HRM Taxonomy Generalization

Investigating whether the four-mode HRM reasoning taxonomy generalizes to five recursive reasoning architectures across 10,000 trajectory simulations.

10,000

Total Trajectories

100%

Non-halting Coverage

4.3%

UT Coverage (halting)

0.55%

CGTR Coverage (halting)

Proposed Extended Modes

Taxonomy Coverage

Four-Mode Coverage by Architecture

Mode Distributions (Aggregated)

Cross-Architecture Agreement

Pairwise Jensen-Shannon Divergence

Mode Distribution by Difficulty (HRM)

Detailed Results

Taxonomy Coverage

Architecture	Halting	Coverage	95% CI
HRM	No	1.000	[1.0, 1.0]
RMT	No	1.000	[1.0, 1.0]
LT	No	1.000	[1.0, 1.0]
UT	Yes	0.043	[0.035, 0.052]
CGTR	Yes	0.006	[0.003, 0.009]

Pairwise JSD (within-cluster in bold)

	HRM	UT	RMT	LT	CGTR
HRM	-	0.610	0.017	0.004	0.677
UT	0.610	-	0.617	0.613	0.008
RMT	0.017	0.617	-	0.005	0.678
LT	0.004	0.613	0.005	-	0.678
CGTR	0.677	0.008	0.678	0.678	-