Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2025
|
| Online Access: | https://doi.org/10.5281/zenodo.17648656 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <p><strong>The Empty Brain Hypothesis (EBH)</strong> introduces a model-agnostic framework for understanding silent epistemic gaps in large language models (LLMs). Unlike the human body, which signals physical hunger, both biological and artificial minds often fail to signal intellectual emptiness. This asymmetry creates “silent voids” that manifest as confident hallucinations, representational drift, or biased overgeneralization—problems that now define the frontier of AI reliability.</p> <p>This paper formalizes that phenomenon through the <strong>Empty Brain Model (EBM)</strong>, a tri-state cognitive architecture comprising:</p> <ol> <li> <p><strong>Satiated State</strong> — coherent reasoning with high alignment to learned distributions</p> </li> <li> <p><strong>Empty State</strong> — epistemic voids revealed through perplexity spikes, semantic drift, and Gap Score escalation</p> </li> <li> <p><strong>Overfed State</strong> — degradation caused by excessive low-quality or biased inputs (“model rot”)</p> </li> </ol> <p>To address these states proactively, the paper develops a system of <strong>bio-inspired “hunger signals”</strong>, combining internal sensors (perplexity, drift metrics, Bayesian surprise <span><span>G(t)G(t)</span><span><span><span>G</span><span>(</span><span>t</span><span>)</span></span></span></span>) with external probes (stochastic error ascent, abstention classifiers). These mechanisms detect emerging voids in real time—before hallucinations manifest.</p> <p>The paper further proposes <strong>feeding protocols</strong> that restore epistemic health through controlled retrieval, diverse data injections, and ethical balance audits. Experimental validation on Llama-3 models demonstrates:</p> <ul> <li> <p><strong>42% reduction</strong> in hallucinations</p> </li> <li> <p><strong>F1 = 0.66</strong> detection accuracy for silent gaps</p> </li> <li> <p><strong>40–67% improvement</strong> in post-remediation tasks</p> </li> </ul> <p>A unique aspect of this work is the <strong>multi-model case study</strong>, in which systems such as Grok, Claude, Gemini, ChatGPT, Perplexity, and NotebookLM first revealed their voids unknowingly and later participated knowingly in cross-model synthesis. This reflexive process—AI evaluating AI—provides rare empirical insight into architectural blind spots.</p> <p>Overall, EBH advances a shift from reactive correction to <strong>proactive epistemic regulation</strong>. It positions LLMs not as oracles but as minds requiring continual nourishment, capable of abstaining, self-probing, and seeking truth intentionally. The framework remains fully <strong>model-agnostic</strong> and applicable across current and future AI architectures.</p>