Source
Chen, Huang, Golowich, Malladi, Block, Ash, Krishnamurthy and Foster (2025), LIT-474 — read as NOTE-223.
What it explains
| practice | what it says to do | what this says it is |
|---|---|---|
| SOTA-284 | select the post-training checkpoint on coverage, not validation cross-entropy | two metrics that agree until missing mass and sequence length pull them apart, which is exactly where the interesting checkpoints are |
| SOTA-210 | report pass@k as well as pass@1 | pass@k is an estimator of the quantity that actually gates post-training, and pass@1 is not |
The account
The cleanest statement is a two-outcome model. Let π*(1) = ε for small ε,
and fit by maximum likelihood on n samples. With constant probability the
dataset contains no 1 at all, the estimator assigns it zero, and expected KL
divergence is infinite.
Nothing has gone wrong with the learner. It saw no evidence of an event and
did not predict it. But KL’s penalty for an under-assigned outcome is
log(π*/π̂), which diverges as π̂ → 0, so a learner that is behaving
correctly is charged without bound.
The coverage profile — how much mass the model puts where the data distribution puts it, weighted by how many Best-of-N draws you get — has no such term. Either the empirical frequency concentrated, or the missing mass is small enough to write off against the mass you did capture. The worst it can charge is what is actually missing.
That asymmetry scales. Over a sequence of H tokens, the opportunities
for a well-generalizing model to miss some tail response multiply, so
sequence-level KL accumulates them: for autoregressive linear models, KL grows
linearly in H for any proper estimator on some data distribution, and the
same linear growth appears empirically. Coverage does not — at convergence it
shows no H dependence at all.
So the two metrics do not merely differ by a constant. They diverge in the direction of longer sequences, which is the direction everything has been moving.
And next-token prediction optimizes the better one anyway. The coverage
bound for the maximum-likelihood estimator splits into a fine-grained term
that reads the class’s complexity at a small scale, scaled by 1/N and
carrying no sequence-length or density-ratio dependence, and a coarse-grained
term that reads complexity at a very large scale and captures the missing
mass. Both are tight. The 1/N scaling means coverage converges faster the
further into the tail you look.
That is the coverage principle: the objective is implicitly doing the right thing, and the number on the screen is measuring something else — something that inherits a penalty the objective’s own generalization behaviour does not.
Every step above is proved in LIT-474: that the coverage profile is necessary
and sufficient for Best-of-N success, the lower bound making sequence-level KL
linear in H for autoregressive linear models, and the coverage bound for
the maximum-likelihood estimator with both terms tight. Its one experiment, a
synthetic graph-reasoning task, shows KL at convergence growing linearly in
H while coverage shows no H dependence, and KL improving through training
while coverage degrades.
Why Proposed
Not because the theorems are in doubt. They are proved, with matching lower bounds for the parts that matter.
Because the distance between the theorems’ setting and a language model is large and the paper says so. The formulation splits data into prompts and responses, which the authors call “closer in spirit to supervised fine-tuning”; the main theorem assumes realizability; the optimizer results are for autoregressive linear models where the loss is convex in the parameterization; the empirical content is one synthetic graph-reasoning task.
A proof about a model class the record’s practices are not about is an
account offered, not an account established — which is what Proposed means
here, and the promotion condition names the measurement that would change it.
What this does not say
It does not say cross-entropy is useless. It upper-bounds coverage, it is estimable, and coverage is not — a footnote in the source concedes that neither KL nor coverage is directly observable. The claim is about when the bound goes slack: with missing mass, and increasingly with sequence length.
It does not prove anything about Adam. The source proves that globally normalized SGD removes the sequence-length dependence, and then says “somewhat speculatively” that Adam-like updates may inherit the benefit — noting that Adam normalizes per coordinate, a difference the authors themselves call important for deep models. This is the sentence most likely to be quoted as a result. It is not one, and SOTA-001 gains nothing from it.
It does not establish anything about RL, only about Best-of-N. The source states plainly that the minimal conditions for RL are unknown, and the bridge is a cited empirical finding that BoN predicts post-RL performance.
And the version it analyses is the conservative one. For reasoning tasks the object that matters is answer-level coverage, which is bounded by the sequence-level profile and can be strictly smaller. The account is stated about a quantity that is pessimistic relative to what a reasoning practitioner needs.