Junk Training Data Degrades AI Model Reasoning

Researchers at Texas A&M University found that training large language models on popular, short and low quality social media posts can reduce reasoning and memory performance, according to Texas A&M Engineering. The study was accepted to the 2026 Conference on Language Modeling.

The team used 1 million public Twitter posts from 2010 and classified content based on engagement, length and writing quality. Researchers trained four models, including Llama3 8B Instruct and three Qwen models, on different proportions of junk and control data before testing reasoning, memory, ethical norms and personality.

Models exposed to more junk data increasingly skipped reasoning steps and sometimes gave answers that violated ethical norms. Popularity had a stronger effect on reasoning, while shorter content had a larger effect on memory and understanding. Later training with higher quality data and other corrective techniques did not fully restore model performance.