Research map / Learning theory and algorithms / Kernels and regression theory
Weak-to-Strong Generalization Through the Data-Centric Lens
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes overlap density, a data-centric mechanism explaining when weak-to-strong generalization occurs, with an overlap detection algorithm and a UCB-based data source selection method, supported by theory and experiments on LLM fine-tuning, weak supervision, and synthetic data.
Problem
Understanding which properties of data enable weak-to-strong generalization, and how to acquire data that promotes it, when a weak model pseudolabels data used to train a strong model.
Approach
Weak-to-strong generalization is driven by overlap points containing both easy patterns (learnable by weak model) and hard patterns (only strong model learns); a detection algorithm (confidence change-point plus inner-product/cosine overlap scores) identifies such points, and a UCB bandit selects the data source with highest overlap density.
Key results
- In LLM and weak supervision settings, strong model performance improves as overlap proportion increases, with three regimes (low, medium, high overlap) observed.
- UCB-based overlap-maximizing data selection yields better weak-to-strong generalization than random sampling on Amazon Polarity and DREAM (averages over 20 seeds).
- Theory gives a regret bound of O(sqrt(K log T / t)) for the UCB data selection algorithm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the overlap proportion in the training data | Llama3 8B weak-to-strong model supervised by Qwen1.5 0.5B · 19 LLM datasets | accuracy | Helped |
| Increase the overlap proportion in the training data | 4-layer MLP weak-to-strong model supervised by Snorkel · 9 WRENCH datasets | accuracy | Helped |
| Select data sources using UCB-based overlap maximization instead of random sampling | Weak-to-strong training · Amazon Polarity and DREAM | generalization | Helped |
| Increase the overlap ratio | Synthetic weak-to-strong model · Hard-only test points | accuracy | Helped |
| Increase the overlap ratio | Synthetic weak-to-strong model · Easy-only test points | accuracy | Mixed |
| Increase the overlap ratio | Synthetic weak-to-strong model · Overlap test points | accuracy | Helped |
| Select data sources using the proposed algorithm instead of random sampling | Synthetic weak-to-strong model · Synthetic data selection | generalization | Helped |
Surprising findings
- If overlap density is very low or overlap detection is very noisy, weak-to-strong generalization does not occur, and the weak-to-strong model can underperform the weak model.
Limitations
- Theory uses a simplified easy/hard feature decomposition assumption (A1-A3) and Gaussian mixture model, which may not hold in real data.
- Overlap detection is noisy and depends on thresholds from change point detection; performance drops when overlap is rare or detection is noisy.
- Only two difficulty levels are modeled; more complex multi-level patterns are left for future work.
Experimental setup
Datasets: WRENCH weak supervision benchmark (9 datasets) 19 NLP datasets from Burns et al. weak-to-strong setup Amazon Polarity DREAM synthetic Gaussian mixture data
Benchmarks: WRENCH 19 datasets from Burns et al. weak-to-strong setup Amazon Polarity DREAM synthetic Gaussian mixture
Metrics: accuracy cumulative regret
Baselines: random sampling oracle data source selection
Builds on: weak-to-strong generalization (Burns et al.) expansion-based pseudolabel correction (Wei & Zou) UCB bandit Snorkel WRENCH
Closest papers by meaning
- Weak-to-Strong Generalization Even in Random Feature Networks, Provably (ICML 2025 · similarity 0.88)
- From Linear to Nonlinear: Provable Weak-to-Strong Generalization through Feature Learning (NeurIPS 2025 · similarity 0.87)
- Theoretical Analysis of Weak-to-Strong Generalization (NeurIPS 2024 · similarity 0.87)
- On the Mechanisms of Weak-to-Strong Generalization: A Theoretical Perspective (NeurIPS 2025 · similarity 0.87)
- Relating Misfit to Gain in Weak-to-Strong Generalization Beyond the Squared Loss (ICML 2025 · similarity 0.87)
- Quantifying the Gain in Weak-to-Strong Generalization (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.