Behavioral Regime as Representational Sculptor: How Exploration–Exploitation Balance, Predictive Horizon, and Neural Dimensionality Co-Determine Internal Geometry

Version 2 — revised in response to an external structural review and an automated critique pass. See "Response to Review" appendix in the PDF for the change log.

A structural pattern appears across several recent preprints spanning computational neuroscience and language modeling: the *regime* in which an agent or model operates—exploratory versus exploitative, short-horizon predictive versus long-horizon predictive, high-dimensional versus compressed—does not merely influence performance on a given task but actively sculpts the *geometry* of internal representations. This synthesis assembles six to eight findings from q-bio.NC and cs.CL preprints to argue, as a heuristic reading rather than a formal derivation, that behavioral regime and representational geometry are mutually constitutive: the statistics of experience generated by a behavioral policy shape the latent space, and that latent space in turn constrains what future behaviors are accessible. Specific findings include: (1) exploratory agents in a maze develop more spatially organized, transition-preserving predictive representations than exploitative agents [corpus:arxiv:2605.27929]; (2) successor representations trained on natural language spontaneously produce syntactically organized geometric structure whose granularity depends on predictive horizon [corpus:arxiv:2605.24585]; (3) random neural networks, when finite-time and contextual variability are incorporated, match empirically observed low dimensionality of neural recordings, and representational orientation similarity across behavioral contexts is more sensitive to connectivity than dimensionality alone [corpus:arxiv:2605.26551]; (4) highly predictive digital twins of mouse V1 exhibit flatter eigenspectra—higher-dimensional representations—than less predictive models [corpus:arxiv:2605.23122]; (5) efficient coding under resource constraints drives neural populations toward criticality and sloppiness via diverging correlation lengths [corpus:arxiv:2605.22598]; (6) contextual role dynamically remaps object representational geometry in human cortex, with action-target objects organized by affordance dimensions and passive objects by semantic dimensions [corpus:arxiv:2605.23111]; and (7) sparse autoencoder features from LLMs recover cortical semantic topography, with semantic features alone recovering 94% of peak encoding performance [corpus:arxiv:2605.23035]. The central falsification path is: if behavioral regime sculpts representational geometry rather than merely selecting among pre-existing representations, then interventions that hold task performance constant while varying behavioral history should produce measurably different representational geometries—testable via orientation similarity of neural manifolds across conditions. ---

Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted from arXiv preprint corpus on the date in the filename.

Cited arXiv preprints: 2605.22598, 2605.22988, 2605.23032, 2605.23035, 2605.23111, 2605.23122, 2605.24585, 2605.26551, 2605.27929, 2605.30882, 2605.31473