Answers are grounded in this paper and cite their sources.
Reveals a new scaling law for LLM efficiency and hallucination control, suggesting a method from nuclear physics.
The training of large language models (LLMs) heavily relies on empirical scaling laws, which lack rigorous first-principles physical support, leading to inefficient resource allocation and unexplainable model hallucinations. Drawing on mature neutron resonance and Doppler broadening theories in nuclear reactor physics, this paper establishes a strict formal isomorphic mapping between the neutron scattering time-of-flight (TOF) spectrometer measurement system and LLM training dynamics, proposing three core physical concepts: knowledge resonance peak, equivalent data temperature, and resolution matching. By mapping model parameters to neutron energy resolution, we derive a dynamic scaling law adaptively adjusted by data noise. Further, combined with the physical characteristics of TOF spectrometers, we introduce Fourier truncation of step signals and Gibbs oscillation effects, proving that LLM boundary hallucinations are essentially inherent mathematical results under finite bandwidth constraints, with a constant ~9% overshoot magnitude consistent with nuclear physics experimental observations. This paper proposes a two-stage hierarchical training paradigm: coarse scan positioning and fine-tuning optimization, aligned with standard experimental procedures verified over 70 years in nuclear physics, which can simultaneously improve training efficiency and model reliability, providing quantifiable and verifiable physical design criteria for LLM training.
No takes yet. Share an insight, caveat, or question.
Wei Wei (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: