Just-in-Time Construal: Efficient Simplified Representations for Mental Simulation

Problem Statement & Algorithm

How do cognitive agents efficiently determine which scene elements to encode for mental simulation? With N elements, there are 2^N possible construals. JIT-C sidesteps this by building representations on-demand during simulation.

Step 1 -- Plan: BFS on current construal to find shortest path to goal.

Step 2 -- Uncertainty Check: At each path position, estimate uncertainty from nearby un-encoded elements. If u > tau, expand.

Step 3 -- Saliency-Gated Encoding: Select top-k un-encoded elements by composite saliency (type prior x path proximity x goal alignment).

Step 4 -- Re-plan: Re-run BFS with expanded construal and continue.

Strategy Comparison (100 Grid Worlds, 37 Elements)

Elements Encoded

Mean Collisions

Success Rate (%)

Strategy	Encoded	Abs. Ratio	Collisions	Path Length
Full Scene	37.0	1.000	0.00	19.2
Oracle	0.5	0.015	3.32	23.0
JIT (tau=1.5)	30.4	0.821	0.00	19.2
JIT (tau=2.5)	24.5	0.663	0.00	19.2
JIT (tau=4.0)	16.3	0.442	0.02	19.2
Random 30%	11.0	0.297	2.61	22.8
Random 50%	18.0	0.486	1.75	21.7

Threshold Sensitivity Analysis

Elements Encoded vs. Threshold

Collisions vs. Threshold

Lowering tau from 10.0 to 0.5 increases encoding from 7.3 to 35.2 elements while eliminating collisions entirely. All threshold values maintain 100% success. Slope: ~-5.1 elements per unit of tau in range [0.5, 4.0].

Sub-Linear Complexity Scaling

JIT-C Encoding vs. Total Elements

Encoding Ratio vs. Complexity

Power law: y = a * x^b, b < 1 (sub-linear)

At 5 elements: 68% encoded

At 40 elements: 80% encoded

Gap grows with complexity

Behavioral Predictions

Distractor Robustness: Collisions by Distractor Fraction

Time-Pressure x Complexity Interaction

Key Results

Efficiency

34-56% fewer elements

Zero collisions at tau=2.5

100% success rate

Graceful Degradation

Smooth cost-accuracy tradeoff

100% success at all thresholds

Pareto-efficient frontier

Scalability

Sub-linear encoding growth

Advantage grows with complexity

Distractor-robust