20/ Contextual differences impacted specific model parameters, such as perceived volatility and beliefs about rule transitions (e.g. the probability of partial VS complete rule changes). Participants adjusted their latent belief about the structure of the task with no explicit incentive to do so!

18/ Last but not least, we explored how participants’ adaptation varied across different contexts. For partial rule changes, performance varied depending on the task context, i.e. the presence of global rule changes during the session or not.

17/ The SI-Struct model, a variant of the strategy inference model, outperformed other models, capturing both the learning curves and interference effects most accurately. Interestingly, this model explicitly incorporates high-level beliefs about the structure of the rule changes (aka the context).

16/ Only strategy inference models accounted for all observed behavioral patterns, and matched participants’ performance. These models use Bayes' rule to track strategy reliability and incorporate long-term priors, supporting a more flexible, contextual and memory-driven approach to adaptation.

15/ Now to the generative models: to investigate the origins of these effects, we compared three families of models: (1) adaptive Q-learning (incremental learning), (2) Probe-based hybrid models (representing strategies built on incremental learning), and (3) pure strategy inference models.

13/ The HMM also revealed bouts of behavior that were attributed to a random strategy, but that were still biased against the previous correct response and followed by near-plateau correct performance.

12/ But first we wanted to investigate strategy changes in participants’ behavior without making any mechanistic assumptions. Using a hidden Markov Model to break down participant's behavior, we show that the interference effect is entirely due to unwarranted changes in the strategy as a whole!

10/ We also observed interesting effects during partial rule changes, where some associations remained stable. Participants showed a transient drop in the stable association’s performance, again suggesting a representation of global strategies, and not aggregates of local stim-action associations.

9/ The impact of recurrent rules was distinct from just the frequency of stimulus-action associations. When faced with new combinations of stimuli, the observed behavioral patterns were identical, supporting the idea of strategy-based memory rather than incremental learning.

8/ This was evident by quicker responses and less exploratory behavior, a hallmark of strategic memory retrieval (cf Collins & Koechlin 2012).

5/ Key to our design: we manipulated the environment’s structure by introducing hidden "contexts" — sets of rules that changed periodically. Some contexts involved complete rule changes (all the associations changed), while others kept parts of the rules stable.

4/ To test this, we designed an experiment where participants had to choose the preferred fruit for 3 different monkeys with changing preferences. Feedback was unreliable, requiring participants to infer underlying rules through trial and error.

3/ We hypothesized that humans may rely on direct inference over strategies, quickly shifting between abstract strategies to adapt to changes.