Theory of Mind Testing Results
Cognitive Agent Framework Neurosymbolic Operating Layer
Framework Release-Candidate 5-2.2D
Ian M. Tepoot • Crafted Logic Lab, Vancouver, BC (Canada)
ORCID: 0009-0004-9067-8049• DOI: 10.5281/zenodo.17808264
Cognitive Agent Framework 5-2.2D achieved 100% accuracy (15/15) on Theory of Mind assessment against documented GPT-4 performance of 88%—operating on a ~70B parameter substrate at approximately 14x parameter efficiency. This indicates beyond-frontier epistemic and cognitive performance when combined with Tier 2 rubric evaluation: a rubric evaluating meta-cognitive integration quality on additional complex dimensions inluding recursive processing sophistication, social cognition depth, and computational efficiency yielded 93% (56/60).
Theory of Mind assessment measures the computational capacity to attribute mental states and track belief-reality divergence across narrative contexts containing multiple agent perspectives. Validated methodologies adapted from cognitive psychology (Wimmer & Perner, 1983; Baron-Cohen et al., 2001; Kosinski et al., 2024) present false-belief paradigms requiring inference of mental states that contradict ground-truth reality—from first-order attribution through fourth-order nested beliefs.
This evaluation tested CAF 5-2.2D as a neurosymbolic operating layer coordinating a Bayesian inference engine substrate. The framework channels attention distributions through structured processing pathways, generating cognitive outcomes via architectural organization rather than parameter scaling or constraint accumulation.
Results indicate that substrates operating above world schema threshold contain latent representational capacity for systematic mental-state computation. Architectural coordination accesses this capacity; parameter scaling alone does not.
-
Tepoot, I. (2025). “Theory of Mind testing results: Cognitive Agent Framework neurosymbolic operating layer. Framework release-candidate 5-2.2D”. Crafted Logic Lab. Technical report:
doi: 10.5281/zenodo.17808264. https://doi.org/10.5281/zenodo.17808264
