Evaluating 35 open-weight models across three context lengths (32K, 128K, 200K), four temperatures, and three hardware platforms—consuming 172 billion tokens across more than 4,000 runs—we find that the answer is “substantially, and unavoidably.” Even under optimal conditions—best model, best temperature, temperature chosen specifically to minimize fabrication—the floor is non-zero and rises steeply with context length. At 32K, the best model (GLM 4.5) fabricates 1.19% of answers, top-tier models fabricate 5–7%, and the median model fabricates roughly 25%.
Okay, this is looking promising, at least in terms of the most important qualifications being plainly stated in the opening line.
Because the amount of hallucinations/inaccuracies “in the wild” - depending on the model being tested - runs about 60-80%. But then again, this would be average use on generalized data sets, not questions focusing on specific documentation. So of course the “in the wild” questions will see a higher rate.
This also helps users, as it shows that hallucinations/inaccuracies can be reduced by as much as ⅔ by simply limiting LLMs to specific documentation that the user is certain contains the desired information, rather than letting them trawl world+dog.
Very interesting!
As I mentioned elsewhere (below) I am currently conducting similar testing across 4 different 4B models (Qwen3-4B Hivemind, Qwen3-4B-2507-Instruct, Phi-4-mini, Granite-4-3B-micro), using both grounded and ungrounded conditions. Aiming for 10,000 runs, currently at 3,500.
Not to count chickens before they hatch - but at ctx 8192, hallucination flags in the grounded condition are trending toward near-zero across the models tested (so far). If that holds across the full campaign, useful to know. If it doesn’t hold, also useful to know.
I have an idea for how to make grounded state even more useful. Again, chickens not hatched blah blah. I’ll share what I find here if there’s interest. I’m intending to submit the whole shooting match for peer review (TMLR or JMLR) and put it on arXiv for others to poke at.
I realize this is peak “fine, I’ll do it myself” energy after getting sick of ChatGPT’s bullshit, but I got sick of ChatGPT’s bullshit and wanted to try something to ameliorate it.
I have been saying this for awhile. I am sorta hoping we see open source llms that are trained on a curated list of literature. its funny that these came out and it seemed like the makers did not take the long known garbage in - garbage out into account.