CurDKV: Value-Guided KV Cache Compression

• CurDKV is a value-guided key-value cache compression method formulated as a CUR-decomposition problem to preserve the end-to-end attention output in autoregressive LLM inference.
• It computes leverage scores via Gaussian projections on keys and values, enabling selective token retention that achieves up to 9.6% higher accuracy and 40% lower generation latency.
• In a distinct differential-geometric context, CurDKV denotes a curvature-derived invariant Killing structure for Petrov type D spacetimes, illustrating its dual usage in research.

CurDKV denotes a value-guided key-value cache compression method for autoregressive LLM inference that tries to preserve the actual attention output rather than merely preserving tokens that receive high attention. In “Value-Guided KV Compression for LLMs via Approximated CUR Decomposition,” the method is formulated as a CUR-decomposition problem over attention-related matrices, with token retention driven by leverage scores that explicitly incorporate value vectors and therefore target preservation of $\mathrm{softmax}(QK^\top)V$ under compression (Sengupta et al., 18 Sep 2025). In a distinct and unrelated usage in the differential-geometric literature, the same label is applied to a curvature-derived invariant Killing structure in Petrov type D vacuum spacetimes, centered on the complex Killing vector $Z = w^{-1/3}\,\delta U$ (Ferrando et al., 2013).

1. Long-context inference setting and problem formulation

CurDKV arises in the setting of autoregressive LLM inference, where the KV cache grows linearly with context length and long-context inference becomes memory- and latency-heavy because every previously generated token contributes a key/value pair. At each decoding step, attention remains

$\mathrm{softmax}(QK^\top)V,$

so any compression policy must decide which cached key-value rows to keep without significantly changing this output (Sengupta et al., 18 Sep 2025).

The central observation is that most prior cache-eviction schemes optimize the wrong proxy. Existing methods predominantly rank tokens by query-key attention scores or related statistics, assuming that attention intensity correlates with semantic importance. CurDKV argues that this only approximates the matrix $QK^\top$, not the end result $\mathrm{softmax}(QK^\top)V$. A token can receive low attention yet still carry a semantically crucial value vector, while a highly attended token may have a less important contribution after multiplication by $V$. The paper therefore identifies “eviction loss” as the real quantity of interest and treats preservation of the attention output as the compression objective rather than preservation of high-attention tokens.

This framing places CurDKV in contrast with attention-score-based eviction heuristics such as SnapKV, H2O, AdaSnapKV, ChunkKV, Streaming LLM, and Knorm. The distinction is not merely algorithmic; it changes the optimization target from a surrogate over attention weights to end-to-end attention reconstruction.

2. Output-preservation objective and theoretical motivation

A recurrent misconception in KV-cache compression is that high attention implies low eviction sensitivity. CurDKV directly disputes this by emphasizing that attention weights tell which keys are queried, not whether removing the associated value vectors will preserve the final output. The paper reports empirically that high attention does not reliably imply low post-eviction reconstruction error, and its theory formalizes why value-aware selection is more principled (Sengupta et al., 18 Sep 2025).

Let

$$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$

where $K',V'$ are compressed versions obtained by zeroing out rows. The paper gives the perturbation bound

$$\|A-A'\|_F \le \sqrt{n}\|V-V'\|_F + 2\sqrt{n}\|V'\|_F.$$

Its intended implication is explicit: the value matrix $V$ is directly responsible for output fidelity, while the effect of $Z = w^{-1/3}\,\delta U$0 is mediated by softmax and thus partially absorbed. At high compression, the retained value approximation $Z = w^{-1/3}\,\delta U$1 becomes the critical term.

On this basis, the paper’s main theoretical claim is that approximating attention scores does not guarantee preservation of the attention output. CurDKV instead states the compression problem as minimizing the end-to-end reconstruction loss of the full attention pipeline. The ablations reinforce the same point: key-only scoring degrades more sharply at high compression, whereas value-based or combined key-value leverage is more robust. Random Gaussian projection helps somewhat, especially under severe compression, but the more important design choice is the explicit use of value information.

3. CUR-decomposition formulation and leverage-score selection

CurDKV casts KV selection as a CUR-decomposition problem. For a matrix $Z = w^{-1/3}\,\delta U$2, CUR approximates it as

$Z = w^{-1/3}\,\delta U$3

where $Z = w^{-1/3}\,\delta U$4 is a subset of columns of $Z = w^{-1/3}\,\delta U$5, $Z = w^{-1/3}\,\delta U$6 is a subset of rows, and $Z = w^{-1/3}\,\delta U$7 is a low-rank linking matrix. The attraction of CUR in this setting is that it keeps actual data rows and columns rather than arbitrary latent singular vectors, making it operationally suitable for cache eviction (Sengupta et al., 18 Sep 2025).

The method uses leverage scores derived from the dominant subspace. For a matrix $Z = w^{-1/3}\,\delta U$8 with rank $Z = w^{-1/3}\,\delta U$9, the row and column leverage scores are

$\mathrm{softmax}(QK^\top)V,$0

Since exact SVD is expensive, CurDKV uses a Gaussian projection approximation. For each head or GQA group, it projects the key and value matrices with a random matrix $\mathrm{softmax}(QK^\top)V,$1 whose entries satisfy

$\mathrm{softmax}(QK^\top)V,$2

If $\mathrm{softmax}(QK^\top)V,$3 and $\mathrm{softmax}(QK^\top)V,$4 are the projected matrices, the algorithm computes row-wise norms as approximate leverage values: $\mathrm{softmax}(QK^\top)V,$5 then combines them multiplicatively: $\mathrm{softmax}(QK^\top)V,$6

Selection proceeds by retaining the top-$\mathrm{softmax}(QK^\top)V,$7 indices by the normalized combined score, while always preserving the first $\mathrm{softmax}(QK^\top)V,$8 tokens as attention sinks. The resulting subsets become $\mathrm{softmax}(QK^\top)V,$9 and $QK^\top$0, which replace the full cache during inference. The paper’s interpretation is that these selected rows are the most representative of the dominant attention-output subspace, with the value component explicitly included because it determines the output vectors passed to the next layer.

4. Empirical performance on LongBench and Ruler

Empirically, CurDKV is reported to be strong across long-context benchmarks on LLaMA-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3. The abstract reports up to a $QK^\top$1 higher accuracy than SnapKV and ChunkKV under aggressive compression budgets on LLaMA and Mistral, together with up to $QK^\top$2 lower generation latency at high compression (Sengupta et al., 18 Sep 2025).

On LongBench, the detailed results state that CurDKV beats SnapKV by about $QK^\top$3 at one compression setting on LLaMA and remains ahead at harsher budgets as well. On the average LongBench score at $QK^\top$4 budget, CurDKV reaches about $QK^\top$5 on LLaMA versus $QK^\top$6 for SnapKV; on Mistral it gets $QK^\top$7 versus $QK^\top$8 for SnapKV. At $QK^\top$9 compression, CurDKV still holds up better than many baselines, and AdaCurDKV often pushes performance further by reallocating budget across heads.

On Ruler’s needle-in-a-haystack tasks, CurDKV is particularly strong under moderate compression. At $\mathrm{softmax}(QK^\top)V$0 compression, it attains an average of $\mathrm{softmax}(QK^\top)V$1 on LLaMA and $\mathrm{softmax}(QK^\top)V$2 on Mistral, outperforming attention-based and norm-based alternatives. At $\mathrm{softmax}(QK^\top)V$3 compression, performance drops, but CurDKV and AdaCurDKV remain significantly stronger than most baselines. The paper also notes that ChunkKV can sometimes be competitive on some highly structured retrieval cases, but CurDKV is described as more robust overall across diverse tasks and models.

Setting	CurDKV	Comparator from the same report
LongBench average, LLaMA, $\mathrm{softmax}(QK^\top)V$4 budget	$\mathrm{softmax}(QK^\top)V$5	SnapKV: $\mathrm{softmax}(QK^\top)V$6
LongBench average, Mistral, $\mathrm{softmax}(QK^\top)V$7 budget	$\mathrm{softmax}(QK^\top)V$8	SnapKV: $\mathrm{softmax}(QK^\top)V$9
Ruler average, LLaMA, $V$0 compression	$V$1	attention-based and norm-based alternatives outperformed
Ruler average, Mistral, $V$2 compression	$V$3	attention-based and norm-based alternatives outperformed

These results position CurDKV as a compression method whose gains are concentrated in aggressive or moderate compression regimes where preservation of the actual attention output becomes more consequential than preservation of raw attention mass.

5. Runtime characteristics, systems compatibility, and deployment regime

CurDKV reduces KV-cache size roughly linearly with compression, and generation latency drops substantially because fewer cached tokens must be attended to. The tradeoff is an increase in prefill latency due to leverage-score computation and token selection, but the paper characterizes this overhead as modest and amortized for long contexts (Sengupta et al., 18 Sep 2025).

For long sequences such as $V$4 tokens, the reported timing trend is that generation time can fall from about $V$5–$V$6 seconds to under $V$7–$V$8 seconds at high compression, while prefill time rises from around $V$9 seconds to about $$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$0–$$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$1 seconds in the high-compression regime. This produces the paper’s stated practical speed-accuracy tradeoff: higher prefill cost in exchange for substantially cheaper generation under long-context workloads.

A notable systems property is compatibility with FlashAttention. Because CurDKV does not require explicit access to attention weights, it remains compatible with fused attention implementations that avoid materializing the full attention matrix. Methods that rely on observed attention scores can break under FlashAttention, whereas CurDKV avoids that dependency. The method is also designed for Grouped Query Attention, operating per GQA group with shared key/value matrices across heads, so compression can be applied without redundantly replicating cache entries and without sacrificing GQA’s efficiency.

The paper identifies the most suitable deployment regime as long-context inference under tight memory budgets, especially when large quality drops are unacceptable, when compression must be question-agnostic, and when FlashAttention or GQA compatibility matters.

6. Distinct geometric usage in type D relativity

In an unrelated differential-geometric context, the label “CurDKV” is used for the curvature-derived invariant complex Killing structure that organizes the intrinsic classification of Petrov type D vacuum spacetimes. There the central object is the invariant complex Killing vector

$$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$2

constructed from the Weyl eigenvalue $$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$3 and the canonical self-dual bivector $$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$4 (Ferrando et al., 2013).

Within that framework, the self-dual Weyl tensor has canonical form

$$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$5

and the Bianchi identities are written as

$$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$6

The vector $$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$7 is introduced from the integrability conditions of these identities and is used to encode first covariant derivatives of the Riemann tensor, express second derivatives through the Killing 2-form $$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$8, and distinguish intrinsic subclasses such as Kerr-NUT metrics, regular generalized C-metrics, and strict Ehlers-Kundt C-metrics. The paper proves that two covariant derivatives of the Riemann tensor always suffice for Cartan-Karlhede classification, with the exceptional strict C-metrics requiring, in addition to first-order data, a constant second-order invariant (Ferrando et al., 2013).

The same geometric structure appears in the conformal initial-data characterization of vacuum Petrov type D spacetimes. There, the conformal data

$$A=\mathrm{softmax}(QK^\top)V,\qquad A'=\mathrm{softmax}(QK'^\top)V',$$9

must encode both the algebraic Petrov-D degeneracy of the rescaled Weyl curvature and the existence of the distinguished Killing structure associated with type D solutions (García-Parrado, 2018). In that literature, the role played by the curvature-derived Killing field is analogous to its role in the intrinsic classification: it is the invariant structure that organizes the passage from curvature conditions to spacetime characterization.

Because these two usages belong to separate research domains, “CurDKV” is best treated as a homonymous term: in contemporary LLM systems work it denotes value-guided KV compression via approximated CUR decomposition, whereas in the type D vacuum literature it designates a curvature-derived invariant Killing structure.