S₀ Tuning

• Optimizes the initial hidden state (S₀) of recurrent layers while keeping the model weights fixed.
• Injects the learned states before each forward pass.
• Saves and loads the state tensors separately from the base model.
• HumanEval score on Qwen3.5-4B: 48.8% for the base model, 72.2% with S₀ tuning (10 seeds).