At 1 bpd on LLaMA-70B, Qwen-72B, and Mixtral-8x7B, KV-COBRA variants remained much closer to FP16 than the collapsing integer-grid baselines on perplexity, and KV-COBRA_KL was the best compressed method for PPL and LongBench in every family.

At 1 bpd on LLaMA-70B, Qwen-72B, and Mixtral-8x7B, KV-COBRA variants remained much closer to FP16 than the collapsing integer-grid baselines on perplexity, and KV-COBRA_KL was the best compressed method for PPL and LongBench in every family. · CiteArk

Measurement results

0/68 assessed

0 supported0 disputed or conflicting68 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
bpd: 16 · side: K-only · model: LLaMA-3.1-70B · method: FP16	5.38 perplexity	—	Not assessed
bpd: 16 · side: K-only · model: LLaMA-3.1-70B · method: FP16	60.4%	—	Not assessed
bpd: 16 · side: K-only · model: LLaMA-3.1-70B · method: FP16	14.3 F1	—	Not assessed
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_KL	6.97 perplexity	—	Not assessed
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_KL	58.4%	—	Not assessed
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_KL	13.5 F1	—	Not assessed
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_MSE	9.5 perplexity	—	Not assessed
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_MSE	58.5%	—	Not assessed
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_MSE	10.5 F1	—	Not assessed
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KQ-SVD	9.79 perplexity	—	Not assessed
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KQ-SVD	57%	—	Not assessed
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KQ-SVD	11.2 F1	—	Not assessed

At 1 bpd on LLaMA-70B, Qwen-72B, and Mixtral-8x7B, KV-COBRA variants remained much closer to FP16 than the collapsing integer-grid baselines on perplexity, and KV-COBRA_KL was the best compressed method for PPL and LongBench in every family.

Reported and observed measurements

Assessments (0)

Runs (0)

Measurement results