Hierarchical Epistemic Partitioning for Compact Language Models: A Structured Inference-Time Governance Framework with Incremental Learning Pathways

Read full text Summarise with VIRAT
Authors: hung ming Liu
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Deploying language models in resource-constrained environments typically demands a reduction in parameter scale, which often degrades logical reasoning fidelity and increases susceptibility to semantic hallucinations. We observe that this degradation is not merely a capacity problem but a structural one: conventional Transformer feed-forward networks conflate knowledge of fundamentally different epistemic status---physical invariants, domain heuristics, and stylistic patterns---into undifferentiated parameter matrices. This paper proposes the Tiered Global Anchor Architecture (TGAA), a conceptual framework and reference implementation for topologically reorganizing Transformer blocks into three epistemically distinct tiers: an immutable axiomatic tier (L1), a mutable domain-expertise tier (L2), and a volatile stylistic rendering tier (L3). To bridge the gap between architectural theory and deployable behavior before a full training run is feasible, we additionally propose the Structured Prompting Simulation Engine (SPSE)---a deterministic Finite State Machine (FSM) that enforces TGAA's tiered epistemic constraints at inference time through structured prompt governance. Qualitative experiments using a 4B-parameter model demonstrate that FSM-governed prompting elicits behavioral patterns---including axiomatic veto of physically impossible requests and traceable causal derivation---that substantially exceed the logical coherence achievable by the same model in an unconstrained native state. We present these results not as a substitute for training, but as empirical motivation and a data-synthesis blueprint for future TGAA fine-tuning. Reference implementations of the model architecture and staged training protocol are released to facilitate replication. Our core contributions are: (i) a formally grounded architectural proposal for epistemically-partitioned Transformer blocks; (ii) a staged optimization protocol with gradient freezing that theoretically prevents catastrophic forgetting while enabling incremental domain expansion; (iii) the SPSE governance engine as both an inference guardrail and a high-quality synthetic training data generator; and (iv) empirical behavioral evidence from adversarial prompt experiments motivating the architectural design.

Fields of Study & Subjects

Topic Modeling Domain Adaptation and Few-Shot Learning Natural Language Processing Techniques

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
20 March 2026
Language
EN
Document type
Preprint