Source
Albalak et al. (2023), LIT-117 — ARXIV-2312.02406.
The temperature is real, and it is inside the bandit
This practice’s body used to argue that a temperature and a bandit are different kinds of thing — “a temperature reshapes a distribution somebody already chose; a bandit discovers the distribution”. Reading LIT-117 for #114 shows that is wrong about this bandit.
Exp3’s policy is a Gibbs distribution — a softmax over importance-weighted rewards — mixed with a uniform distribution for exploration:
πₜ(Dᵢ) = (1 − K·ℰₜ) · exp(ℰₜ₋₁·R̂ᵢ) / Σⱼ exp(ℰₜ₋₁·R̂ⱼ) + ℰₜ
The exploration rate ℰₜ multiplies the reward inside the exponent, which is
exactly the role an inverse temperature plays, and it varies over training.
So “dynamic temperature scaling for mixing” is an imprecise name for something
ODM genuinely has.
Why superseded rather than rejected
Because it names a component, not a practice. The temperature is one term in a policy whose other parts — the domain arms, the training-loss reward, the importance weighting — are what make the method work, and a recommendation to tune the temperature alone would be advice about a knob detached from the machine it belongs to.
SOTA-103 now describes the whole method, which is what the two of these were between them. The record’s older body already suspected this, saying the two “are one practice split in the import rather than two”. That was right; the reason given for it was not.
The correction is the point
Two of the three retirements in this cluster stand on the paper containing a
string zero times. This one does not, and it would have been retired on the
same evidence if the reading had stopped at a keyword search. The word
temperature appears nowhere in LIT-117; the mechanism it names is in
equation form on the page.