Draft-KV addresses a flaw in evaluating between language models: accuracy can improve because of the trained interface rather than because the receiver uses the sender’s message. Across five method-dataset pairs, replacing messages with ones from unrelated questions changed accuracy by at most 0.60 points, despite system gains as large as 15.44 points.
The method sends key-value states generated while a larger model drafts an answer. A small trainable interface projects these states into side memory for a frozen receiver, with progressive training and protection against harmful mismatched messages. On MMLU-Redux, a frozen Qwen2.5-0.5B-Instruct receiver rose from 37.45% alone to 78.04% with a Qwen3-8B sharer, but scored 36.40% with reassigned messages. This intervention indicates that the gain depends on question-specific content rather than interface adaptation alone.