Deeper Transformers Are Forgetting What They Learned. MoDA Is the Fix.

Deep transformers forget early-layer signals. MoDA fixes this with depth KV caching — better performance, only 3.7% extra compute. Here's how it works