This is highly speculative & requires formalizing, but my hope is that the framework promotes these investigations! 🤞 Something I'm so excited to work on with others including new students ☺️

Assuming a very simple rule for planning termination, that we need a plan with an expected return above some threshold, the construal itself can make that threshold much harder to reach. Thus, search persists and is negative (ie worry).

One may not be risk averse yet select goals in ways that predict worry. E.g., an individual w low self-esteem may desire to "sound confident" & plan for an interview by simulating speech production - response transitions others would ignore. this construal may impact how we find a good-enough policy

I've reflected more on how I could offer new models using the framework and how to fit for instance Chris's work into the framework. Chris's model's risk aversion param he identifies is reflective of personality (top level). Through CVaR, it quite directly distorts task structure & planning

Chris and Sam's work is great (Chris and I were in a joint RLDM symposium). The goal in this very short review is to spotlight more recent planning models that can be applied for anxiety model development.

The proposal synthesizes seminal work on planning models from @markkho.bsky.social@fredcallaway.bsky.social@katenuss.bsky.social 's framework on meta-learning to study how self-world construals develop. Would love your feedback!

Thanks Ross -- will check it out!

Thanks, Sam. I recall reading your paper and wondering if forgetting of structure was a confound. After revisiting, forgetting cannot accommodate 50-trial load having greater revaluation mag than 30-trial load (& possibly rest effects). Did you ever probe for memory of structure post phase-4?

Also, I don't believe their model has been tested in human participants. I suppose it's harder to test empirically questions about offline model-updating via a DYNA-like process.

Thanks, Fred. We found Chelu et al & Namboodiri's on PR serendipitously after our ideas coalesced. Great suggestion to reconcile our findings! One other point is that they focus on accuracy (RMSE) whereas we on efficiency (representational load). Many questions abound (e.g., how is PR learnt)!