Provides empirical evidence that LLMs hallucinate not from a lack of internal uncertainty, but because that uncertainty is 'functionally silent' during output generation.
arXiv · March 17, 2026 · 2603.13911
The Takeaway
The study shows that models internalize uncertainty in high-dimensional regions but fail to couple it to the logit layer due to training objectives. This suggests that solving hallucinations requires architectural changes to integrate existing uncertainty signals rather than simply more or better data.
From the abstract
We show that language models hallucinate not because they fail to detect uncertainty, but because of a failure to integrate it into output generation. Across architectures, uncertain inputs are reliably identified, occupying high-dimensional regions with 2-3$\times$ the intrinsic dimensionality of factual inputs. However, this internal signal is weakly coupled to the output layer: uncertainty migrates into low-sensitivity subspaces, becoming geometrically amplified yet functionally silent. Topol