"Weak-to-Strong Generalization Through the Data-Centric Lens"

Data points with both simple and complex patterns enable weak models to teach strong ones.

Overlap density reveals why smaller models can effectively train larger ones.

This paper introduces a data-centric mechanism called "overlap density" that explains how weaker models can effectively supervise stronger models. The mechanism identifies data points containing both easy patterns (learnable by weak models) and hard patterns (only learnable by strong models), enabling effective weak-to-strong generalization.

-----

https://arxiv.org/abs/2412.03881

🤔 Original Problem:

→ Current research focuses heavily on algorithmic improvements for weak-to-strong generalization but lacks understanding of what data characteristics enable this phenomenon.

-----

🔍 Solution in this Paper:

→ The paper proposes "overlap density" as a key mechanism where data points contain both easy and hard patterns.

→ Easy patterns are learnable by weak models while hard patterns are only accessible to strong models.

→ The researchers develop an overlap detection algorithm to identify such data points.

→ They introduce a UCB-based data selection strategy to maximize overlap density when choosing between multiple data sources.

-----

💡 Key Insights:

→ Weak-to-strong generalization success depends on the presence of overlapping data points

→ Higher overlap density leads to better generalization performance

→ Data source selection can be optimized to maximize overlap density

-----

📊 Results:

→ Empirically validated across LLM experiments showing strong correlation between overlap density and generalization performance

→ Achieved theoretical regret bound of O(√(K log T/t)) for data source selection

→ Demonstrated three clear regimes: low overlap (poor generalization), medium overlap (comparable to weak model), high overlap (approaching strong model performance)