Abstract
ABSTRACT: Retrieval-of-Analogy-Mapped-Synthetics (RAMS): Generating High-Density Synthetic Reasoning Data for Large Language Model Training This disclosure describes a novel application of the Retrieval-of-Analogy-Mapped-Synthetics (RAMS) framework for the large-scale generation of synthetic training data focused on abstract reasoning, fluid intelligence, and out-of-distribution (OOD) semantic generalization. Building upon the previously disclosed Forward-RAMS prototype, the system automatically identifies structurally and functionally analogous concepts across semantically distant domains, while employing aggressive negative-token filtering to eliminate near-neighbor associations, synonyms, and domain-specific jargon. This constraint compels the underlying model to discover deeply isolated relational structures rather than rely upon superficial lexical similarity. The proposed approach enables rapid production of high-quality Semantically Distant Analogies (SDAs), each accompanied by explicit, traceable reasoning that explains the correspondence between source-domain and target-domain concepts. Because the number of specialized concepts across human knowledge domains is effectively unlimited, the framework provides a scalable mechanism for generating vast quantities of synthetic reasoning examples. Additional system features include automated validation filters for detecting hallucinated or repetitive mappings, structural novelty screening, and automated endpoint detection to identify analogy saturation. Optional human review can further enhance dataset quality. Unlike conventional retrieval and synthetic-data generation approaches that emphasize positive vector similarity, RAMS deliberately seeks structurally isomorphic relationships between highly unrelated knowledge domains. The resulting datasets may offer a rich source of training material for frontier large language models by encouraging abstraction, transfer learning, analogical reasoning, and cross-domain generalization. This work suggests that semantically distant analogy generation may represent a practical and highly scalable pathway for creating synthetic reasoning corpora capable of supplementing or extending traditional human-authored training datasets.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Gillis, Herbert, "POTENTIAL APPLICATION FOR RETRIEVAL-OF-ANALOGY-MAPPED SYNTHETICS (RAMS): DEVELOPING SYNTHETIC TRAINING DATA FOR LARGE LANGUAGE MODELS (LLMs).", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12100