Preference tuning on weak-versus-weaker response pairs can improve a strong LLM as much as tuning on strong supervision, via the relative quality delta.
Standard tail bounds due to Laurent & Massart (2000) give that Pr x(t,i) 2 ≥ 4 √ d ≤ e−4d
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
The Delta Learning Hypothesis: Preference Tuning on Weak Data can Yield Strong Gains
Preference tuning on weak-versus-weaker response pairs can improve a strong LLM as much as tuning on strong supervision, via the relative quality delta.