IC-1319Larger LLaMA and LLaMA2 models show better calibration on phrase-level tasks but not consistently on sentence- and paragraph-level tasks

Xin Liu, Muhammad Khalifa, Lu Wang

SourceLitCab: Lightweight Language Model Calibration over Short- and Long-form Responses

The paper evaluates calibration (ECE, Brier score) of LLaMA-7B, LLaMA-13B, LLaMA-30B, LLaMA2-7B, and LLaMA2-13B across phrase-, sentence-, and paragraph-level generation tasks. On phrase-level tasks, the larger models in each family achieve lower ECE and Brier scores than their smaller counterparts. However, on sentence- and paragraph-level tasks, the scaling trend breaks down: for example, LLaMA-30B has higher ECE than LLaMA-7B on WikiQA (0.142 vs. 0.108) and WikiGen (0.165 vs. 0.102), indicating that the within-family scaling benefit does not extend to longer generations.

Evidence
correlational
Key metric
Phrase-level NQ ECE: LLaMA-7b 0.144, LLaMA-13b 0.123, LLaMA-30b 0.169; Sentence-level WikiQA ECE: LLaMA-7b 0.108, LLaMA-13b 0.114, LLaMA-30b 0.142; Paragraph-level WikiGen ECE: LLaMA-7b 0.102, LLaMA-13b 0.124, LLaMA-30b 0.165
Caveat
The scaling trend is inconsistent across individual tasks within the same length category; the paper states the effect 'may not hold true' rather than being uniformly absent.
Model
LLaMA, Llama 2 / Llama 2 base
Concepts
Scale-dependent behaviour
Datasets
Natural Questions / NaturalQA [eval], SciQ / SciQA [eval], TriviaQA [eval], TruthfulQA / TruthfulQA MC1 [eval], WikiQA [eval], BioGen [eval], WikiGen [eval], QAMPARI [eval]
Methods
Expected Calibration Error / Integral Calibration Error (ECE) [eval], Brier score [eval]
Related work
Kadavath et al. (2022) [context]
Related findings
IC-1320, IC-1321
Extraction
automatic-extraction