Across the full 0.5–4 bpd K-only sweep, allocator choice mattered little from about 2 bpd upward, while KV-COBRA—especially its KL ordering on LongBench—degraded more gracefully at low rates than the evaluated baselines.

bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: MSE · datasetOrTask: mean(Wikitext-2, PTB, C4)

bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: KL · datasetOrTask: mean(Wikitext-2, PTB, C4)

bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: MSE · datasetOrTask: mean(Wikitext-2, PTB, C4)

bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · datasetOrTask: mean(Wikitext-2, PTB, C4)

bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: MSE · datasetOrTask: mean(Wikitext-2, PTB, C4)

bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: KL · datasetOrTask: mean(Wikitext-2, PTB, C4)

bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: MSE · datasetOrTask: mean(five tasks)

bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: KL · datasetOrTask: mean(five tasks)

bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: MSE · datasetOrTask: mean(five tasks)

bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · datasetOrTask: mean(five tasks)

bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: MSE · datasetOrTask: mean(five tasks)

bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: KL · datasetOrTask: mean(five tasks)