The best model that best explains confidence serial dependence is the model that 1) systematically jitters criteria based on preceding confidence, 2) uses only explicit confidence responses., and 3) reaches as far as two trials back (wish we could test this further...) for information.

We built and compared several computational models with different configurations for the criterion-updating mechanisms, e.g., we asked whether decision criteria, based on preceding information, are dynamically updated trial to trial, or systematically jittered on every trial from a fixed template.

Interestingly, this serial dependence requires a continuity of confidence responses: without that in the preceding trial, the serial dependence disappeared, even when observers were prompted with "easy/hard" preceding difficulty or were "yoked" to give the high/low ratings in preceding trials.

We replicated the well-known "confidence leak" @dobyrahnev.bsky.social found: confidence at target trial was higher if it was preceded by high-confidence (easy) trials, and lower if by low-confidence (hard) ones, despite matched task performance. It's more than just a leak; it's serial dependence.

In 3 experiments, we varied task difficulty over a short sequences of trials (e.g., hard->hard->[medium], easy->[medium]) to prompt observers to give a specific pattern of confidence ratings before the [target trial] at constant difficulty level and, thus, matched task performance.