parameter-interpolation

GRIP introduces a reward-guided parameter interpolation framework to merge reasoning and instruction models, improving accuracy-efficiency trade-off for LLM reasoning without retraining.