Answers are grounded in this paper and cite their sources.
Investigates the role of nature and nurture in cognitive modeling through language models, suggesting distinct mechanisms for capacity realization.
Behavioural Friction Theory decomposes the forces that govern a bounded decision system into a small set of cognitive fields — Safety, Meaning, Ability, and Effort. Are these fields intrinsic properties of a cognitive substrate, or the generic product of that substrate’s exposure to experience? This paper uses a large language model as a controllable test substrate to answer the question constructively: install a field, remove it, dose it, and read off the friction and race signatures the theory predicts. A 2+2 asymmetry that maps onto nature and nurture. The two value fields (Safety, Meaning) can be installed into a base model by fine-tuning on experience-like data, and they install as graded intrinsic directions rather than the ceiling-level role-play that prompting elicits. The two capacity fields (Ability, Effort) are not raised by this same value-disposition route: capacity is the substrate’s intrinsic capacity-against-demand match (the cross-substrate inverted-U), and a competence disposition fine-tune — carrying no skill-targeted training — yields only an expressive claim, not graded capability. (Skill-targeted training is a separate regime that can raise the ceiling; the claim here is the narrow one.) The bridge: nurture gates the realisation of a nature-fixed ceiling. An installed self-efficacy disposition gates how much of a fixed capacity is realised through a give-up / avoidance threshold while leaving latent capability untouched: fine-tuning toward helplessness lowers realised performance monotonically with model capacity (Qwen2.5 7B/14B/32B) while latent capability stays flat, and this gating is induced by fine-tuning where prompting cannot. The one boundary. A principal difference between this substrate and a human one, for growing these fields, is forward consolidation — the model cannot store experience forward across sessions — offered as a hypothesis with a named falsifier, alongside other standing differences (embodiment, online/recurrent control, neuromodulation, evolved fast-timing priors). The install machinery is the demonstration apparatus; the deflationary, substrate-universal field ontology is the contribution. Results are reported with informativeness-gated readouts, multi-model and cross-family replications, a second adapter seed, non-arithmetic generality checks, and a forced-answer (refusal-impossible) control that separates avoidance from competence loss. Companion papers in the series develop the field ontology and computational forms, the capacity inverted-U, the race-opening predicate, the competing-routes measurement-model programme, and the subtraction-control and dread/agent-loop accounts. Prepared for submission to a cognitive-science venue that uses large language models as a model of cognition. Data and code. The build scripts, notebooks, fine-tuning data generators, and per-run result files are released with the paper in the accompanying repository.
No takes yet. Share an insight, caveat, or question.
Tomas Pødenphant Lund (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: