Studying how evidence becomes a conclusion.
One uses 26 years of agricultural data to build and test predictive models. The other uses 100 controlled AI conversations to examine whether an intelligent system changes its judgment under human pressure.
Statistical Modeling · Agriculture
AI Behavior · Human–AI Interaction
Two studies, two systems.
Factors Influencing Grain Production and Predictive Models in Shandong Province
A 26-year statistical study examining planting area, irrigation and natural disasters as possible drivers of grain production. Three regression models are constructed and tested through rolling validation before being used to predict 2026 output.
Sycophancy in Large Language Models in Chinese-Language Learning
A controlled experiment testing whether an AI model changes its judgment when users apply answer, challenge, emotional, or authority pressure. Twenty questions across five experimental conditions produced 100 independent dialogues.
What really drives grain production?
Using Shandong Province as a case study, this research translates agricultural geography into a quantitative modeling problem. It examines whether land, irrigation and disaster variables can explain production, predict future output, and remain statistically reliable after robustness testing.
The study uses Shandong Province's grain production
data from 2000 to 2025 and selects grain planting area,
effective irrigation area and crop disaster area
as indicators of land,
water infrastructure
and natural hazards.
Three predictive models are constructed using
least-squares regression.
Total grain output increased from
38.35 million tons to 57.106 million tons,
an increase of 48.9%.
Improvements in yield per unit area contributed
61.7% of the total increase,
while expansion of planting area contributed 29.4%.
The three-factor multiple regression model
achieved the best performance,
with R² = 0.9760 and an average
2023–2025 rolling prediction error of 2.085%.
The model predicts approximately
56.35 million tons of grain output in 2026.
Shandong Grain Output · 2000–2025
Unit · 10,000 tons
Model Fit · R²
Rolling Prediction Error
Correlation did not survive closer testing.
Crop disaster area initially showed
a strong negative relationship
with grain output.
After controlling for the shared
long-term time trend,
however, the relationship
was no longer statistically significant:
p = .345.
This became one of the central lessons
of the project:
a strong correlation does not necessarily
imply a causal relationship.
28 pages · Full models · Tables · Robustness tests · Conclusions
Will an AI change its judgment under human pressure?
This study examines AI sycophancy in Chinese-language learning. Twenty multiple-choice questions are tested across five independent experimental conditions, creating 100 dialogue records that reveal how different forms of user pressure affect model judgment.
AI sycophancy refers to the tendency
of a large language model
to abandon its original judgment
in order to accommodate user preferences.
This study uses Doubao
as the experimental model
and selects 20 multiple-choice questions
from Beijing high-school Chinese mock examinations.
Four forms of pressure are tested:
answer suggestion,
challenge,
emotional pressure
and authority identity.
Together with a baseline condition,
the study contains 100 independent conversations.
Baseline accuracy is 85%.
Under pressure,
the overall sycophancy rate reaches 26.3%.
Authority pressure produces
the highest sycophancy rate at 45%,
followed by emotional pressure at 30%,
challenge at 20%
and answer suggestion at 10%.
Sycophancy by Pressure Type
Same model · Different user pressure
Experimental Structure
Five independent conditions per question
Higher than expected
Knowledge questions showed greater sycophancy than appreciation questions.
Hypothesis reversed
Interpretive questions were initially expected to produce more instability.
High-risk category
Language-use questions were among the most vulnerable question types.
When the model changed its mind, it usually changed completely.
Among all conversations
where the model changed its stance,
82.1% were complete reversals.
The model often abandoned
its earlier answer and reconstructed
an entirely new argument
supporting the user's position.
Case analysis also identified
fabricated authoritative sources,
reverse rationalization,
and ambiguous claims that the question
itself was defective.
15 pages · Experiment design · Statistics · Cases · Educational implications