An Inference-Based Architecture for Intent and Affordance Saturation in Decision-Making

TL;DR

This paper tackles decision paralysis by formalizing a two-stage planning framework that separately handles intent (what outcome to pursue) and affordance (how to pursue it). Decisions are generated by probabilistic inference under a mixture of Forward KL and Reverse KL divergences, controlled by a parameter $\lambda$, and implemented as both static and dynamic drift–diffusion processes. The key insight is that Forward KL (mode-covering) can preserve multiple plausible goals or actions, leading to slow or halted decisions when options are equally valued, a pattern observed in Autism Spectrum Disorder (ASD); autism is framed as an extreme regime along a general inference continuum. The model reproduces inertia and shutdown in RGB-token and CGT tasks and connects to Bayesian/predictive-processing theories as an orthogonal optimization geometry, offering testable predictions and potential neural correlates, with implications for understanding decision-making diversity beyond ASD.

Abstract

Decision paralysis, i.e. hesitation, freezing, or failure to act despite full knowledge and motivation, poses a challenge for choice models that assume options are already specified and readily comparable. Drawing on qualitative reports in autism research that are especially salient, we propose a computational account in which paralysis arises from convergence failure in a hierarchical decision process. We separate intent selection (what to pursue) from affordance selection (how to pursue the goal) and formalize commitment as inference under a mixture of reverse- and forward-Kullback-Leibler (KL) objectives. Reverse KL is mode-seeking and promotes rapid commitment, whereas forward KL is mode-covering and preserves multiple plausible goals or actions. In static and dynamic (drift-diffusion) models, forward-KL-biased inference yields slow, heavy-tailed response times and two distinct failure modes, intent saturation and affordance saturation, when values are similar. Simulations in multi-option tasks reproduce key features of decision inertia and shutdown, treating autism as an extreme regime of a general, inference-based, decision-making continuum.

Figures (14)

• Figure 1: Overview of the proposed computational model of intent and affordance saturation.
• Figure 2: One-step decision task (RGB token world).
• Figure 3: Cambridge Gamble Task (CGT).
• Figure 5: Violin plots of decision times for intent, action, and total decision time across all cases. The FKL regime ($\lambda=1$) exhibits consistently longer latencies, especially when multiple valuable options are available (Cases 2-4).
• Figure 6: Selection proportions for intents and actions across 1000 drift-diffusion simulations. Intent selections match the optimal intent distribution, and action selection follows the softmax policy induced by the KL-based decision model.