Texas A&M researchers find language models trained on junk social media posts start skipping reasoning steps, and better data doesn't fully fix them
Texas A&M Engineering reported on October 19th, 2026 that researchers in the Department of Computer Science and Engineering trained four open language models on mixes of junk and control text. The junk came from 1 million public Twitter posts collected in 2010. A junk post was highly engaged and short. The main change in the models was thought skipping, where a model skips steps before it answers. The researchers say later training on higher quality data did not fully restore performance. The paper was accepted to the 2026 Conference on Language Modeling. Larger closed models were not tested.
How to take part
The findings are described on the university's engineering news page. A research result is not a public proceeding and has no window for comment.
Where
1 county. In College Station-Bryan.
Timeline
- filed
Texas A&M Engineering reports the findings
- Today
How this decision moved
One dated line per check, oldest first. A line that says nothing changed means somebody looked and it had not.
- 2026-10-20
Admitted on the university's own account. The release reports no benchmark scores, so the size of the effect isn't stated here.
The evidence
Every fact above rests on one of these. The words are the source's own.
Tu, assistant professor in the Department of Computer Science and Engineering at Texas A&M University , found that when large language models (LLMs) train on junk data, they can get brain rot, too.Texas A&M Engineering News, AI models can get brain rot, too Primary source, official · news.engineering.tamu.edu
This research was recently accepted to the 2026 Conference on Language Modeling.Texas A&M Engineering News, AI models can get brain rot, too Primary source, official · news.engineering.tamu.edu
the main cognitive change in LLMs trained on junk data was thought-skipping, meaning the model will skip steps, or not “think” at all, before providing an answer.Texas A&M Engineering News, AI models can get brain rot, too Primary source, official · news.engineering.tamu.edu
Even after we use some higher quality data to train the model or deploy some techniques to improve the model’s response, it is very hard to recover the modelTexas A&M Engineering News, AI models can get brain rot, too Primary source, official · news.engineering.tamu.edu
the researchers used a dataset of 1 million public Twitter posts collected in 2010 that included engagement data, like the number of retweets.Texas A&M Engineering News, AI models can get brain rot, too Primary source, official · news.engineering.tamu.edu
researchers designated posts as junk if they received more than a combined 500 likes, retweets or replies and contained fewer than 30 tokens.Texas A&M Engineering News, AI models can get brain rot, too Primary source, official · news.engineering.tamu.edu
Once the junk and control datasets were assembled, the researchers trained four LLMs on various proportions of junk and control data. Through training, LLMs learn to make predictions based on a dataset. The four LLMs they selected — Llama3 8B Instruct, Qwen2.5 7B Instruct, Qwen2.5 0.5B Instruct and Qwen3 4B Instruct — are diverse in size and represent different model families.Texas A&M Engineering News, AI models can get brain rot, too Primary source, official · news.engineering.tamu.edu
Testing for brain rot in larger models like ChatGPT or Gemini, which are not open source, is also needed.Texas A&M Engineering News, AI models can get brain rot, too Primary source, official · news.engineering.tamu.edu
Questions about this decision
Answered from the record itself. Every answer is assembled from stored fields, so an answer the record has no basis for is left out rather than guessed.
What is this decision?
Texas A&M Engineering reported on October 19th, 2026 that researchers in the Department of Computer Science and Engineering trained four open language models on mixes of junk and control text. The junk came from 1 million public Twitter posts collected in 2010. A junk post was highly engaged and short. The main change in the models was thought skipping, where a model skips steps before it answers. The researchers say later training on higher quality data did not fully restore performance. The paper was accepted to the 2026 Conference on Language Modeling. Larger closed models were not tested.
Who decides it?
Texas A&M University decides. The record names the deciding body for every entry it carries.
Can the public take part?
The findings are described on the university's engineering news page. A research result is not a public proceeding and has no window for comment. No dated public window is on the record. The deciding body is named and reachable.
Where in Texas does it apply?
It covers Brazos County in the College Station-Bryan, TX area.
Has it been decided?
It has been decided. The dates on the item page carry when.
What happens next?
No future date is on the record. The last dated step on it was filed on October 19th.
When did it start?
The earliest date on its record is October 19th, 2026.
What kind of decision is it?
It is filed under research and science.
What sources back it?
One source backs it. It is primary.
Is it on the ERCOT grid?
Yes. It sits inside the ERCOT interconnection.
When was it last checked?
Every fact on it was last verified against its source on October 20th, 2026.
Cite this
Texas AI Docket, Texas A&M researchers find language models trained on junk social media posts start skipping reasoning steps, and better data doesn't fully fix them. Tracked since October 19th, 2026. Last verified October 20th, 2026. https://texasaidocket.com/item/tx-2026-0223/. Reuse permitted under CC BY 4.0 with attribution. The same entry is in the docket JSON as item tx-2026-0223.
Beat
Filed under Research and science, with every other decision on that beat.