Wahooo!!!

This was a team effort!!

Massive credit to @banerjeesounak.bsky.social for leading this work and a big thanks to @eugenevinitsky.bsky.social, @daphne-cornelisse.bsky.social, and our partners at @toyota-tri.bsky.social for the fantastic collaboration 🙌🎉

One reason I’m excited about this work is that it is step towards scaling cognitive models to naturalistic domains, like driving 🚗

This pushes cognitive modeling approaches forward while helping bridge the gap between cognitive science theory and real-world autonomous systems!

We pose the problem of *attention-aware inverse planning* and compare to standard inverse RL, which assumes agents have unlimited attention

Further, we examine AAIP in tabular problems and with simulated agents placed in real-world scenarios selected from the Waymo Open Dataset!

In short, attentional biases -> suboptimal behavior

Our new paper inverts this process and asks: Can we infer the biases of attention-limited agents from their observed behavior?

Previous work also examined how *attentional biases* in construing tasks can produce *behavioral suboptimalities*

For example, learned biases can cause us to overlook obvious shortcuts that would make tasks much easier to solve, aka functional fixedness🔨🧩

journals.sagepub.com/doi/full/10....

We build on previous work on *resource-rational planning* and *value-guided construal*

The key idea? People are cognitively limited, so they plan by forming simplified (but useful!) world models 🌍➡️🗺️

This captures how attention-limited humans think/act!

www.nature.com/articles/s41...