IC-894FLAN-T5 base's average per-token probability increases with token index during generation on WMT translation tasks

Neha Gupta, Harikrishna Narasimhan, Wittawat Jitkrittum, Ankit Singh Rawat, Aditya Krishna Menon, Sanjiv Kumar

SourceLanguage Model Cascades: Token-Level Uncertainty And Beyond

The paper measures the average prediction probability of each token position across all examples for FLAN-T5 base on WMT fr→en. The result shows that as the token index increases, the average probability increases, meaning the model becomes progressively more confident as it generates. This structural property causes the sum of log-probabilities (chow-sum) to be systematically lower for longer outputs, and the average (chow-average) to over-correct toward shorter outputs. The effect is visible in the deferral curves where chow-sum preferentially defers longer predictions and chow-average preferentially defers shorter ones.

Evidence
correlational
Caveat
Measured on a single dataset (WMT fr→en) and a single model size (base); the paper does not report the same analysis for other sizes or tasks in the main text.
Model
FLAN-T5
Datasets
WMT 14: fr→en [eval]
Related findings
IC-895
Extraction
automatic-extraction