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

Recent activity

—Analysis made new progress
—Analysis environment started
—Analysis made new progress
—Research inventory and experiment plan prepared
—Waiting for experiment selection

ArkGraph

Explore ArkGraph and select the steps to run.

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

Public

Authors:Sihyeon Ha, Jaeho Lee, Yo-Seb Jeon

arXiv 2026 Machine Learning (cs.LG)Artificial Intelligence

More

Reproduction

Start with coverage and measured results; open a run only when you need evidence or technical details.

RunMatchRepeat

0/21

claims supported by evidence

0

Supported

0

Challenged or mixed

0

Contradicted

0

Inconclusive

21

Not assessed

Run history

Each row is one recorded execution. Commands, logs, hashes, and signatures are available in its details.

No reproduction runs yet

Runs will appear here as the paper's experiment plans are executed.

Claim–experiment reproduction matrix

See each experiment's execution state, blocker, recovery action, and evidence destination while keeping operations separate from scientific conclusions. There are also 4 claims with no independent reproduction scheduled in this plan.

A successful execution does not by itself validate a paper claim

Target state says whether the platform completed the work. The scientific conclusion is determined only by immutable evidence and Assessments. Resource shortages and platform failures are never presented as scientific contradictions.

This task predates target-level records. The states below are conservative projections; no attempts, resource decisions, or evidence links are invented.

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.

CiteArk reconstructionc063 plans0 runs

Scientific conclusionNot assessed

LLaMA-3.1-8B: joint K+V Wikitext-2 perplexity sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: joint K+V Wikitext-2 perplexity sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: joint K+V Wikitext-2 perplexity sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Equal spectral distortion was not equally harmful across depth: the most sensitive layer quartile caused much more perplexity damage per surrogate-distortion unit than the least sensitive, with different depth profiles by model, and full-depth errors compounded superadditively on LLaMA.

CiteArk reconstructionc171 plan0 runs

Scientific conclusionNot assessed

Measure layer-quartile sensitivity and cross-depth compounding

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Every tested compressor produced lower perplexity when keys were compressed pre-RoPE rather than post-RoPE at the canonical bit rates; the paper attributes this to position-stationary pre-RoPE covariance and the loss of a useful shared low-rank basis after position-dependent rotations.

Information insufficientc121 plan0 runs

Scientific conclusionNot assessed

Compare pre-RoPE and post-RoPE key compression

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

The effective-bpd accounting assigns method-specific per-token metadata overheads: +1.0 bpd to KIVI and GEAR, +0.16 to KVQuant, and zero to the rotation/SVD methods; fixed bases are excluded as amortized at T=32,768.

CiteArk reconstructionc101 plan0 runs

Scientific conclusionNot assessed

Recompute effective-bpd metadata overheads

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

The CUDA-graph-captured per-layer attention primitive was faster than FP16-SDPA through rank 64 at all tested lengths and bit widths, but slower at full rank 128.

CiteArk reconstructionc141 plan0 runs

Scientific conclusionNot assessed

Benchmark the CUDA-graph attention primitive on RTX 6000 Ada

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

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.

CiteArk reconstructionc0519 plans0 runs

Scientific conclusionNot assessed

Mistral-7B-v0.3: full K-only 0.5–4-bpd perplexity sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: full K-only 0.5–4-bpd perplexity sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: K-only 0.5/1-bpd perplexity

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: K-only 0.5/1-bpd LongBench F1

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: full K-only 0.5–4-bpd zero-shot accuracy sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: K-only 0.5/1-bpd zero-shot accuracy

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: full K-only 0.5–4-bpd LongBench F1 sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: K-only 0.5/1-bpd LongBench F1

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: K-only 0.5/1-bpd LongBench F1

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: full K-only 0.5–4-bpd LongBench F1 sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: K-only 0.5/1-bpd zero-shot accuracy

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: full K-only 0.5–4-bpd zero-shot accuracy sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: full K-only 0.5–4-bpd LongBench F1 sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: full K-only 0.5–4-bpd zero-shot accuracy sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: K-only 0.5/1-bpd perplexity

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: K-only 0.5/1-bpd perplexity

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: full K-only 0.5–4-bpd perplexity sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Compare pre-RoPE and post-RoPE key compression

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: K-only 0.5/1-bpd zero-shot accuracy

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

At 0.5 bpd in K-only mode, the paper reports per-dataset perplexity, per-task zero-shot accuracy, and per-task LongBench F1 for both KV-COBRA objectives on all three 7–8B models.

CiteArk reconstructionc039 plans0 runs

Scientific conclusionNot assessed

Mistral-7B-v0.3: K-only 0.5/1-bpd perplexity

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: K-only 0.5/1-bpd LongBench F1

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: K-only 0.5/1-bpd zero-shot accuracy

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: K-only 0.5/1-bpd LongBench F1

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: K-only 0.5/1-bpd LongBench F1

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: K-only 0.5/1-bpd zero-shot accuracy

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: K-only 0.5/1-bpd perplexity

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: K-only 0.5/1-bpd perplexity

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: K-only 0.5/1-bpd zero-shot accuracy

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Attention heads are strongly heterogeneous in spectral distortion and in their jointly selected rank/bit operating points; C2 consequently assigns non-flat head budgets and substantially flattens residual distortion.

CiteArk reconstructionc011 plan0 runs

Scientific conclusionNot assessed

Reconstruct allocation heterogeneity and Hadamard diagnostics

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

The joint K+V LongBench sweep reports all K/V splits for KV-COBRA_MSE, KV-COBRA_KL, and KQ-SVD; the suite mean generally favors providing more bits to V, while the V_bpd=2 cells are especially weak on LLaMA and Mistral.

CiteArk reconstructionc073 plans0 runs

Scientific conclusionNot assessed

Qwen2.5-7B-Instruct: joint K+V LongBench F1 sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: joint K+V LongBench F1 sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: joint K+V LongBench F1 sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

At 1 bpd in K-only mode, the paper reports per-dataset perplexity, per-task zero-shot accuracy, and per-task LongBench F1 for both KV-COBRA objectives on all three 7–8B models.

CiteArk reconstructionc049 plans0 runs

Scientific conclusionNot assessed

Mistral-7B-v0.3: K-only 0.5/1-bpd perplexity

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: K-only 0.5/1-bpd LongBench F1

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: K-only 0.5/1-bpd zero-shot accuracy

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: K-only 0.5/1-bpd LongBench F1

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: K-only 0.5/1-bpd LongBench F1

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: K-only 0.5/1-bpd zero-shot accuracy

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: K-only 0.5/1-bpd perplexity

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: K-only 0.5/1-bpd perplexity

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: K-only 0.5/1-bpd zero-shot accuracy

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

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 reconstructionc183 plans0 runs

Scientific conclusionNot assessed

Mixtral-8x7B: 1-bpd scale and family check

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-70B: 1-bpd scale and family check

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-72B: 1-bpd scale and family check

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

For KV-COBRA_KL on individual LongBench tasks, the best K/V split was V-heavy in 40 of 45 model–task–budget cells; the five exceptions were QASPER or MultiFieldQA-en cells that became balanced or K-heavy at higher budgets.

CiteArk reconstructionc153 plans0 runs

Scientific conclusionNot assessed

Qwen2.5-7B-Instruct: joint K+V LongBench F1 sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: joint K+V LongBench F1 sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

LLaMA-3.1-8B: joint K+V LongBench F1 sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

KV-COBRA calibrates a per-head SVD, jointly enumerates rank and integer bit width in C1, performs five damped C2 redistribution rounds, optionally reorders directions by query–key variance product for the KL variant, and fuses the Hadamard rotation into the stored basis; inference then uses the frozen per-head choices.

CiteArk reconstructionc211 plan0 runs

Scientific conclusionNot assessed

Reconstruct allocation heterogeneity and Hadamard diagnostics

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

V has a much flatter spectrum than K, so C1 selects full rank for almost every V head at 2 bpd; C2 and KL reordering add no measurable V-side gain, and V-only compression remains much worse than K-only below 3 bpd before approaching FP16 around 4–5 bpd.

CiteArk reconstructionc111 plan0 runs

Scientific conclusionNot assessed

Measure V-side spectral flatness and compression limits

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

C1+C2 allocations were stable under disjoint corpora and fourfold smaller or larger calibration sets, with downstream LongBench scores in a narrow band; a Python-source calibration also changed HumanEval pass@1 by at most 1.6 points.

CiteArk reconstructionc162 plans0 runs

Scientific conclusionNot assessed

Test calibration robustness across corpora and sample counts

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Probe Python calibration transfer to HumanEval

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

On RULER-4k retrieval, attention-KL reordering outperformed the MSE ordering at 0.5 and 1 bpd for every 7–8B model, with its largest objective separation among the paper's benchmarks; at 2 bpd it came within 0.7–3.6 points of FP16.

CiteArk reconstructionc193 plans0 runs

Scientific conclusionNot assessed

LLaMA-3.1-8B: K-only RULER-4k objective comparison

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: K-only RULER-4k objective comparison

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: K-only RULER-4k objective comparison

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

For joint K+V compression on RULER-4k, the best splits reverse the PPL/LongBench pattern and are balanced to K-heavy; K_bpd=1 loses roughly 15–30 accuracy points, and KQ-SVD is near zero on Qwen across almost every split.

CiteArk reconstructionc203 plans0 runs

Scientific conclusionNot assessed

LLaMA-3.1-8B: joint K+V RULER-4k accuracy sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Mistral-7B-v0.3: joint K+V RULER-4k accuracy sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Qwen2.5-7B-Instruct: joint K+V RULER-4k accuracy sweep

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Across the multi-seed K-only cells in Table 3 and Figure 4, run-to-run dispersion was generally small but reached its maximum at 0.5 bpd.

CiteArk reconstructionc131 plan0 runs

Scientific conclusionNot assessed

Aggregate run-to-run dispersion over all K-only cells

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

The Bennett shape is accurate near 3–4 bits but increasingly underestimates deployed-quantizer distortion at low bits; replacing it with measured per-head distortion makes exact DP improve over the shipped allocator on all three models at 1 bpd.

CiteArk reconstructionc091 plan0 runs

Scientific conclusionNot assessed

Test Bennett finite-rate error and measured-distortion DP

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Exact dynamic programming optimizes the integer Bennett-distortion surrogate but does not consistently improve real-model perplexity; across 15 cells it is far slower than the damped allocator and sometimes helps, ties, or hurts.

CiteArk reconstructionc081 plan0 runs

Scientific conclusionNot assessed

Compare damped C2 allocation with exact dynamic programming

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Hadamard rotation made retained-channel marginals substantially closer to Gaussian on the calibration data, although mild tail deviation remained and the paper treats Gaussianity as an empirical working assumption rather than a guarantee.

CiteArk reconstructionc021 plan0 runs

Scientific conclusionNot assessed

Reconstruct allocation heterogeneity and Hadamard diagnostics

An experiment plan exists, but the current processing task has no matching execution target.

Not in current task

Community reproductions

Executed locally and uploaded by users. The platform verifies file signatures; conclusions come from the uploaded runs.

No community runs yet.