Hendler, O.; Mondal, k.; Dushnik Shamir, N.; Volotsky, S.; Shamir, M.; Segev, R.

To quantify animals' perceptual capabilities, studies typically assess behavioral accuracy, the proportion of correct choices accumulated over trials on a given task. However, recent works on humans and rodents have shown that task decisions exhibit dynamic shifts from trial to trial, thus casting doubt on the reliability of behavioral accuracy as a true measure of capabilities. Here, these decision-making dynamics were tested on archerfish, an animal lacking a fully developed cortex, whose behavioral decisions are easy to read out. We conducted a series of experiments involving a two-alternative choice task where a target and a non-target shape were randomly placed in two possible positions. Fitting dynamic generalized linear models to each fish's binary choice data revealed that target position strongly affected accuracy and that this effect fluctuated over a timescale of ~100 trials. The archerfish often repeated their choices regardless of the reward: they frequently selected one target or non-target shape on numerous consecutive trials, which is suggestive of high object recognition capacity. Importantly, the findings indicated that similar latent decision variables underlying mammalian decision-making, such as choice history, were also operational in the archerfish. Then, to investigate behavioral accuracy in more detail, we introduced unrewarded probe trials. Unlike the findings reported in rodents, archerfish performance remained stable during these unrewarded trials. Finally, a decision-making paradigm with stimuli at multiple locations yielded results that were consistent with the simpler task variant. More generally, these findings suggest that an animal's decision-making dynamics can mask its true perceptual capabilities when performing an object recognition task, with broad implications for the ways in which behavioral assays are designed and animal performance is interpreted across taxa.