Soft Data Duplication Blows Out Gradients: Root Cause of GRPO Training Collapse

ivan_bezdomny · x · 2026-08-02

The author diagnosed the root cause of the previous GRPO training collapse in Gemma. GRPO generates multiple outputs per input to compute gradients. When consecutive batches contain many of the same rare tokens (e.g., a specific bill name), it creates gradients in the exact same direction. This double-step backprop, amplified by momentum, blows out the model weights irrecoverably.

Related event: Developer Reveals LLM Fine-Tuning Fragility and Training Crashes(3 posts)→

Original post →

More from Research

Research channel →