(a) Neuronal object values provide a suitable foundation for local choice probability and global matching behaviour. Left: Average value-related activities for object A (red) and B (green) tracked local choice probability (bars), calculated for discrete 16 value bins (d.f.=14). Right: Local choice probabilities, binned according to object value and aggregated over trial blocks, were related to experienced reward ratios, consistent with matching behaviour. (b) Neuronal sensitivity to object value is related to behavioural valuation accuracy. Unsigned neuronal regression coefficients for value coding plotted against behavioural ‘valuation accuracy’ (defined as session-specific correlation between estimated values and true reward probabilities). Stronger neuronal value coding predicted more accurate behavioural valuation. Valuation accuracy was also related to better choices (that is, choosing the object with higher true reward probability, R=0.178, P=0.0037, linear regression). (c) Relationship between neuronal object values and performance. The strength of neuronal value coding dropped transiently when the animals would commit an error and subsequently increased on the next correct trial. Linear regressions of normalized population activity on object value, calculated across 273 value-coding responses (18,117 trials), separately for pre-error, error and post-error trials. *P<0.05, t-test for dependent samples. Data in all plots are taken across all task periods.