NOEMA: Procedure-Oriented Post-Training for Epistemic-State Control in Mid-Sized Language Models

Project NOEMA investigates epistemic-state control in mid-sized language models under incomplete information. Rather than treating reasoning solely as final-answer generation, NOEMA trains models to preserve what is established, disconfirmed, or unresolved; localize unresolved conditions; infer the evidence required for further resolution; and maintain the resulting epistemic state through natural-language rendering. In experiments with Qwen3-14B, the final Stage 1 adapter achieved correct final judgments on all 29 cases in a fresh-base evaluation, and subsequent structured-IR training completed the evaluated Full-flow procedure without residual procedural failure in the tested setting. A cross-family study with Llama3.2-11B achieved semantic Full-flow success on 11 of 12 final evaluation cases. The sole remaining failure was localized to condition-level epistemic-state classification, while subsequent routing remained internally consistent with the predicted state. This record contains both the English and Japanese editions of Version 1.0 of the paper. DOI: 10.5281/zenodo.22789884