O96a/weak-supervision-reasoning

Weak Supervision Reasoning Explorer

Interactive demo exploring when LLMs can learn to reason with weak supervision, based on paper 2604.18574.

Hypothesis: Models that generalize under weak supervision exhibit a prolonged pre-saturation phase during which training reward and downstream performance climb together, while rapid saturation indicates memorization.

Key Findings from Paper

• —Reward Saturation Dynamics: Models that generalize show prolonged pre-saturation
• —Reasoning Faithfulness: Intermediate steps logically supporting final answers predict generalization
• —SFT is Critical: Supervised fine-tuning on explicit reasoning traces enables weak supervision generalization

Features

• —Visualize reward saturation curves
• —Compare reasoning faithfulness across models
• —Interactive weak supervision scenarios