Look Before You Leap: Thermodynamic Arbitration of Parametric and Non-Parametric Knowledge in LLM Agents via Self-Regulating Memory Architectures

Akash Das, Ishan Roy

Keywords: Metacognitive Agentic Memory Arbitration, Epistemic Uncertainty, Thermodynamic Regularization, Contrastive Learning, Efficient Inference, Hallucination Mitigation

TL;DR: MARTA introduces a thermodynamic control layer that empowers frozen LLMs to 'look before they leap,' dynamically arbitrating between internal knowledge and external retrieval to strictly decouple semantic similarity from epistemic utility

Abstract:

The architecture of modern LLMs consists of one profound cognitive polarization. LLMs have an intuition (implicitly) in the vast range of their parameters, yet rely on a disconnected, explicit mechanism to reach the outside world. We have not managed to bridge this gap using agentic frameworks, where models are compelled into pathological ”induced amnesia.” With the prevailing “Retrieve- Always” paradigm, agents must always distrust their own internal knowledge, and every user interaction is a ”tabula rasa” event which must be checked externally. Therefore, they create processes of reflexive dependence that tend to be thermodynamically wasteful, cognitively fragile, and susceptible to the noise of irrelevant context. In this research, we propose a return to first principles, operationalizing the biological maxim to ”Look Before You Leap.” Here we introduce MARTA (Metacognitive Adaptive Retrieval and Thought Architecture), a neuro-symbolic framework proposed to cure the rift between Parametric and Non-Parametric knowledge. In contrast to perceiving retrieval as an essential step, MARTA presents it as a thermodynamic cost— it is a “leap” only taken when perceived internal inadequacy in the level of “looking” occurs. By allowing the agent to gauge the entropy of its own thoughts before engaging in action, we turn the receiver of text into a deliberative one — an agent that mediates their own certainty. Our method makes clear: if we allow agents access to silence and introspection,
they restore an equilibrium that brings efficiency and truth to the center, that takes
us closer to machines that know, and are aware of themselves

Submission Number: 17