Two distinct populations of internal features drive how an LLM handles being wrong versus being unsure.
April 23, 2026
Original Paper
Are LLM Uncertainty and Correctness Encoded by the Same Features? A Functional Dissociation via Sparse Autoencoders
arXiv · 2604.19974
The Takeaway
Sparse autoencoders reveal that a model's internal feeling of uncertainty is mechanically separate from its actual correctness. Many features are confounded, meaning they mix signals of doubt with signals of inaccuracy in a way that degrades performance. Suppressing these specific confounded features through targeted intervention makes the model significantly more accurate. This functional dissociation proves that an AI can be confident and wrong for very specific structural reasons. Practitioners can now clean up these internal signals to build models that are both smarter and more honest about their own limits.
From the abstract
Large language models can be uncertain yet correct, or confident yet wrong, raising the question of whether their output-level uncertainty and their actual correctness are driven by the same internal mechanisms or by distinct feature populations. We introduce a 2x2 framework that partitions model predictions along correctness and confidence axes, and uses sparse autoencoders to identify features associated with each dimension independently. Applying this to Llama-3.1-8B and Gemma-2-9B, we identi