Read from LIT-382 — ARXIV-2310.01798.
What it establishes
A definition and a measurement against it. Intrinsic self-correction is the model revising “based solely on its inherent capabilities, without the crutch of external feedback”. Under that definition, on reasoning, models struggle — and performance sometimes degrades after the revision step.
The mechanism is not subtle and is the reason the result should not have been surprising. A model with no external signal has nothing to tell a correct answer from an incorrect one. A revision pass is therefore about as likely to overwrite a right answer as a wrong one, and if the initial answers are mostly right, the expected effect of revision is negative.
Why the definition outlives the result
Because a great deal of the literature it was written against is measuring something else. If an oracle says “this is wrong, try again”, or a verifier filters, or a second model critiques, the loop contains information the model did not have — and the gain is attributable to that information rather than to any capacity for self-correction. The paper’s contribution is drawing that line and insisting results declare which side they are on.
That makes it a lens rather than a verdict, which is also why the title’s “yet” is doing real work. It is a claim about models as they were, under one specific protocol.
What it does to the record
No practice is sourced here, deliberately — the finding is that an intervention does not work, and the useful instruction it supports belongs to the paper that supplies the alternative (LIT-383).
Two things it is worth holding for. First, as the baseline that makes SCoRe’s numbers mean something: a gain measured against prompted self-correction is only interesting because prompted self-correction is known not to work. Second, as a reading instrument — any self-correction claim in this corpus should be checked for whether external feedback was in the loop before its number is believed.