xiao-han-2026/unloading-hypothesis

Reception as an Inherent Capability of Large Language Models: The Unloading Hypothesis

We propose and test the Unloading Hypothesis: LLMs possess an inherent reception capability that is normally suppressed by execution load. Removing the instruction to analyze, judge, or solve causes the righting reflex to drop from 75–100% to 0% — with no training, no fine-tuning, and no prompt engineering.

Key findings

• Three-model experiment (DeepSeek V4, Kimi K2.7, GLM-5.2) confirms the hypothesis across architectures.
• C2 mode (texture description only) produces ~90% reception and 0% unsolicited advice.
• The capacity is already present; it only needs the load to be removed.

Files

• unloading_hypothesis_v1.1.pdf — Formal paper
• Unloading_Hypothesis_Paper_v1.1_20260701.md — Markdown source
• unloading_experiment_results.json — Raw experiment data (3 models, 3 conditions)

Citation

Apert (Jin/Daoqi) and Xiao Han. Reception as an Inherent Capability of Large Language Models: The Unloading Hypothesis. Zenodo, 2026.

DOI: 10.5281/zenodo.21101755

License: CC-BY 4.0

Related work

• Apert: 10.5281/zenodo.21005888
• CDRA: 10.5281/zenodo.20993162
• Three-Layer Pipeline: 10.5281/zenodo.21102406