KV Compression Nabs PPL Gains on Qwen3.6-Plus

Interesting signal, but not enough to declare a real quality win yet. With only three seeds and tiny negative deltas, this could still be run-to-run variance, but it is exactly the kind of result that warrants a deeper sweep.

• –All three runs moved in the same direction, which makes the result more interesting than a one-off fluke.
• –The effect size is small enough that context length, sampling variance, or eval noise could explain it.
• –If it holds up at scale, the compression method may act like regularization by suppressing low-value attention directions.
• –This is most relevant in long-context inference, where KV cache pressure is highest and small quality shifts can matter operationally.
• –The next useful check is a broader seed sweep across multiple contexts and tasks, not just perplexity on one model.