M-CLTR: Measurable Controlled Latent Trajectory Reasoning

This paper introduces M-CLTR (Measurable Controlled Latent Trajectory Reasoning), an exploratory diagnostic framework designed to isolate, measure, and analyze control dynamics within the latent reasoning trajectories of language models. Rather than presenting a finalized architecture or claiming state-of-the-art performance, this work establishes an experimental scaffold to untangle three frequently conflated dimensions of latent reasoning: knowledge availability, theoretical steerability (via an iterative optimization oracle), and the training feasibility of a practical learned controller. Through a sequence of synthetic verification tasks, the empirical findings converge on three core insights. First, the latent control space is mathematically viable, as evidenced by consistent and robust oracle performance advantages. Second, the underlying quality of the pre-existing knowledge base strongly dominates final accuracy numbers. Third, a standard learned controller remains weak, struggling to match the oracle path and often being washed out by token-level knowledge strength. M-CLTR contributes a clean methodological lens to the field of test-time compute and latent steering, clarifying the mechanical trade-offs between internal path overwrites and state restarts. Fully committed to open science, the manuscript includes transparent links to the originating large language model (LLM) conversation logs and the public Google Colab experimental notebooks.