Learn2Discern Benchmark: LLMs Barely Tell Reliable Sources from Popular Ones

In July 2026, researchers published a paper on arXiv about "information discernment" in large language models and introduced a benchmark called Learn2Discern (L2D). Across 13 models and nearly 670,000 trials, they found that models tell reliable sources from unreliable ones at roughly the level of random guessing.

What Learn2Discern Tests

Learn2Discern evaluates two skills a model needs when working with external data: source discernment, meaning whether the model trusts a reliable source more, and truth discernment, meaning whether it updates its answer more when a new claim moves it closer to the truth. The framework rests on three normative axioms with interpretable metrics. According to the authors, real users share these principles: a pre-registered survey with a quota sample (n=299) showed that violating them reduces trust in the model and willingness to use it.

Key Results

Models fail on both dimensions: results for both source reliability and closeness to the truth stay near the random level.

• 13 models, nearly 670,000 trials
• Models rely on a source's popularity twice as much as on its reliability
• They update their answers about equally, whether a claim moves them toward the truth or away from it
• They integrate external knowledge best where their own priors are already the most accurate
• User survey: n=299, quota sample, pre-registered

"Models perform near chance on both source reliability and closeness

to the truth," the study's abstract on arXiv says.

Why Newer Models Don't Fix This

Larger and more recent models improve only truth discernment, not source discernment. This is a blind spot that more model complexity does not cure: the ability to tell a credible source from a popular one does not grow with size. Worse, according to Learn2Discern, models incorporate external knowledge most effectively where their priors are already the most accurate, which means they help themselves most where help is needed least. The authors also found some good news: simple inference-time interventions improve both skills at once, without fine-tuning the model.

What This Means

As LLMs replace traditional search, the ability to tell a credible source from a merely popular one becomes a critical alignment property. So far, judging by the 13 models tested, they have not learned this skill, and model size does not solve the problem. The authors have released the dataset and the survey publicly as a testbed for this task.

Frequently Asked Questions

What is Learn2Discern?

Learn2Discern (L2D) is an experimental framework and benchmark for evaluating how LLMs weigh external information: whether they trust reliable sources and whether they move closer to the truth. It is built on three normative axioms with interpretable metrics; the authors have opened the dataset and the survey for reuse.

Which models were tested?

The study covered 13 models across nearly 670,000 trials. All of them showed source and truth discernment near the random level, and newer, larger models improved only truth discernment, not source discernment.

Why does this matter for AI search?

When LLMs replace search engines, they decide for themselves which source to believe. If a model leans on a source's popularity twice as much as on its reliability, it risks amplifying claims that are widespread but wrong.