Uptown Chunk You Up: Skill Automatization as Emergent Temporal Chunking from Adaptive Credit Assignment

Abstract

Human skill acquisition exhibits a transition from slow, deliberative behavior to fast, automatic performance, often mediated by the formation and consolidation of “chunks” that compress sequences of actions into cohesive units. While temporal-difference methods such as TD(λ) provide a principled mechanism for distributing credit across time, existing approaches rely on fixed or heuristic choices of λ, limiting their ability to adapt to task structure. In this work, we propose a cognitively grounded framework in which temporal credit assignment is dynamically modulated as a function of decision complexity. We introduce the λ-corridor hypothesis, which proposes that environments with extended low-conflict regions (corridors) favor long-horizon credit propagation and chunk formation, while high-conflict decision points (junctions) require localized credit assignment. Building on this insight, we develop an agent architecture that integrates a fast habitual policy with a slow memory system, arbitrated by a meta-controller network. This network modulates λ online, enabling the agent to interpolate between local and temporally extended learning regimes. Empirical results across a range of graph structures show that the ideal credit assignment horizon depends strongly on local environment structure. Allowing λ to adapt dynamically leads to more efficient learning and more reliable decisions. These results suggest that chunking is not a distinct mechanism, but instead emerges naturally from how agents generalize across time. In this view, our framework extends Shepard’s universal law of generalization beyond static psychological space to also include temporal structure.