Overhang Tower: Resource-Rational Adaptation in Sequential Physical Planning

Resource-Rational Adaptation in Sequential Physical Planning: A Critical Review of "Overhang Tower"

Introduction

"Overhang Tower: Resource-Rational Adaptation in Sequential Physical Planning" (2604.09072) addresses the computational architecture underlying human sequential physical planning under resource constraints. The paper scrutinizes two long-standing debates within cognitive science and AI: (i) whether physical prediction is underpinned by simulation-based (Intuitive Physics Engine, IPE) mechanisms or fast cue-based visual heuristics (CNNs), and (ii) whether sequential planning employs deliberative lookahead or myopic, single-step strategies. The authors provide a unifying empirical framework by analyzing how both mechanisms adapt dynamically to cognitive resource limitations using a novel block-stacking task, Overhang Tower.

Experimental Design and Framework

The Overhang Tower task constitutes a sequential, interactive physical planning paradigm in which participants place blocks on a 2D grid to maximize the final overhang while maintaining dynamic stability. Critically, participants must engage in temporally extended planning, since greedily maximizing immediate overhang rapidly exhausts the "stability budget" due to physical constraints, necessitating the use of counter-weighting and causal ordering to achieve optimal results.

Participants were randomly assigned to either a time-constrained (5-second per placement) or unconstrained condition. Detailed behavioral traces, including action sequences and mouse movements, allowed the reconstruction of implicit planning search strategies.

To model human performance, the authors implement two independent factors: (i) the internal physics prediction module (either IPE or a visual-heuristic CNN), and (ii) the planning module (either myopic or deliberative forward-search of variable depth). The IPE is implemented via Monte Carlo simulation (with Gaussian perturbations), while the visual-heuristic model is a trained Inception-V4 CNN classifier for stability using photorealistic renderings. The planning module ranges from single-step greedy selection to multi-step lookahead (D-step forward search).

Empirical Results

Dual Transition in Physical Prediction and Planning

Key results reveal a dual adaptation in both the physical prediction mechanism and planning strategy, constituting strong empirical support for a resource-rational, hierarchical cognitive architecture. Specifically:

• Physical Prediction Adaptation: In early construction stages with low scene complexity, participant actions are best explained by simulations of physical dynamics (IPE). As complexity increases with additional blocks, the predictive advantage of the IPE degrades, and the visual-heuristic model (CNN) becomes a superior fit to human decisions. This mechanistic crossover reflects the compounding uncertainty of simulations with depth, supporting that humans adaptively switch to cheaper, more robust heuristics when simulation cost outweighs benefit. Quantitatively, the CNN exceeds IPE in log-likelihood fit to human actions as construction progresses (significant at $p < 0.001$).
• Planning Strategy Adaptation: Under time constraints, participants predominantly employ myopic, vertically conservative strategies, whereas unconstrained participants are more likely to engage in non-greedy, laterally extended, high-reward configurations—often requiring precise order-dependent placements (counter-weighting and vertical anchoring). This is formalized by Order Dependency $\Gamma$, a metric quantifying the proportion of valid block-order permutations compatible with a stable structure. Higher $\Gamma$ (more path-dependent solutions) appeared in unconstrained trials ($p = 0.022$), directly implicating lookahead-based planning.
• Performance Benchmarks: Fully myopic models produced significantly lower rewards than human participants, while models incorporating two or three-step lookahead closely matched human performance in time-constrained and unconstrained settings, respectively.

Dissociation and Coupling

The results establish a dissociation of prediction and planning processes while also demonstrating that both must adapt jointly for robust sequential physical reasoning. Myopic planning, regardless of prediction fidelity, fails to realize high-reward but path-dependent solutions; deep lookahead without robust prediction is likewise inadequate once complexity induces simulation noise or intractability.

Theoretical Implications

This work provides a strong evidence base for a hierarchical, resource-rational cognitive architecture in physical planning. The ability to dynamically shift between simulation and heuristic prediction as a function of complexity advances prior accounts that treated physical prediction and planning strategy as isolated modules. The fact that the cognitive system deploys expensive mechanisms (IPE-style simulation, deep lookahead) selectively, optimizing for expected gain under task demands and computational costs, substantiates resource-rationality as a critical design principle for both biological and artificial agents.

The formalization of Order Dependency as a diagnostic metric for lookahead highlights the importance of precise action sequencing in physical construction and offers a tool for analyzing other multi-step planning domains.

Implications for Artificial Intelligence and Future Directions

Several implications arise for AI and cognitive modeling:

1. Hybrid Architectures: The evidence supports the design of AI systems that flexibly arbitrate between simulation and learned heuristics, optimizing computational expenditure based on complexity and resource constraints.
2. Model-based RL and Physical Reasoning: Autonomous agents may benefit from adaptive planning horizons and multi-fidelity prediction mechanisms, especially in settings involving extended temporal dependencies and under noisy dynamics.
3. Systematic Training and Transfer: Future work could generalize these findings to dynamic or stochastic physical environments, evaluate learning curves with increasing task exposure, and investigate how meta-learning or curriculum design can endow agents with efficient resource allocation policies.
4. Dissociating Error Sources: The decomposition of failures into prediction and planning-induced errors could inform diagnostics and targeted improvements in embodied AI agents, as well as new behavioral paradigms for human studies.
5. Broader Applicability: The paradigm and modeling framework may be extensible to domains outside physical reasoning, wherever sequential decision-making under uncertainty and limited computation is paramount (e.g., strategic games, robotics, or complex manipulation tasks).

Conclusion

The study delivers a rigorous demonstration of a hierarchical, resource-rational adaptation mechanism in human sequential physical planning. By experimentally dissociating and modeling both physical prediction fidelity and planning horizon, the authors show that humans flexibly trade computational cost against predictive reliability in a manner tightly coupled to cognitive budget and task demands. These findings have substantive implications for the development of adaptive, efficient AI agents and for formal models of human cognition that aspire to ecological validity and computational rationality.