Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Keep it in Sight: Reward Function Compression Through Goal Visibility

Xitong Chen¹, Chaofan Li², Gaia Molinaro¹, Anne Collins¹; ¹University of California, Berkeley, ²Ohio State University, Columbus

Presenter: Xitong Chen

Humans often learn from abstract goals, but less efficiently than from explicit rewards. The "reward function compression" hypothesis proposes that this inefficiency arises because learners gradually acquire a compressed goal evaluation function. Alternatively, this cost could simply arise from an immediate dual-task effect, where goal maintenance imposes a cost on decisions. To arbitrate between these hypotheses, we extended a goal-based learning protocol to offload goal or non-goal information to a visible reminder. Our results (N = 61) support the reward compression hypothesis. Visible goals gradually reduced learning efficiency costs, eventually approaching the efficiency of point-based learning performance. An equivalent memory-load reduction (visible non-goals) did not boost learning as much, highlighting the role of outcome evaluation in goal-directed learning. Computational modeling revealed that outcome evaluation efficiency, proxied by participants' "reward collection" response time, specifically influenced participants' learning processes, rather than decision noise. These results show that learning from goals depends on our ability to train efficient estimators of outcome values, which is facilitated by keeping goals in sight.

Topic Area: Memory, Learning & Knowledge Structures