A fine-tuning method that trains on a model's own correct answers, with gold or paraphrased answers otherwise, improves task accuracy and cuts generalization loss versus standard SFT.
A notion of twins
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
abstract
Given a combinatorial structure, a ``twin'' is a pair of disjoint substructures which are isomorphic (or look the same in some sense). In recent years, there have been many problems about finding large twins in various combinatorial structures. For example, given a graph $G$, one can ask what is the largest $s$ such that there exist disjoint subsets $I,J\subset V(G)$ on $s$ vertices, such that the induced subgraphs $G[I],G[J]$ are isomorphic. We are motivated by two different problems of finding twins in two kinds of ordered objects (strings and permutations). We introduce a new variant of ``twin problem'' which generalizes both of these. By considering this generalization, we are able to improve some bounds obtained by Dudek, Grytczuk, and Ruci\'nski, and give a negative answer to a conjecture of theirs.
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Selective Self-to-Supervised Fine-Tuning for Generalization in Large Language Models
A fine-tuning method that trains on a model's own correct answers, with gold or paraphrased answers otherwise, improves task accuracy and cuts generalization loss versus standard SFT.