KV-COBRA: KV Cache Compression via Co-Optimized Bit-Rank Allocation · CiteArk

ArkGraph

浏览 ArkGraph，选择本次要执行的步骤。

KV-COBRA: KV Cache Compression via Co-Optimized Bit-Rank Allocation

公开

作者：Sihyeon Ha、Jaeho Lee、Yo-Seb Jeon

arXiv 2026 机器学习 (cs.LG)人工智能

复现结果

全部结论 ↗

尚无复现结果

论文中的结论可在「研究结论」中查看。

完整研究主张目录

0 / 25 条结论已通过验证

其余仍在验证中

关键指标论文报告值 → 复现观测值 · 点击行展开详情

In the joint K+V Wikitext-2 sweep, V-heavy splits generally produced lower perplexity; KV-COBRA_KL supplied the best cell in seven of nine model-by-total-budget blocks, while V_bpd=2 caused severe LLaMA and Mistral degradation and KQ-SVD diverged on Qwen cells with K_bpd=1.整组结论：等待复现0/111 项指标获得支持 · 0 项已有评估111项测量结果展开结果表

实验方案已生成，还没有运行记录

测量结果

已评估 0/111 项

获得支持 0有争议或冲突 0尚未评估 111

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
指标与条件	论文报告	最近可用实测	证据判断
kBpd: 1 · vBpd: 4 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 5	17.35 perplexity	—	尚未评估
kBpd: 2 · vBpd: 3 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 5	38.84 perplexity	—	尚未评估
kBpd: 3 · vBpd: 2 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 5	732.1 perplexity	—	尚未评估
kBpd: 1 · vBpd: 5 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6	15.79 perplexity	—	尚未评估
kBpd: 2 · vBpd: 4 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6	17.88 perplexity	—	尚未评估
kBpd: 3 · vBpd: 3 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6	19.17 perplexity	—	尚未评估
kBpd: 4 · vBpd: 2 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6	550.4 perplexity	—	尚未评估
kBpd: 1 · vBpd: 6 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7	15.21 perplexity	—	尚未评估
kBpd: 2 · vBpd: 5 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7	15.48 perplexity	—	尚未评估
kBpd: 3 · vBpd: 4 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7	6.69 perplexity	—	尚未评估
kBpd: 4 · vBpd: 3 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7	17.26 perplexity	—	尚未评估
kBpd: 5 · vBpd: 2 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7	505.7 perplexity	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
	2.5 fold	—	尚未评估
	5.9 fold	—	尚未评估
model: LLaMA-3.1-8B	7.72 perplexity	—	尚未评估
model: LLaMA-3.1-8B	4.03 perplexity	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
method: KIVI · headDimension: 128	1 bits per dimension	—	尚未评估
method: KVQuant · headDimension: 128	0.16 bits per dimension	—	尚未评估
method: GEAR · headDimension: 128	1 bits per dimension	—	尚未评估
method: TurboQuant · headDimension: 128	0 bits per dimension	—	尚未评估
method: SVDq · headDimension: 128	0 bits per dimension	—	尚未评估
method: KQ-SVD · headDimension: 128	0 bits per dimension	—	尚未评估
method: KV-COBRA · headDimension: 128	0 bits per dimension	—	尚未评估
contextLength: 32768 · residualTokens: 128	0.07 bits per dimension	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
bitWidth: 2 · queryHeads: 32 · retainedRank: 16 · headDimension: 128 · sequenceLength: 32000	1.83 fold	—	尚未评估
bitWidth: 4 · queryHeads: 32 · retainedRank: 16 · headDimension: 128 · sequenceLength: 32000	1.84 fold	—	尚未评估
bitWidth: 2 · queryHeads: 32 · retainedRank: 32 · headDimension: 128 · sequenceLength: 32000	1.77 fold	—	尚未评估
bitWidth: 4 · queryHeads: 32 · retainedRank: 32 · headDimension: 128 · sequenceLength: 32000	1.73 fold	—	尚未评估
bitWidth: 2 · queryHeads: 32 · retainedRank: 64 · headDimension: 128 · sequenceLength: 32000	1.4 fold	—	尚未评估
bitWidth: 4 · queryHeads: 32 · retainedRank: 64 · headDimension: 128 · sequenceLength: 32000	1.31 fold	—	尚未评估
bitWidth: 2 · queryHeads: 32 · retainedRank: 128 · headDimension: 128 · sequenceLength: 32000	0.44 fold	—	尚未评估
bitWidth: 4 · queryHeads: 32 · retainedRank: 128 · headDimension: 128 · sequenceLength: 32000	0.32 fold	—	尚未评估
bitWidth: 2 · queryHeads: 32 · retainedRank: 16 · headDimension: 128 · sequenceLength: 64000	1.79 fold	—	尚未评估
bitWidth: 4 · queryHeads: 32 · retainedRank: 16 · headDimension: 128 · sequenceLength: 64000	1.85 fold	—	尚未评估
bitWidth: 2 · queryHeads: 32 · retainedRank: 32 · headDimension: 128 · sequenceLength: 64000	1.82 fold	—	尚未评估
bitWidth: 4 · queryHeads: 32 · retainedRank: 32 · headDimension: 128 · sequenceLength: 64000	1.74 fold	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: MSE · datasetOrTask: mean(Wikitext-2, PTB, C4)	31.17 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: KL · datasetOrTask: mean(Wikitext-2, PTB, C4)	37.31 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: MSE · datasetOrTask: mean(Wikitext-2, PTB, C4)	36.5 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · datasetOrTask: mean(Wikitext-2, PTB, C4)	36.24 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: MSE · datasetOrTask: mean(Wikitext-2, PTB, C4)	16.33 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: KL · datasetOrTask: mean(Wikitext-2, PTB, C4)	17.17 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: MSE · datasetOrTask: mean(five tasks)	48.8%	—	尚未评估
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: KL · datasetOrTask: mean(five tasks)	48.5%	—	尚未评估
bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: MSE · datasetOrTask: mean(five tasks)	53.4%	—	尚未评估
bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · datasetOrTask: mean(five tasks)	52.5%	—	尚未评估
bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: MSE · datasetOrTask: mean(five tasks)	48.7%	—	尚未评估
bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: KL · datasetOrTask: mean(five tasks)	48.5%	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: MSE · datasetOrTask: Wikitext-2	16.29 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: KL · datasetOrTask: Wikitext-2	17.6 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: MSE · datasetOrTask: Wikitext-2	9.05 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · datasetOrTask: Wikitext-2	8.08 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: MSE · datasetOrTask: Wikitext-2	10.06 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: KL · datasetOrTask: Wikitext-2	10.37 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: MSE · datasetOrTask: PTB	30.75 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: KL · datasetOrTask: PTB	34.74 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: MSE · datasetOrTask: PTB	81.88 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · datasetOrTask: PTB	86.2 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: MSE · datasetOrTask: PTB	19.5 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: Qwen2.5-7B-Instruct · objective: KL · datasetOrTask: PTB	20.37 perplexity	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
heads: 256 · model: LLaMA-3.1-8B	100 fold	—	尚未评估
	1,000 fold	—	尚未评估
	16 dimensions	—	尚未评估
	100 dimensions	—	尚未评估
	2 bits	—	尚未评估
	5 bits	—	尚未评估
	2 fold	—	尚未评估
	8 fold	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
kBpd: 1 · vBpd: 4 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 5	8.64 F1	—	尚未评估
kBpd: 2 · vBpd: 3 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 5	6.75 F1	—	尚未评估
kBpd: 3 · vBpd: 2 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 5	3.25 F1	—	尚未评估
kBpd: 1 · vBpd: 5 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6	9.86 F1	—	尚未评估
kBpd: 2 · vBpd: 4 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6	8.31 F1	—	尚未评估
kBpd: 3 · vBpd: 3 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6	9.71 F1	—	尚未评估
kBpd: 4 · vBpd: 2 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6	3.73 F1	—	尚未评估
kBpd: 1 · vBpd: 6 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7	8.08 F1	—	尚未评估
kBpd: 2 · vBpd: 5 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7	8.89 F1	—	尚未评估
kBpd: 3 · vBpd: 4 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7	13.59 F1	—	尚未评估
kBpd: 4 · vBpd: 3 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7	8.98 F1	—	尚未评估
kBpd: 5 · vBpd: 2 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7	3.66 F1	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · objective: MSE · datasetOrTask: Wikitext-2	16.45 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · objective: KL · datasetOrTask: Wikitext-2	13.27 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: Mistral-7B-v0.3 · objective: MSE · datasetOrTask: Wikitext-2	7.47 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · datasetOrTask: Wikitext-2	5.65 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: Qwen2.5-7B-Instruct · objective: MSE · datasetOrTask: Wikitext-2	8.14 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: Qwen2.5-7B-Instruct · objective: KL · datasetOrTask: Wikitext-2	8.17 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · objective: MSE · datasetOrTask: PTB	25.01 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · objective: KL · datasetOrTask: PTB	21.12 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: Mistral-7B-v0.3 · objective: MSE · datasetOrTask: PTB	59.38 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · datasetOrTask: PTB	32.86 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: Qwen2.5-7B-Instruct · objective: MSE · datasetOrTask: PTB	14.79 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: Qwen2.5-7B-Instruct · objective: KL · datasetOrTask: PTB	15.07 perplexity	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
bpd: 16 · side: K-only · model: LLaMA-3.1-70B · method: FP16	5.38 perplexity	—	尚未评估
bpd: 16 · side: K-only · model: LLaMA-3.1-70B · method: FP16	60.4%	—	尚未评估
bpd: 16 · side: K-only · model: LLaMA-3.1-70B · method: FP16	14.3 F1	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_KL	6.97 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_KL	58.4%	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_KL	13.5 F1	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_MSE	9.5 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_MSE	58.5%	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KV-COBRA_MSE	10.5 F1	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KQ-SVD	9.79 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KQ-SVD	57%	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-70B · method: KQ-SVD	11.2 F1	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
task: NarrativeQA · model: LLaMA-3.1-8B · method: KV-COBRA_KL · totalBpd: 5	1 bits per dimension	—	尚未评估
task: NarrativeQA · model: LLaMA-3.1-8B · method: KV-COBRA_KL · totalBpd: 5	4 bits per dimension	—	尚未评估
task: NarrativeQA · model: LLaMA-3.1-8B · method: KV-COBRA_KL · totalBpd: 6	1 bits per dimension	—	尚未评估
task: NarrativeQA · model: LLaMA-3.1-8B · method: KV-COBRA_KL · totalBpd: 6	5 bits per dimension	—	尚未评估
task: NarrativeQA · model: LLaMA-3.1-8B · method: KV-COBRA_KL · totalBpd: 7	1 bits per dimension	—	尚未评估
task: NarrativeQA · model: LLaMA-3.1-8B · method: KV-COBRA_KL · totalBpd: 7	6 bits per dimension	—	尚未评估
task: NarrativeQA · model: Mistral-7B-v0.3 · method: KV-COBRA_KL · totalBpd: 5	1 bits per dimension	—	尚未评估
task: NarrativeQA · model: Mistral-7B-v0.3 · method: KV-COBRA_KL · totalBpd: 5	4 bits per dimension	—	尚未评估
task: NarrativeQA · model: Mistral-7B-v0.3 · method: KV-COBRA_KL · totalBpd: 6	2 bits per dimension	—	尚未评估
task: NarrativeQA · model: Mistral-7B-v0.3 · method: KV-COBRA_KL · totalBpd: 6	4 bits per dimension	—	尚未评估
task: NarrativeQA · model: Mistral-7B-v0.3 · method: KV-COBRA_KL · totalBpd: 7	3 bits per dimension	—	尚未评估
task: NarrativeQA · model: Mistral-7B-v0.3 · method: KV-COBRA_KL · totalBpd: 7	4 bits per dimension	—	尚未评估

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
指标与条件	论文报告	最近可用实测	证据判断
models: all three 7–8B models · bitWidth: 4 · aggregation: median across heads	1 ratio	—	尚未评估
models: all three 7–8B models · bitWidth: 4	31%	—	尚未评估
models: all three 7–8B models · bitWidth: 3 · aggregation: median across heads	1.1 ratio	—	尚未评估
models: all three 7–8B models · bitWidth: 3	33%	—	尚未评估
models: all three 7–8B models · bitWidth: 2 · aggregation: median across heads	1.7 ratio	—	尚未评估
models: all three 7–8B models · bitWidth: 2	41%	—	尚未评估
models: all three 7–8B models · bitWidth: 1 · aggregation: median across heads	6.9 ratio	—	尚未评估
models: all three 7–8B models · bitWidth: 1	36%	—	尚未评估
bpd: 16 · side: K-only · model: LLaMA-3.1-8B · configuration: FP16	5.71 perplexity	—	尚未评估
bpd: 16 · side: K-only · model: Mistral-7B-v0.3 · configuration: FP16	4.77 perplexity	—	尚未评估
bpd: 16 · side: K-only · model: Qwen2.5-7B-Instruct · configuration: FP16	6.07 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · configuration: water-filling + Bennett estimate (shipped)	16.33 perplexity	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
	5 rounds	—	尚未评估
	1%	—	尚未评估
scope: paper wording: across all three 7–8B models	200 milliseconds	—	尚未评估
rankStep: 2 · headDimension: 128	32 candidates	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
	0.46 ratio	—	尚未评估
side: V	128 dimensions	—	尚未评估
	4 bits per dimension	—	尚未评估
	5 bits per dimension	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · method: KV-COBRA_KL · calibration: reference n=32 WikiText-2	10.8 F1	—	尚未评估
bpd: 1 · side: K-only · model: Mistral-7B-v0.3 · method: KV-COBRA_KL · calibration: reference n=32 WikiText-2	10.6 F1	—	尚未评估
bpd: 1 · side: K-only · model: Qwen2.5-7B-Instruct · method: KV-COBRA_KL · calibration: reference n=32 WikiText-2	9.4 F1	—	尚未评估
calibration: disjoint WikiText-2 draw	0.992 correlation	—	尚未评估
calibration: disjoint WikiText-2 draw	1 correlation	—	尚未评估
calibration: disjoint WikiText-2 draw	97%	—	尚未评估
calibration: disjoint WikiText-2 draw	100%	—	尚未评估
calibration: disjoint WikiText-2 draw	0.45 dimensions	—	尚未评估
calibration: disjoint WikiText-2 draw	1.17 dimensions	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · method: KV-COBRA_KL · calibration: disjoint WikiText-2 draw	9.6 F1	—	尚未评估
bpd: 1 · side: K-only · model: Mistral-7B-v0.3 · method: KV-COBRA_KL · calibration: disjoint WikiText-2 draw	10 F1	—	尚未评估
bpd: 1 · side: K-only · model: Qwen2.5-7B-Instruct · method: KV-COBRA_KL · calibration: disjoint WikiText-2 draw	9.4 F1	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: MSE · contextLength: 4096	20.9%	—	尚未评估
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · objective: KL · contextLength: 4096	44.3%	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · objective: MSE · contextLength: 4096	67.1%	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · objective: KL · contextLength: 4096	76%	—	尚未评估
bpd: 2 · side: K-only · model: LLaMA-3.1-8B · objective: KL · contextLength: 4096	92.5%	—	尚未评估
model: LLaMA-3.1-8B · configuration: FP16 · contextLength: 4096	93.5%	—	尚未评估
bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: MSE · contextLength: 4096	36.7%	—	尚未评估
bpd: 0.5 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · contextLength: 4096	52%	—	尚未评估
bpd: 1 · side: K-only · model: Mistral-7B-v0.3 · objective: MSE · contextLength: 4096	69.1%	—	尚未评估
bpd: 1 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · contextLength: 4096	79.3%	—	尚未评估
bpd: 2 · side: K-only · model: Mistral-7B-v0.3 · objective: KL · contextLength: 4096	90.6%	—	尚未评估
model: Mistral-7B-v0.3 · configuration: FP16 · contextLength: 4096	91.3%	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
kBpd: 1 · vBpd: 4 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 5 · contextLength: 4096	67.33%	—	尚未评估
kBpd: 2 · vBpd: 3 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 5 · contextLength: 4096	87.61%	—	尚未评估
kBpd: 3 · vBpd: 2 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 5 · contextLength: 4096	92.07%	—	尚未评估
kBpd: 1 · vBpd: 5 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6 · contextLength: 4096	68.65%	—	尚未评估
kBpd: 2 · vBpd: 4 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6 · contextLength: 4096	88.33%	—	尚未评估
kBpd: 3 · vBpd: 3 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6 · contextLength: 4096	93.71%	—	尚未评估
kBpd: 4 · vBpd: 2 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 6 · contextLength: 4096	93.11%	—	尚未评估
kBpd: 1 · vBpd: 6 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7 · contextLength: 4096	68.6%	—	尚未评估
kBpd: 2 · vBpd: 5 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7 · contextLength: 4096	88.24%	—	尚未评估
kBpd: 3 · vBpd: 4 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7 · contextLength: 4096	93.72%	—	尚未评估
kBpd: 4 · vBpd: 3 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7 · contextLength: 4096	93.88%	—	尚未评估
kBpd: 5 · vBpd: 2 · model: LLaMA-3.1-8B · sides: K+V · method: KV-COBRA_MSE · totalBpd: 7 · contextLength: 4096	93.43%	—	尚未评估

指标与条件	论文报告	最近可用实测	证据判断
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · solver: damped proportional update	15.56 perplexity	—	尚未评估
bpd: 0.5 · side: K-only · model: LLaMA-3.1-8B · solver: exact DP over Bennett surrogate	17.02 perplexity	—	尚未评估
bpd: 0.5 · model: LLaMA-3.1-8B	25 fold	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · solver: damped proportional update	15.55 perplexity	—	尚未评估
bpd: 1 · side: K-only · model: LLaMA-3.1-8B · solver: exact DP over Bennett surrogate	15.71 perplexity	—	尚未评估
bpd: 1 · model: LLaMA-3.1-8B	40 fold	—	尚未评估
bpd: 2 · side: K-only · model: LLaMA-3.1-8B · solver: damped proportional update	6.78 perplexity	—	尚未评估
bpd: 2 · side: K-only · model: LLaMA-3.1-8B · solver: exact DP over Bennett surrogate	7.38 perplexity	—	尚未评估
bpd: 2 · model: LLaMA-3.1-8B	49 fold	—	尚未评估
bpd: 3 · side: K-only · model: LLaMA-3.1-8B · solver: damped proportional update	5.98 perplexity	—	尚未评估
bpd: 3 · side: K-only · model: LLaMA-3.1-8B · solver: exact DP over Bennett surrogate	5.93 perplexity	—	尚未评估
bpd: 3 · model: LLaMA-3.1-8B	50 fold	—	尚未评估

论文报告	最近可用实测	证据判断
0.3 excess kurtosis	—	尚未评估
0.2 excess kurtosis	—	尚未评估
1.1 excess kurtosis	—	尚未评估
2.5 sigma	—	尚未评估

ArkGraph

KV-COBRA: KV Cache Compression via Co-Optimized Bit-Rank Allocation

研究结论

尚无复现结果

关键指标论文报告值 → 复现观测值 · 点击行展开详情

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

测量结果

定性结论