Source
Albalak et al. (2023), LIT-117 — ARXIV-2312.02406.
Perplexity as a signal, and what ODM does with it
LIT-117’s mechanism is a multi-armed bandit over data domains whose reward is a perplexity-derived quantity computed during training, on the batches the run is already taking. That is what makes it online: no separate proxy runs, no held-out sweep, negligible added cost.
The signal it uses is learnability rather than quality. A domain whose loss is falling fastest is where the next tokens buy the most, and that changes as training proceeds — a domain worth over-sampling early can be exhausted later. Perplexity is a proxy for that, not for whether the text is good.
Where the title overstates it
“Quality assessment” is not what a perplexity number is. A high-perplexity document can be noise or it can be rare and valuable, and the two are indistinguishable to the model’s own uncertainty — which is why perplexity-based filtering, as distinct from perplexity-based weighting, is a blunt instrument that preferentially discards unusual text.
The defensible version is the one the source supports: use perplexity to decide how much to sample from a domain, continuously, and not to decide what to discard. Those are different practices with different failure modes, and the title names the weaker one.