We will release all code and data in the coming weeks to encourage further research. To read up on all the omitted details, check out the whole paper here:

👉 arxiv.org/pdf/2601.03746 📄

7/7

TL;DR: Multiple factors of source credibility influence how LLMs resolve knowledge conflicts, following a consistent, internal credibility hierarchy. But these preferences can be easily overwritten by simple repetition. Fine-tuning can mitigate this vulnerability.

6/7

To address this, we propose a novel knowledge-distillation training approach that makes models agnostic to repeated information. This reduces repetition bias by up to 99.8%, while retaining up to 88.8% of the original source preference.

5/7

By explicitly providing source information, we disentangle whether models truly favor majorities as often reported, or whether they are simply influenced by repeated information. Our findings strongly suggest the latter, leaving models vulnerable to adversarial manipulation.

4/7

When comparing different source types, we find that source information significantly affects conflict resolution, with models following a highly consistent source credibility hierarchy: institutional sources (government, newspaper) over individual ones (person, social media user).

3/7

We evaluate 13 models from 4 model families on fully synthetic conflicts and sources and measure shifts in answer probabilities when conflicting answers are attributed to different sources, all grounded in interdisciplinary frameworks.

2/7