The Hidden Flaw in LLM Few-Shot Predictions — And How a New ACL 2026 Method Exposes It
Introduces 'self-function vectors' to separate aleatoric from epistemic uncertainty in LLMs.
A new paper accepted to ACL 2026 tackles a critical issue in In-Context Learning (ICL): reliably measuring when an LLM is uncertain due to inherent data ambiguity (aleatoric uncertainty) versus model limitations (epistemic uncertainty). The authors—Jinseok Chung, Minkyoung Song, Hyunji Jung, and Namhoon Lee—introduce 'self-function vectors,' which leverage mechanistic interpretability to model the latent concept learned during in-context prompting. By embedding these vectors into a Bayesian framework, they directly estimate aleatoric uncertainty without relying on brittle input perturbations or decoding tweaks, which existing methods often require.
To validate their approach, the researchers developed the first dedicated evaluation protocol for aleatoric uncertainty in ICL. They manipulated synthetic data in controlled ways to isolate uncertainty sources, then extended tests to real-world datasets. Results show their method outperforms existing alternatives in measuring prediction confidence and proves practical for hallucination detection. This work bridges quantitative uncertainty decomposition with mechanistic understanding of model behavior, offering a more trustworthy path for LLM deployment in high-stakes applications.
- Introduces 'self-function vectors' derived from internal LLM representations to estimate aleatoric uncertainty in In-Context Learning within a Bayesian framework.
- Proposes first rigorous evaluation protocol for aleatoric vs. epistemic uncertainty separation, validated on synthetic and real-world datasets.
- Demonstrates practical use for hallucination detection, outperforming existing uncertainty estimation methods.
Why It Matters
Enables more trustworthy LLM outputs by precisely identifying when data ambiguity, not model error, causes uncertainty.