A large model can use a demonstration to produce a useful teaching response for itself. A small model may misunderstand the same demonstration, creating weak supervision and forgetting older skills.
SynthSelf, posted October 9, uses a minimally larger same-family aligner to rewrite supervision for smaller models, with rejection sampling and off-policy training. Its evaluations separately examine acquisition and retention. Read the research paper.
A correct target can still be a poor teaching target
The paper associates failure with shifts away from the student's original response distribution. A demonstration can be sound yet demand a style of computation the student cannot reproduce reliably.
The same-family aligner is intended to bridge that mismatch. Its assistance should be counted when describing the method as self-learning; the small model is not improving unaided.
For edge deployments, test both the newly requested skill and old tasks after each update. A specialised benchmark gain can conceal a less useful everyday model.
The reported comparisons support the approach at the tested scales. They do not establish that distribution matching alone guarantees retention. The broader insight is pedagogical: choosing a powerful teacher is insufficient. The teaching signal must also be usable by the particular student expected to learn from it.
Questions, news tips or corrections? Contact [email protected].
