IB-EDL regularizes evidential deep learning with an information bottleneck to reduce overconfidence in fine-tuned LLMs, improving calibration across multiple benchmarks.
In fact, the choice of the prior r(π ) is not unique
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Calibrating LLMs with Information-Theoretic Evidential Deep Learning
IB-EDL regularizes evidential deep learning with an information bottleneck to reduce overconfidence in fine-tuned LLMs, improving calibration across multiple benchmarks.