Computational Neuroscience (CCN) conference at New York University.
On August 5, 2026 METATOOL researcher Polina Arbuzova (Humboldt-Universität zu Berlin) took part in the 2026 Cognitive Computational Neuroscience (CCN) conference at New York University , where she presented the poster “The Claw Machine Task: Modelling Confidence in Tool Use”, co-authored with Paul F. Kinghorn and Poppy Collis (University of Sussex) and Verena V.
Hafner (Humboldt-Universität zu Berlin).
CCN is an annual international meeting that brings together researchers across neuroscience, cognitive science, and artificial intelligence to advance computationally grounded theories of mind and brain. Alongside the main conference, she also attended the “Computational Consciousness Science” and “Metacognitive Science” satellite meetings.”
Advanced tool use requires evaluating a range of uncertainties. We introduce the Claw Machine experimental paradigm, a novel tool-use task in which reward structure and outcome uncertainty are manipulated independently, eliciting separate confidence ratings for action selection and physical execution. Data are reported from a preliminary sample of 36 out of 40 planned participants. A Bayesian negative entropy model, parameterised by inverse temperature β indexing the degree to which reward influences confidence, captures this dissociation: action selection confidence tracks expected utility and varies with reward structure, while execution confidence reflects object properties and outcome variability independently of reward. Model-fitted β did not correlate significantly with independent risk-taking (BART), suggesting reward sensitivity in structured tasks may be dissociable from general risk tolerance under ambiguity.
Interested in the full article?
Comments