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

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

LLMs Lack Dissociable Planning Failures Observable in Humans and Deep RL on Tower of Hanoi

Austin Tudor David Andrews¹, Jascha Achterberg¹, Rui Ponte Costa¹; ¹University of Oxford

Presenter: Austin Tudor David Andrews

Planning is thought to rely on coordinated, brain-wide computations that generate and execute action sequences under environmental constraints (Goel & Grafman, 1995; Grafman et al., 1992; Schmahmann & Sherman, 1997). Experimental studies have identified key contributions from regions such as the prefrontal cortex (PFC) and cerebellum. We propose that modular deep reinforcement learning (RL) agents capable of solving human-level planning tasks can provide a framework for understanding how different brain regions contribute to planning. We employ MuZero and the Tower of Hanoi (ToH) task to investigate the roles of distinct neural systems in goal-directed planning and behaviour. By performing targeted network ablations, we show that ablation of the MuZero value network reproduces behavioural patterns observed in patients with PFC damage, whereas ablation of the MuZero policy network mirrors deficits associated with cerebellar atrophy (CA). In contrast, we show that LLMs capture PFC deficits, but fail to capture cerebellar-like deficits.

Topic Area: Decision-Making, Cognitive Control & Event Cognition