CommitKV: Lifecycle-Aware KV Cache Compression via Commit Transitions for Multi-Turn Agents

# CommitKV: Lifecycle-Aware KV Cache Compression via Commit Transitions for Multi-Turn Agents
Source: [https://arxiv.org/html/2608.07855](https://arxiv.org/html/2608.07855)
Weizhong Huang1,†Jinchao Zhang2,\*Xiawu Zheng1,\* 1Xiamen University2WeChat AI, Tencent Inc\., China \*Corresponding authors

###### Abstract

Multi\-turn Reasoning\-and\-Acting \(ReAct\) agents accumulate growing trajectories of reasoning, tool calls, and observations\. Their key\-value \(KV\) caches grow accordingly, increasing memory use and attention cost during model inference\. Existing KV cache compression methods reduce these costs by evicting states with low attention scores\. However, low attention in the current turn does not imply future irrelevance, as temporarily inactive information may become important later\. Snapshot\-based eviction methods therefore do not explicitly distinguish temporarily dormant information from information that appears to have completed its role\. In this paper, we presentCommitKV, which identifies KV lifecycles through commit transitions\. Specifically,CommitKVfirst divides completed agent events into token pages and compares each eligible page’s deletion effect before a tool\-call commit and after the commit’s returned observation has been incorporated\. Based on these paired measurements,CommitKVdistinguishes dormant pages from high\-to\-low completion candidates\. It then applies a greedy joint test, accepting candidates for retirement only when their combined post\-commit effect remains bounded\. Finally, at a later compression checkpoint, accepted pages are excluded, a bounded set of pages awaiting post\-commit measurement is protected, and the remaining KV states are retained within the cache budget using the same token indices for keys, values, and absolute positions\. These mechanisms ensure thatCommitKVcan distinguish dormant information from information that has completed its observed role and can be safely removed\. Experiments on various benchmarks show thatCommitKVreduces agent memory use, accelerates end\-to\-end inference, and achieves higher accuracy than existing KV cache compression methods\.

22footnotetext:This work was done when Weizhong Huang was an intern at WeChat AI\.## 1Introduction

Large language model \(LLM\)\-based agents\(Yanget al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib81); Guoet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib84); Abdinet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib166); Agarwalet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib171); Lyuet al\.,[2026b](https://arxiv.org/html/2608.07855#bib.bib53),[a](https://arxiv.org/html/2608.07855#bib.bib3)\)solve complex tasks through reasoning and tool use\(Guoet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib76); Chenget al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib74)\)\. In the widely used ReAct paradigm, an agent alternates between reasoning, tool calls, and returned observations\(Yaoet al\.,[2023](https://arxiv.org/html/2608.07855#bib.bib22)\)\. As the trajectory grows, so does its key\-value \(KV\) cache\. Cross\-turn reuse avoids repeatedly prefilling the full history, but the persistent cache still increases memory and attention costs\(Gaoet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib205); Zhenget al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib107)\)\. KV cache compression mitigates this problem by retaining cached states under a fixed budget\(Shiet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib68); WEIet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib193)\)\.

Existing methods typically estimate the importance of cached tokens from attention\-based or query\-aware scores\(Tanget al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib71); Caiet al\.,[2025b](https://arxiv.org/html/2608.07855#bib.bib93); Fenget al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib94); Genget al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib95)\)\. For example, SnapKV\(Liet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib92)\)aggregates attention scores from a prompt\-end observation window, R\-KV\(Caiet al\.,[2025a](https://arxiv.org/html/2608.07855#bib.bib101)\)balances attention\-based importance and key redundancy, and TriAttention\(Maoet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib182)\)estimates importance using pre\-RoPE\(Suet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib123)\)query–key geometry\.

Despite their different scoring functions, these methods follow the same principle: they estimate token importance from the current inference state and evict low\-scoring tokens\. This snapshot\-based strategy overlooks how information evolves across agent turns\. As illustrated in Figure[1](https://arxiv.org/html/2608.07855#S1.F1), an earlier observation and a completed tool\-call page may both receive low scores at the current checkpoint\. However, the earlier observation may only be temporarily dormant and become useful again, whereas the tool\-call page may have completed its role after the returned observation becomes available\. Therefore, a low current score alone cannot determine whether a page is safe to remove, potentially causing the premature eviction of useful states\.

![Refer to caption](https://arxiv.org/html/2608.07855v1/x1.png)Figure 1:Why snapshot scores are insufficient for KV\-cache compression\. At checkpointjj, the earlier observation pageE1E\_\{1\}and tool\-call pageE2E\_\{2\}may both have low current importance, althoughE1E\_\{1\}contains Paris, which is needed later, whileE2E\_\{2\}may have completed its role after the weather observation returns\.CommitKVcompares their pre\- and post\-commit deletion effects: low\-to\-lowE1E\_\{1\}is not evicted by the lifecycle rule, whereas high\-to\-lowE2E\_\{2\}becomes a completion candidate and is marked for retirement only if it passes joint validation; its KV states are removed at a later checkpoint\.To resolve this ambiguity, we presentCommitKV, a lifecycle\-aware KV cache compression method for multi\-turn ReAct agents\. First,CommitKVpartitions completed tool calls and returned observations into event pages and treats the end of each generated tool call as a*commit*measurement boundary\. It measures the deletion effect of the same page near the commit and after the returned observation is incorporated into the next turn, forming a*commit transition*\. Next, it combines absolute deletion effects with percentile ranks to identify high\-to\-low completion candidates while avoiding the premature removal of low\-to\-low dormant pages\. Finally,CommitKVjointly validates completion candidates and temporarily protects pages awaiting post\-commit measurement\. At each compression checkpoint, it excludes retired pages, retains the protected pending pages, and selects the remaining KV states under the cache budget, applying the same token indices to the keys, values, and absolute positions\. In this way,CommitKVremoves information that has completed its observed role while preserving information that may become useful in future turns\.

To evaluateCommitKV, we conduct extensive experiments on six LLMs across eight benchmarks and multiple KV\-cache budgets\.CommitKVoutperforms the strongest compressed baseline by up to 22\.24 percentage points in average accuracy, while achieving up to 5\.00×\\timespeak KV\-memory savings and a 5\.62×\\timesend\-to\-end speedup\. These results demonstrate thatCommitKVeffectively preserves agent performance while substantially reducing memory and latency\.

## 2Related Work

#### LLM agents\.

LLM\(Xieet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib35),[2026](https://arxiv.org/html/2608.07855#bib.bib2)\)agents interleave reasoning, tool use, and environment feedback across turns\(Yaoet al\.,[2023](https://arxiv.org/html/2608.07855#bib.bib22); Mastermanet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib204); Zhanget al\.,[2025b](https://arxiv.org/html/2608.07855#bib.bib75)\)\. Their growing trajectories increase context length and persistent KV\-cache costs\. Recent methods manage agent states using phase\- and intent\-aware importance, region\-specific decay, episode\-level eviction, context pruning, or cache recovery\(Liuet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib192); Liet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib195); Matam and Kim,[2026](https://arxiv.org/html/2608.07855#bib.bib194); Kariyappa and Suh,[2026](https://arxiv.org/html/2608.07855#bib.bib153); Kimet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib197); Haoet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib203); Rusliet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib201)\)\.CommitKVinstead treats a tool\-call commit as a lifecycle boundary, retiring completed pages while preserving dormant ones\.

#### KV cache compression\.

KV cache compression reduces long\-context inference costs by evicting or compacting cached states\. Eviction criteria include attention\(Zhanget al\.,[2023](https://arxiv.org/html/2608.07855#bib.bib65); Liuet al\.,[2023](https://arxiv.org/html/2608.07855#bib.bib66)\), query relevance\(Tanget al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib71)\), adaptive policies\(Fenget al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib94); Geet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib72)\), anchor\-direction projection\(Genget al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib95)\), key similarity\(Parket al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib96)\), value awareness\(Changet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib97)\), and structured pages or chunks\(Chitty\-Venkataet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib196); Huet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib199)\)\. SnapKV uses prompt\-end attention\(Liet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib92)\), R\-KV balances importance and redundancy\(Caiet al\.,[2025a](https://arxiv.org/html/2608.07855#bib.bib101)\), and TriAttention uses pre\-RoPE geometry\(Maoet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib182)\)\. Other techniques\(Maet al\.,[2023a](https://arxiv.org/html/2608.07855#bib.bib32),[2024b](https://arxiv.org/html/2608.07855#bib.bib70); Zhenget al\.,[2021](https://arxiv.org/html/2608.07855#bib.bib133); Zhanget al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib41); Huanget al\.,[2025b](https://arxiv.org/html/2608.07855#bib.bib181),[a](https://arxiv.org/html/2608.07855#bib.bib146)\)include cache merging\(Zhanget al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib86); Liuet al\.,[2024a](https://arxiv.org/html/2608.07855#bib.bib85)\), quantization\(Liuet al\.,[2024b](https://arxiv.org/html/2608.07855#bib.bib52); Hooperet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib103); Heet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib102); Zhanget al\.,[2025a](https://arxiv.org/html/2608.07855#bib.bib67); Maet al\.,[2024a](https://arxiv.org/html/2608.07855#bib.bib19),[2023b](https://arxiv.org/html/2608.07855#bib.bib13)\), hybrid error correction\(Kanget al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib104)\), low\-rank projection\(Changet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib87)\), and low\-rank keys with value\-cache offloading\(Sunet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib88)\)and others\(Maet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib17)\)\. Most eviction methods score a single inference state, conflating dormant and completed low\-influence pages\.CommitKVinstead compares deletion effects before a commit and after the returned observation, distinguishing low\-to\-low dormant pages from high\-to\-low completion candidates and jointly validating the latter before retirement\.

![Refer to caption](https://arxiv.org/html/2608.07855v1/x2.png)Figure 2:Overview ofCommitKV\. \(1\) For each tool\-call commit, the deletion effect of the same eligible event page is measured in the pre\- and post\-commit query windows\. \(2\) The paired deletion effects and percentile ranks determine the page’s lifecycle state; only a high\-to\-low page becomes a completion candidate\. \(3\) A greedy joint test constructs the retirement set𝒟c\\mathcal\{D\}\_\{c\}\. At checkpointjj, tokens in𝒟c\\mathcal\{D\}\_\{c\}are excluded, pending indices𝒩j\\mathcal\{N\}\_\{j\}are protected withinBNB\_\{N\}, and the retained\-token index setℐj\\mathcal\{I\}\_\{j\}is selected under budgetBBand applied consistently to keys, values, and absolute positions\.

## 3Methodology

### 3\.1Problem Formulation

#### Multi\-turn ReAct agent\.

We consider a tool\-using multi\-turn ReAct agent\(Yaoet al\.,[2023](https://arxiv.org/html/2608.07855#bib.bib22)\)that answers a query𝒒\\boldsymbol\{q\}through repeated reasoning and tool use\. At turntt, the agent receives the interaction history𝒙t\\boldsymbol\{x\}\_\{t\}and generates reasoning𝒛t\\boldsymbol\{z\}\_\{t\}followed by a decision𝒂t\\boldsymbol\{a\}\_\{t\}, which is either a tool call or a final answer\. Define

𝒙t\\displaystyle\\boldsymbol\{x\}\_\{t\}=\(𝒒,𝒛1,𝒂1,𝒐1,…,𝒛t−1,𝒂t−1,𝒐t−1\),\\displaystyle=\(\\boldsymbol\{q\},\\boldsymbol\{z\}\_\{1\},\\boldsymbol\{a\}\_\{1\},\\boldsymbol\{o\}\_\{1\},\\ldots,\\boldsymbol\{z\}\_\{t\-1\},\\boldsymbol\{a\}\_\{t\-1\},\\boldsymbol\{o\}\_\{t\-1\}\),\(1\)𝒚t\\displaystyle\\boldsymbol\{y\}\_\{t\}=\(𝒙t,𝒛t,𝒂t\)\.\\displaystyle=\(\\boldsymbol\{x\}\_\{t\},\\boldsymbol\{z\}\_\{t\},\\boldsymbol\{a\}\_\{t\}\)\.If𝒂t\\boldsymbol\{a\}\_\{t\}invokes a tool, the returned observation𝒐t\\boldsymbol\{o\}\_\{t\}is appended to the history for the next turn\. Otherwise,𝒂t\\boldsymbol\{a\}\_\{t\}is the final answer and terminates the interaction\.

#### KV reuse across turns\.

To avoid repeatedly prefilling the complete history, the server retains and reuses KV states across turns\(Gaoet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib205); Zhenget al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib107)\)\. At the first turn, it prefills𝒙1=𝒒\\boldsymbol\{x\}\_\{1\}=\\boldsymbol\{q\}\. For each subsequent turnt\>1t\>1,

𝒙t=\(𝒚t−1,𝒐t−1\),\\boldsymbol\{x\}\_\{t\}=\(\\boldsymbol\{y\}\_\{t\-1\},\\boldsymbol\{o\}\_\{t\-1\}\),\(2\)where the KV states of𝒚t−1\\boldsymbol\{y\}\_\{t\-1\}are already cached\. The server therefore prefills only the new observation𝒐t−1\\boldsymbol\{o\}\_\{t\-1\}and appends the KV states of the generated𝒛t\\boldsymbol\{z\}\_\{t\}and𝒂t\\boldsymbol\{a\}\_\{t\}during decoding\.

#### Token\-level KV cache compression\.

KV reuse causes the cache to grow with the interaction history\. Consider anMM\-layer LLM\(Vaswaniet al\.,[2017](https://arxiv.org/html/2608.07855#bib.bib16)\)withHHKV heads of widthdd\. After turntt, letLt=\|𝒚t\|L\_\{t\}=\|\\boldsymbol\{y\}\_\{t\}\|\. The KV tensors are

𝑲t,𝑽t∈ℝM×H×Lt×d,\\boldsymbol\{K\}\_\{t\},\\boldsymbol\{V\}\_\{t\}\\in\\mathbb\{R\}^\{M\\times H\\times L\_\{t\}\\times d\},\(3\)where the four axes represent the layer, KV head, token number, and head width\.

FullKV retains allLtL\_\{t\}tokens, whereas token\-level compression reduces the token axis\. We call each runtime boundary at which tokens may be removed a*compression checkpoint*\. Suppose thatSjS\_\{j\}tokens with absolute positions𝒑j=\(pj,1,…,pj,Sj\)\\boldsymbol\{p\}\_\{j\}=\(p\_\{j,1\},\\ldots,p\_\{j,S\_\{j\}\}\)are resident before checkpointjj\. The policy selectsℐj⊆\{1,…,Sj\}\\mathcal\{I\}\_\{j\}\\subseteq\\\{1,\\ldots,S\_\{j\}\\\}and applies it to the keys, values, and positions:

𝑲~j\\displaystyle\\widetilde\{\\boldsymbol\{K\}\}\_\{j\}=𝑲j​\[:,:,ℐj,:\],\\displaystyle=\\boldsymbol\{K\}\_\{j\}\[:,:,\\mathcal\{I\}\_\{j\},:\],\\qquad𝑽~j\\displaystyle\\widetilde\{\\boldsymbol\{V\}\}\_\{j\}=𝑽j​\[:,:,ℐj,:\],\\displaystyle=\\boldsymbol\{V\}\_\{j\}\[:,:,\\mathcal\{I\}\_\{j\},:\],\(4\)𝒑~j\\displaystyle\\widetilde\{\\boldsymbol\{p\}\}\_\{j\}=𝒑j​\[ℐj\],\\displaystyle=\\boldsymbol\{p\}\_\{j\}\[\\mathcal\{I\}\_\{j\}\],\\qquad\|ℐj\|\\displaystyle\|\\mathcal\{I\}\_\{j\}\|≤B\.\\displaystyle\\leq B\.Here,BBis the token budget, and𝒑~j\\widetilde\{\\boldsymbol\{p\}\}\_\{j\}denotes the absolute positions of the retained tokens after compression\. Applying the same index set preserves the correspondence between each retained KV state and its original absolute position\. These positions must remain unchanged because RoPE\(Suet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib123)\)uses them to encode positional relationships\.

#### Agent events and commit transitions\.

To track how information exchanged through tool use changes across turns,CommitKVtreats each completed tool call𝒂i\\boldsymbol\{a\}\_\{i\}and each returned observation𝒐i\\boldsymbol\{o\}\_\{i\}as an*agent event*\. Each event is partitioned into contiguous*event pages*\. We useEEto denote a contiguous token span taken from either𝒂i\\boldsymbol\{a\}\_\{i\}or𝒐i\\boldsymbol\{o\}\_\{i\}, with\|E\|≤G\|E\|\\leq G, whereGGis the maximum page size\. The pageEEis the basic unit whose influence is measured and whose corresponding KV states may later be removed\.

The importance of an event page may change after the agent receives the corresponding observation\. For example, information used to formulate a tool call may become less useful once the requested result is available\. SnapKV\(Liet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib92)\)uses a short observation window at the end of the prompt to aggregate attention from recent query tokens and identify prefix KV positions that are likely to remain important during subsequent generation\. Inspired by this window\-based importance estimation,CommitKVintroduces paired windows on opposite sides of each tool–observation boundary to measure how the importance of the same event page changes after the observation becomes available\.

For a tool call𝒂t\\boldsymbol\{a\}\_\{t\}, letccdenote the*tool\-call commit*at the end of its generation\. The tool call ends at position\|𝒚t\|\|\\boldsymbol\{y\}\_\{t\}\|\. After𝒐t\\boldsymbol\{o\}\_\{t\}is appended, the next\-turn input becomes𝒙t\+1=\(𝒚t,𝒐t\)\\boldsymbol\{x\}\_\{t\+1\}=\(\\boldsymbol\{y\}\_\{t\},\\boldsymbol\{o\}\_\{t\}\), and the next decoding begins at position\|𝒙t\+1\|\+1\|\\boldsymbol\{x\}\_\{t\+1\}\|\+1\. Given a query\-window sizeWW, we define the paired query windows

𝒬c−\\displaystyle\\mathcal\{Q\}\_\{c\}^\{\-\}=\{\|𝒚t\|−W\+1,…,\|𝒚t\|\},\\displaystyle=\\bigl\\\{\|\\boldsymbol\{y\}\_\{t\}\|\-W\+1,\\ldots,\|\\boldsymbol\{y\}\_\{t\}\|\\bigr\\\},\(5\)𝒬c\+\\displaystyle\\mathcal\{Q\}\_\{c\}^\{\+\}=\{\|𝒙t\+1\|\+1,…,\|𝒙t\+1\|\+W\}\.\\displaystyle=\\bigl\\\{\|\\boldsymbol\{x\}\_\{t\+1\}\|\+1,\\ldots,\|\\boldsymbol\{x\}\_\{t\+1\}\|\+W\\bigr\\\}\.
The pre\-commit query window𝒬c−\\mathcal\{Q\}\_\{c\}^\{\-\}contains the finalWWself\-attention query positions ending at the completion of the tool call\. The post\-commit query window𝒬c\+\\mathcal\{Q\}\_\{c\}^\{\+\}contains the firstWWself\-attention query positions generated during the next turn, after the returned observation has been incorporated into the input\. At each query position, the attention mechanism reads information from the resident KV states\. Measuring the effect of an event page over a query window therefore indicates how strongly that page contributes to the corresponding attention outputs\.

Comparing the influence of the same page across𝒬c−\\mathcal\{Q\}\_\{c\}^\{\-\}and𝒬c\+\\mathcal\{Q\}\_\{c\}^\{\+\}forms a*commit transition*\. Because the two query windows lie immediately before and after the returned observation becomes available, this comparison captures how the role of the page changes across the tool–observation boundary\. The sequence of commit transitions observed for a page across tool interactions defines its*event lifecycle*\.

### 3\.2Lifecycle States from Commit Transitions

The preceding subsection defines a commit transition by comparing the influence of the same event page before a tool\-call commit and after the returned observation\.CommitKVuses this transition to distinguish pages that may have completed their observed role from pages that are merely inactive\. Specifically, a high\-to\-low change nominates a page for lifecycle\-aware removal at a later compression checkpoint, whereas a low\-to\-low page is not removed by this rule because it may become useful in future turns\. To identify these changes,CommitKVmeasures the effect of deleting each page in the pre\- and post\-commit query windows\.

#### Deletion effect for event page\.

CommitKVrecords the attention weights and outputs produced by one attention layer\. Consider one attention head; the following computation is performed independently for every head\. For a query positionp∈𝒬p\\in\\mathcal\{Q\}, letap​ia\_\{pi\}be the attention weight assigned to resident tokenii, let𝒗i\\boldsymbol\{v\}\_\{i\}be its value vector, and let𝒐p\\boldsymbol\{o\}\_\{p\}be the original attention\-head output\.

For the previously defined event pageEE, deleting its tokens and renormalizing the remaining attention weights gives

𝒐p∖E=𝒐p−∑i∈Epi≤pap​i​𝒗i1−∑i∈Epi≤pap​i,\\boldsymbol\{o\}\_\{p\}^\{\\setminus E\}=\\frac\{\\boldsymbol\{o\}\_\{p\}\-\\sum\_\{\\begin\{subarray\}\{c\}i\\in E\\\\ p\_\{i\}\\leq p\\end\{subarray\}\}a\_\{pi\}\\boldsymbol\{v\}\_\{i\}\}\{1\-\\sum\_\{\\begin\{subarray\}\{c\}i\\in E\\\\ p\_\{i\}\\leq p\\end\{subarray\}\}a\_\{pi\}\},\(6\)where𝒐p∖E\\boldsymbol\{o\}\_\{p\}^\{\\setminus E\}is the output after deletingEE\. For a query window𝒬\\mathcal\{Q\}, we first compute the maximum relative output change for each head and then take the maximum across heads:

Rh​\(E;𝒬\)\\displaystyle R\_\{h\}\(E;\\mathcal\{Q\}\)=maxp∈𝒬⁡‖𝒐p−𝒐p∖E‖2‖𝒐p‖2,\\displaystyle=\\max\_\{p\\in\\mathcal\{Q\}\}\\frac\{\\\|\\boldsymbol\{o\}\_\{p\}\-\\boldsymbol\{o\}\_\{p\}^\{\\setminus E\}\\\|\_\{2\}\}\{\\\|\\boldsymbol\{o\}\_\{p\}\\\|\_\{2\}\},\(7\)R​\(E;𝒬\)\\displaystyle R\(E;\\mathcal\{Q\}\)=maxh⁡Rh​\(E;𝒬\)\.\\displaystyle=\\max\_\{h\}R\_\{h\}\(E;\\mathcal\{Q\}\)\.A largerR​\(E;𝒬\)R\(E;\\mathcal\{Q\}\)means that deletingEEcauses a larger relative change in the attention output\.CommitKVevaluates this score in the pre\- and post\-commit query windows\. Comparing the two scores provides the signal used to assign the lifecycle state ofEEand to nominate pages for removal at a later compression checkpoint\.

#### Influence comparison across a commit\.

The deletion effect defined above measures the influence of an event page within one query window\. To determine how this influence changes after a returned observation,CommitKVcompares the deletion effect across the pre\- and post\-commit windows\. For each tool\-call commitcc,

Rc−​\(E\)=R​\(E;𝒬c−\),Rc\+​\(E\)=R​\(E;𝒬c\+\)\.R\_\{c\}^\{\-\}\(E\)=R\(E;\\mathcal\{Q\}\_\{c\}^\{\-\}\),\\qquad R\_\{c\}^\{\+\}\(E\)=R\(E;\\mathcal\{Q\}\_\{c\}^\{\+\}\)\.\(8\)Here,Rc−​\(E\)R\_\{c\}^\{\-\}\(E\)measures the influence ofEEbefore the returned observation, whileRc\+​\(E\)R\_\{c\}^\{\+\}\(E\)measures its influence after the observation has been processed\. Their comparison therefore shows how the role ofEEchanges across the commit\.

Because the two windows contain different query tokens, their deletion effects may have different numerical scales\. Using only a fixed threshold may therefore give inconsistent labels across the two windows\. To obtain a relative measure, letℰc\\mathcal\{E\}\_\{c\}denote the set of event pages whose deletion effects can be computed in both𝒬c−\\mathcal\{Q\}\_\{c\}^\{\-\}and𝒬c\+\\mathcal\{Q\}\_\{c\}^\{\+\}\. For eachE∈ℰcE\\in\\mathcal\{E\}\_\{c\}, letPc−​\(E\)P\_\{c\}^\{\-\}\(E\)andPc\+​\(E\)P\_\{c\}^\{\+\}\(E\)denote the percentile ranks ofRc−​\(E\)R\_\{c\}^\{\-\}\(E\)andRc\+​\(E\)R\_\{c\}^\{\+\}\(E\)among the pages inℰc\\mathcal\{E\}\_\{c\}, respectively\. A larger percentile means thatEEis more influential than a larger fraction of the comparison pages\.

Neither the deletion effect nor the percentile is sufficient alone\. A page may rank highly even when all deletion effects are very small, while a low\-ranked page may still have a large deletion effect when all pages are important\. We therefore classify a page using both its absolute deletion effect and its relative percentile:

Hc±​\(E\)=1\\displaystyle H\_\{c\}^\{\\pm\}\(E\)=1⟺Rc±​\(E\)≥τuse​and​Pc±​\(E\)≥ρuse,\\displaystyle\\Longleftrightarrow R\_\{c\}^\{\\pm\}\(E\)\\geq\\tau\_\{\\mathrm\{use\}\}\\ \\text\{and\}\\ P\_\{c\}^\{\\pm\}\(E\)\\geq\\rho\_\{\\mathrm\{use\}\},\(9\)Lc±​\(E\)=1\\displaystyle L\_\{c\}^\{\\pm\}\(E\)=1⟺Rc±​\(E\)≤τdead​and​Pc±​\(E\)≤ρdead\.\\displaystyle\\Longleftrightarrow R\_\{c\}^\{\\pm\}\(E\)\\leq\\tau\_\{\\mathrm\{dead\}\}\\ \\text\{and\}\\ P\_\{c\}^\{\\pm\}\(E\)\\leq\\rho\_\{\\mathrm\{dead\}\}\.Here,Hc±​\(E\)H\_\{c\}^\{\\pm\}\(E\)andLc±​\(E\)L\_\{c\}^\{\\pm\}\(E\)are binary indicators of whetherEEhas high or low influence, respectively\. The superscripts−\-and\+\+refer to the pre\- and post\-commit windows\. The thresholdsτuse\\tau\_\{\\mathrm\{use\}\}andτdead\\tau\_\{\\mathrm\{dead\}\}are applied to the absolute deletion effect, whileρuse\\rho\_\{\\mathrm\{use\}\}andρdead\\rho\_\{\\mathrm\{dead\}\}are applied to the percentile rank\. We useτdead<τuse\\tau\_\{\\mathrm\{dead\}\}<\\tau\_\{\\mathrm\{use\}\}andρdead<ρuse\\rho\_\{\\mathrm\{dead\}\}<\\rho\_\{\\mathrm\{use\}\}, leaving pages that satisfy neither condition unclassified rather than forcing an unreliable decision\.

#### Lifecycle\-state classification\.

The purpose of the paired measurements is not to remove every page with a low score\. Instead,CommitKVlooks for a clear decrease in influence, which provides evidence that a page has completed its observed role\. The lifecycle state ofEEat commitccis defined as

λc​\(E\)=\{𝖼𝗈𝗆𝗉𝗅𝖾𝗍𝗂𝗈𝗇𝖼𝖺𝗇𝖽𝗂𝖽𝖺𝗍𝖾Hc−​\(E\)=Lc\+​\(E\)=1,𝖽𝗈𝗋𝗆𝖺𝗇𝗍Lc−​\(E\)=Lc\+​\(E\)=1,𝗇𝖾𝗐𝗅𝗒​𝖺𝖼𝗍𝗂𝗏𝖾Lc−​\(E\)=Hc\+​\(E\)=1,𝗌𝗍𝗂𝗅𝗅​𝖺𝖼𝗍𝗂𝗏𝖾Hc−​\(E\)=Hc\+​\(E\)=1,𝗎𝗇𝖼𝖾𝗋𝗍𝖺𝗂𝗇otherwise\.\\lambda\_\{c\}\(E\)=\\left\\\{\\begin\{array\}\[\]\{@\{\}l@\{\\;\}l@\{\}\}\\begin\{array\}\[c\]\{@\{\}l@\{\}\}\\mathsf\{completion\}\\\\ \\mathsf\{candidate\}\\end\{array\}&H\_\{c\}^\{\-\}\(E\)=L\_\{c\}^\{\+\}\(E\)=1,\\\\ \\mathsf\{dormant\}&L\_\{c\}^\{\-\}\(E\)=L\_\{c\}^\{\+\}\(E\)=1,\\\\ \\mathsf\{newly\\ active\}&L\_\{c\}^\{\-\}\(E\)=H\_\{c\}^\{\+\}\(E\)=1,\\\\ \\mathsf\{still\\ active\}&H\_\{c\}^\{\-\}\(E\)=H\_\{c\}^\{\+\}\(E\)=1,\\\\ \\mathsf\{uncertain\}&\\text\{otherwise\}\.\\end\{array\}\\right\.\(10\)
A high\-to\-low transition marksEEas a completion candidate because the page was influential before the commit but became unimportant after the returned observation was available\. This change suggests that the page has completed its role in the current tool interaction\. In contrast, a low\-to\-low page is marked as dormant rather than completed: its information may simply not have been needed yet\. A low\-to\-high page becomes newly active after the observation, while a high\-to\-high page remains active across the commit\. Pages without a clear high or low label are marked as uncertain to avoid making decisions from weak evidence\. Only completion candidates proceed to the joint validation in Section[3\.3](https://arxiv.org/html/2608.07855#S3.SS3)\. Pages classified as dormant, newly active, still active, or uncertain are not added to the retirement set by the lifecycle rule\.

### 3\.3Lifecycle\-Aware Joint Compression

The lifecycle state indicates whether an event page may have completed its role, but it does not directly determine whether the page can be safely removed\. Before retiring any page,CommitKVmust account for the cumulative effect of removing multiple completion candidates and preserve pages whose pre\-commit effects have been measured but whose post\-commit effects are not yet available\.CommitKVtherefore performs two operations: it jointly validates completion candidates to construct the retirement set and temporarily protects pages awaiting post\-commit measurements\.

#### Retirement\-set construction\.

The post\-commit deletion effect of each completion candidate is initially measured by deleting that page alone\. However, small individual effects can accumulate when several pages are removed together, potentially causing a much larger change in the attention output\.CommitKVtherefore evaluates completion candidates jointly before adding them to the retirement set\.

At commitcc, the candidates are sorted by increasingRc\+​\(E\)R\_\{c\}^\{\+\}\(E\), with decreasingRc−​\(E\)R\_\{c\}^\{\-\}\(E\)as the second ordering criterion\. This order prioritizes pages that have little influence after the returned observation but were clearly influential before it, providing stronger evidence that they have completed their role\.

Starting from an empty retirement set𝒟c=∅\\mathcal\{D\}\_\{c\}=\\varnothing,CommitKVexamines the candidates in this order\. A candidate pageEEis added only if its joint deletion with the pages already accepted in𝒟c\\mathcal\{D\}\_\{c\}satisfies

Rc\+​\(𝒟c∪\{E\}\)≤τjoint\.R\_\{c\}^\{\+\}\\left\(\\mathcal\{D\}\_\{c\}\\cup\\\{E\\\}\\right\)\\leq\\tau\_\{\\mathrm\{joint\}\}\.\(11\)Here,Rc\+​\(𝒜\)R\_\{c\}^\{\+\}\(\\mathcal\{A\}\)denotes the post\-commit deletion effect obtained by deleting all tokens belonging to the page set𝒜\\mathcal\{A\}at the same time\. It is computed using Eqs\. \([6](https://arxiv.org/html/2608.07855#S3.E6)\) and \([7](https://arxiv.org/html/2608.07855#S3.E7)\), treating the tokens in𝒜\\mathcal\{A\}as one combined deletion set\. The thresholdτjoint\\tau\_\{\\mathrm\{joint\}\}limits the cumulative change caused by the accepted pages\. If addingEEviolates this threshold,EEis not included in𝒟c\\mathcal\{D\}\_\{c\}\. Passing this test only marks a page for retirement; its KV states remain resident until a later compression checkpoint\.

Algorithm 1CommitKVfor a multi\-turn ReAct agent1:budget

BB, pending fraction

η\\eta, page size

GG, window size

WW, and lifecycle thresholds

2:

BN←⌊η​B⌋B\_\{N\}\\leftarrow\\lfloor\\eta B\\rfloor;

𝒟c←∅\\mathcal\{D\}\_\{c\}\\leftarrow\\varnothing
3:foragent turn

t=1,2,…t=1,2,\\ldotsdo

4:Prefill

𝒒\\boldsymbol\{q\}if

t=1t=1, otherwise

𝒐t−1\\boldsymbol\{o\}\_\{t\-1\}; decode

𝒛t,𝒂t\\boldsymbol\{z\}\_\{t\},\\boldsymbol\{a\}\_\{t\}
5:if

t\>1t\>1and

𝒬ct−1\+\\mathcal\{Q\}\_\{c\_\{t\-1\}\}^\{\+\}is availablethen

6:Form

ℰct−1\\mathcal\{E\}\_\{c\_\{t\-1\}\}; compute

Rct−1\+​\(E\)R\_\{c\_\{t\-1\}\}^\{\+\}\(E\)and

Pct−1±​\(E\)P\_\{c\_\{t\-1\}\}^\{\\pm\}\(E\)
7:Identify and sort completion candidates by Eqs\. \([9](https://arxiv.org/html/2608.07855#S3.E9)\)–\([10](https://arxiv.org/html/2608.07855#S3.E10)\)

8:

𝒜←∅\\mathcal\{A\}\\leftarrow\\varnothing
9:foreach candidate

EEin increasing

Rct−1\+​\(E\)R\_\{c\_\{t\-1\}\}^\{\+\}\(E\)and decreasing

Rct−1−​\(E\)R\_\{c\_\{t\-1\}\}^\{\-\}\(E\)do

10:if

Rct−1\+​\(𝒜∪\{E\}\)≤τjointR\_\{c\_\{t\-1\}\}^\{\+\}\(\\mathcal\{A\}\\cup\\\{E\\\}\)\\leq\\tau\_\{\\mathrm\{joint\}\}then

11:

𝒜←𝒜∪\{E\}\\mathcal\{A\}\\leftarrow\\mathcal\{A\}\\cup\\\{E\\\}
12:endif

13:endfor

14:

𝒟c←𝒟c∪𝒜\\mathcal\{D\}\_\{c\}\\leftarrow\\mathcal\{D\}\_\{c\}\\cup\\mathcal\{A\}
15:endif

16:if

𝒂t\\boldsymbol\{a\}\_\{t\}is a final answerthen

17:return

𝒂t\\boldsymbol\{a\}\_\{t\}
18:endif

19:Set commit

ctc\_\{t\}; partition newly completed events into pages with

\|E\|≤G\|E\|\\leq G
20:Compute

Rct−​\(E\)R\_\{c\_\{t\}\}^\{\-\}\(E\)over

𝒬ct−\\mathcal\{Q\}\_\{c\_\{t\}\}^\{\-\}by Eqs\. \([5](https://arxiv.org/html/2608.07855#S3.E5)\)–\([8](https://arxiv.org/html/2608.07855#S3.E8)\)

21:Protect pending pages by decreasing

Rct−​\(E\)R\_\{c\_\{t\}\}^\{\-\}\(E\), subject to

BNB\_\{N\}
22:ifcompression checkpoint

jjis reachedthen

23:Form

𝒩j\\mathcal\{N\}\_\{j\}, exclude

𝒟c\\mathcal\{D\}\_\{c\}, and select

ℐj\\mathcal\{I\}\_\{j\}with

𝒩j⊆ℐj\\mathcal\{N\}\_\{j\}\\subseteq\\mathcal\{I\}\_\{j\}and

\|ℐj\|≤B\|\\mathcal\{I\}\_\{j\}\|\\leq B
24:Apply

ℐj\\mathcal\{I\}\_\{j\}by Eq\. \([4](https://arxiv.org/html/2608.07855#S3.E4)\)

25:endif

26:endfor

#### Protecting incomplete transitions\.

A commit transition requires both a pre\-commit and a post\-commit measurement\. The pre\-commit measurement becomes available when the tool call is completed, whereas the post\-commit measurement can only be collected after the returned observation has been processed and the next\-turn queries have been generated\. During this interval, the page must remain in the KV cache\. Otherwise, its post\-commit deletion effect cannot be measured and its lifecycle state cannot be determined\.

At checkpointjj, let𝒫j⊆\{1,…,Sj\}\\mathcal\{P\}\_\{j\}\\subseteq\\\{1,\\ldots,S\_\{j\}\\\}denote all resident token indices belonging to pages awaiting post\-commit measurements, and let𝒩j⊆𝒫j\\mathcal\{N\}\_\{j\}\\subseteq\\mathcal\{P\}\_\{j\}denote the subset protected byCommitKV\. To preserve these incomplete transitions without consuming an unbounded portion of the KV cache,CommitKVprotects at most

\|𝒩j\|≤BN,BN=⌊η​B⌋,\|\\mathcal\{N\}\_\{j\}\|\\leq B\_\{N\},\\qquad B\_\{N\}=\\lfloor\\eta B\\rfloor,\(12\)whereBBis the total KV\-cache token budget andη∈\[0,1\]\\eta\\in\[0,1\]specifies the fraction reserved for pending measurements\. If\|𝒫j\|\>BN\|\\mathcal\{P\}\_\{j\}\|\>B\_\{N\}, pages with larger pre\-commit deletion effects are protected first until\|𝒩j\|≤BN\|\\mathcal\{N\}\_\{j\}\|\\leq B\_\{N\}\. These pages receive priority because a strong pre\-commit effect is necessary for identifying a meaningful high\-to\-low transition\.

After the retirement and protection decisions,CommitKVdetermines the KV states retained at compression checkpointjj\. Token indices belonging to pages in𝒟c\\mathcal\{D\}\_\{c\}are excluded, while the protected pending indices𝒩j\\mathcal\{N\}\_\{j\}are retained\. The resulting retained\-token index setℐj\\mathcal\{I\}\_\{j\}satisfies

𝒩j⊆ℐj,\|ℐj\|≤B\.\\mathcal\{N\}\_\{j\}\\subseteq\\mathcal\{I\}\_\{j\},\\qquad\|\\mathcal\{I\}\_\{j\}\|\\leq B\.\(13\)The same index setℐj\\mathcal\{I\}\_\{j\}is applied to the cached keys, values, and their absolute positions, keeping them aligned after compression\.

Table 1:Accuracy comparison across eight datasets and three model backbones under a 4096\-token KV\-cache budget\. All results are reported as percentages \(%\\%\), and Average is the unweighted mean over the eight datasets\.Boldmarks the best compressed result for each backbone\.

### 3\.4Overall Method Overview

Figure[2](https://arxiv.org/html/2608.07855#S2.F2)summarizes the completeCommitKVpipeline, and Algorithm[1](https://arxiv.org/html/2608.07855#alg1)gives its turn\-by\-turn procedure\. At each turn, the server reuses the cached history and prefills𝒒\\boldsymbol\{q\}at the first turn or only the new observation𝒐t−1\\boldsymbol\{o\}\_\{t\-1\}thereafter\. Once𝒬c\+\\mathcal\{Q\}\_\{c\}^\{\+\}is available,CommitKVformsℰc\\mathcal\{E\}\_\{c\}, the set of event pages whose deletion effects can be computed in both𝒬c−\\mathcal\{Q\}\_\{c\}^\{\-\}and𝒬c\+\\mathcal\{Q\}\_\{c\}^\{\+\}\. It then computesRc\+​\(E\)R\_\{c\}^\{\+\}\(E\), combines it with the storedRc−​\(E\)R\_\{c\}^\{\-\}\(E\), assignsλc​\(E\)\\lambda\_\{c\}\(E\), and jointly validates completion candidates to construct𝒟c\\mathcal\{D\}\_\{c\}\. If𝒂t\\boldsymbol\{a\}\_\{t\}is a tool call, newly completed events are partitioned into pages, and pages with largerRc−​\(E\)R\_\{c\}^\{\-\}\(E\)are protected for the next post\-commit measurement, subject toBNB\_\{N\}\. At checkpointjj, their resident token indices form𝒩j\\mathcal\{N\}\_\{j\}\.CommitKVretains𝒩j\\mathcal\{N\}\_\{j\}, excludes token indices belonging to accepted pages in𝒟c\\mathcal\{D\}\_\{c\}, producing the retained\-token index setℐj\\mathcal\{I\}\_\{j\}\. Finally,ℐj\\mathcal\{I\}\_\{j\}is applied to the keys, values, and absolute positions\.

## 4Experiments

### 4\.1Experimental Setup

#### Models\.

We evaluate six open\-weight backbones spanning dense Transformers, hybrid architectures that combine linear attention with standard self\-attention, distilled or specialized reasoning models and mixture\-of\-experts \(MoE\) models: Qwen3\-14B, Qwen3\.6\-27B\(Yanget al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib81)\), DeepSeek\-R1\-Distill\-Llama\-8B\(Guoet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib84)\), GPT\-OSS\-20B\(Agarwalet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib171)\), Phi\-4\-Reasoning\(Abdinet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib166)\), and InternLM3\-8B\-Instruct\(Caiet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib198)\)\. The main text reports results for the first three models, while Appendix presents the remaining results\.

#### Benchmarks\.

The evaluation covers two complementary types of benchmarks to assess both reasoning preservation and tool\-mediated evidence use\. The reasoning\-intensive group contains GPQA\(Reinet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib177)\), Bamboogle\(Presset al\.,[2023](https://arxiv.org/html/2608.07855#bib.bib77)\), MATH\-500\(Hendryckset al\.,[2021](https://arxiv.org/html/2608.07855#bib.bib159)\), and AIME25\(Zhang and Team,[2025](https://arxiv.org/html/2608.07855#bib.bib169)\), which evaluate multi\-step scientific, compositional, and mathematical reasoning\. The tool\-use and evidence\-intensive group contains FRAMES\(Krishnaet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib130)\), GAIA\(Mialonet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib134)\), ToolHop\(Yeet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib160)\), and xbench\-DeepSearch\(Chenet al\.,[2025](https://arxiv.org/html/2608.07855#bib.bib179)\), which evaluate tool use and evidence aggregation across multiple turns\.

#### Baselines\.

We compareCommitKVwith representative KV\-cache compression methods that provide publicly available implementations and can be reliably integrated into the same multi\-turn inference framework\. Specifically, we compare withFullKV,SnapKV\(Liet al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib92)\),R\-KV\(Caiet al\.,[2025a](https://arxiv.org/html/2608.07855#bib.bib101)\), andTriAttention\(Maoet al\.,[2026](https://arxiv.org/html/2608.07855#bib.bib182)\)\. SnapKV, R\-KV, and TriAttention are commonly used KV cache compression baselines for reasoning models\. During each ReAct agent turn, we adopt the same settings as R\-KV and TriAttention: a compression policy is triggered every 128 decoded tokens, pruning the KV cache to the specified budget\. Appendix provides further details\.

#### Implementation Details\.

We conduct all experiments on 8×\\timesNVIDIA H20 96GB GPUs\. We implementCommitKVin SGLang\(Zhenget al\.,[2024](https://arxiv.org/html/2608.07855#bib.bib107)\)\. We evaluate KV\-cache budgets of 2048, 4096, and 8192 tokens\. We setW=8W=8,G=16G=16,η=0\.125\\eta=0\.125,τuse=0\.05\\tau\_\{\\mathrm\{use\}\}=0\.05,τdead=τjoint=0\.01\\tau\_\{\\mathrm\{dead\}\}=\\tau\_\{\\mathrm\{joint\}\}=0\.01,ρuse=0\.75\\rho\_\{\\mathrm\{use\}\}=0\.75, andρdead=0\.25\\rho\_\{\\mathrm\{dead\}\}=0\.25\. At each commit, we scan at most 64 pages and protect at most 16 pending pages, with their total token count bounded byBN=⌊η​B⌋B\_\{N\}=\\lfloor\\eta B\\rfloor\. We keep prompts, decoding settings, tool outputs, hardware, and tensor\-parallel configurations fixed across methods\. Appendix provides the complete protocol\.

Table 2:Comparison of accuracy for the Qwen3\-14B model across eight datasets with KV cache budgets of 2048 and 8192 tokens\. Average is the unweighted mean over all eight datasets;boldmarks the best compressed result under each cache budget\.BudgetMethodTool/evidenceReasoningAverageFRAMESGAIAToolHopxbench\-DeepSearchGPQABamboogleMATH\-500AIME252048FullKV36\.2921\.3676\.9020\.0050\.5167\.2076\.0056\.6750\.62SnapKV3\.643\.881\.524\.0013\.1332\.0052\.200\.0013\.80R\-KV5\.706\.805\.583\.0012\.6341\.6051\.003\.3316\.21TriAttention6\.075\.833\.301\.0012\.1229\.6052\.400\.0013\.79CommitKV28\.2815\.5341\.8813\.0047\.4758\.4073\.0030\.0038\.458192SnapKV35\.8018\.4569\.8012\.0030\.0069\.6077\.4033\.3343\.30R\-KV37\.9923\.3063\.7113\.0040\.0068\.8072\.8030\.0043\.70TriAttention32\.7716\.5067\.2614\.0032\.8360\.0073\.8023\.3340\.06CommitKV38\.4719\.4277\.6611\.0048\.4862\.4075\.2046\.6747\.41

### 4\.2Main Results

#### Comparison with baselines\.

Table[1](https://arxiv.org/html/2608.07855#S3.T1)reports the results across eight benchmarks and three model backbones under a 4096\-token KV\-cache budget\.CommitKVachieves the highest average accuracy among all compressed methods, outperforming the strongest baseline by 19\.85%\\%, 5\.49%\\%, and 1\.54%\\%on the three models, respectively\. Furthermore, compared to FullKV,CommitKVincurs minimal performance degradation\.

#### Cache budgets\.

Table[2](https://arxiv.org/html/2608.07855#S4.T2)reports the average accuracy over the eight benchmarks for Qwen3\-14B under 2048\- and 8192\-token KV\-cache budgets\.CommitKVachieves the best accuracy among the compressed methods under both budgets, demonstrating its robustness across different cache constraints\. Notably, under the aggressive 2048\-token budget setting,CommitKVoutperforms the best R\-KV method by 22\.24%\\%and substantially narrows the gap with the FullKV\.

### 4\.3Memory and Latency Analysis

We evaluate the memory and latency of Qwen3\-14B on GAIA, xbench\-DeepSearch, and GPQA under a 2048\-token KV\-cache budget\. We report two complementary efficiency metrics: peak KV\-cache memory and wall\-clock time per completed sample\. Peak KV\-cache memory is the maximum per\-GPU KV\-cache footprint over a complete agent trajectory, while wall\-clock time measures the average end\-to\-end time required to complete one sample over the full multi\-turn execution process\. Lower values are better for both metrics\.

![Refer to caption](https://arxiv.org/html/2608.07855v1/x3.png)Figure 3:Peak per\-GPU KV\-cache memory and wall\-clock time per completed sample for Qwen3\-14B on GAIA, xbench\-DeepSearch, and GPQA with a 2048\-token budget; lower is better\. FullKV is the uncompressed reference\.As shown in Figure[3](https://arxiv.org/html/2608.07855#S4.F3),CommitKVconsistently achieves the lowest peak KV\-cache memory and wall\-clock time across all three benchmarks\. Specifically,CommitKVachieves up to 5\.00×\\timessavings in peak KV memory and a 5\.62×\\timesspeedup in wall\-clock time\. These results show that lifecycle\-aware cache retirement reduces the memory footprint while accelerating end\-to\-end agent execution without introducing prohibitive runtime overhead\.

### 4\.4Ablation Studies

We conduct ablation studies on Qwen3\-14B with a 4096\-token KV\-cache budget\. We compare the completeCommitKVwith five variants\.*w/o Lifecycle Mechanism*removes all lifecycle\-aware operations\.*w/o Pre\-Commit State*identifies retirement candidates using onlyLc\+​\(E\)=1L\_\{c\}^\{\+\}\(E\)=1, instead of the paired high\-to\-low conditionHc−​\(E\)=Lc\+​\(E\)=1H\_\{c\}^\{\-\}\(E\)=L\_\{c\}^\{\+\}\(E\)=1\.*w/o Percentile Calibration*determines high\- and low\-influence states using only the absolute deletion effects and the thresholdsτuse\\tau\_\{\\mathrm\{use\}\}andτdead\\tau\_\{\\mathrm\{dead\}\}\.*w/o Joint Validation*directly retires all completion candidates without applying Eq\. \([11](https://arxiv.org/html/2608.07855#S3.E11)\)\. Finally,*w/o Transition Protection*does not reserve𝒩j\\mathcal\{N\}\_\{j\}, allowing pages awaiting post\-commit measurements to be removed at a compression checkpoint\.

Table 3:Ablation results for Qwen3\-14B with a 4096\-token KV\-cache budget\.Boldmarks the best result\.As shown in Table[3](https://arxiv.org/html/2608.07855#S4.T3), the completeCommitKVperforms best on all three benchmarks\. Disabling the lifecycle mechanism lowers the average accuracy from 48\.04%\\%to 40\.68%\\%\. Removing the pre\-commit state reduces it to 45\.30%\\%, showing that post\-commit influence alone cannot distinguish dormant from completed pages\. Removing percentile calibration and joint validation decreases average accuracy by 5\.08%\\%and 5\.11%\\%, respectively\. Removing transition protection causes a 6\.52%\\%decrease, consistent with pending pages being removed before their post\-commit measurements are collected\. The above ablation experiments demonstrate that all components are effective\.

## 5Conclusion and Future Work

We presentCommitKV, a lifecycle\-aware KV cache compression method for multi\-turn agents\. It compares event pages across tool\-call commits to distinguish dormant from completed information and jointly validates high\-to\-low candidates before retirement\. Experiments demonstrate lower memory use, faster end\-to\-end inference, and better performance than existing methods\. Future work will address the remaining gap to FullKV under small cache budgets\.

## References

- M\. Abdin, S\. Agarwal, A\. Awadallah, V\. Balachandran, H\. Behl, L\. Chen, G\. de Rosa, S\. Gunasekar, M\. Javaheripi, N\. Joshi,et al\.\(2025\)Phi\-4\-reasoning technical report\.arXiv preprint arXiv:2504\.21318\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1),[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px1.p1.1)\.
- S\. Agarwal, L\. Ahmad, J\. Ai, S\. Altman, A\. Applebaum, E\. Arbus, R\. K\. Arora, Y\. Bai, B\. Baker, H\. Bao,et al\.\(2025\)Gpt\-oss\-120b & gpt\-oss\-20b model card\.arXiv preprint arXiv:2508\.10925\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1),[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px1.p1.1)\.
- Z\. Cai, W\. Xiao, H\. Sun, c\. Luo, Y\. Zhang, K\. Wan, Y\. Li, Y\. Zhou, L\. Chang, J\. Gu, Z\. Dong, A\. Anandkumar, A\. Asi, and J\. Hu \(2025a\)R\-kv: redundancy\-aware kv cache compression for reasoning models\.Advances in Neural Information Processing Systems38,pp\. 60980–61005\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p2.1),[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px3.p1.1)\.
- Z\. Cai, Y\. Zhang, B\. Gao, Y\. Liu, Y\. Li, T\. Liu, K\. Lu, W\. Xiong, Y\. Dong, J\. Hu, and W\. Xiao \(2025b\)PyramidKV: dynamic KV cache compression based on pyramidal information funneling\.InSecond Conference on Language Modeling,External Links:[Link](https://openreview.net/forum?id=ayi7qezU87)Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p2.1)\.
- Z\. Cai, M\. Cao, H\. Chen, K\. Chen, K\. Chen, X\. Chen, X\. Chen, Z\. Chen, Z\. Chen, P\. Chu,et al\.\(2024\)Internlm2 technical report\.arXiv preprint arXiv:2403\.17297\.Cited by:[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px1.p1.1)\.
- C\. Chang, W\. Lin, C\. Lin, C\. Chen, Y\. Hu, P\. Wang, N\. Huang, L\. Ceze, M\. Abdelfattah, and K\. Wu \(2025\)Palu: kv\-cache compression with low\-rank projection\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 50222–50249\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Value\-aware stochastic kv cache eviction for reasoning models\.arXiv preprint arXiv:2606\.03928\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- K\. Chen, Y\. Ren, Y\. Liu, X\. Hu, H\. Tian, T\. Xie, F\. Liu, H\. Zhang, H\. Liu, Y\. Gong,et al\.\(2025\)Xbench: tracking agents productivity scaling with profession\-aligned real\-world evaluations\.arXiv preprint arXiv:2506\.13651\.Cited by:[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px2.p1.1)\.
- Y\. Cheng, C\. Zhang, Z\. Zhang, X\. Meng, S\. Hong, W\. Li, Z\. Wang, Z\. Wang, F\. Yin, J\. Zhao,et al\.\(2024\)Exploring large language model based intelligent agents: definitions, methods, and prospects\.arXiv preprint arXiv:2401\.03428\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1)\.
- K\. T\. Chitty\-Venkata, J\. Ye, S\. Raskar, A\. Kougkas, X\. Sun, M\. Emani, V\. Vishwanath, and B\. Nicolae \(2026\)Pagedeviction: structured block\-wise kv cache pruning for efficient large language model inference\.InFindings of the Association for Computational Linguistics: EACL 2026,pp\. 3207–3218\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Feng, J\. Lv, Y\. Cao, X\. Xie, and S\. K\. Zhou \(2025\)Ada\-kv: optimizing kv cache eviction by adaptive budget allocation for efficient llm inference\.Advances in Neural Information Processing Systems38,pp\. 113152–113188\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p2.1),[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- B\. Gao, Z\. He, P\. Sharma, Q\. Kang, D\. Jevdjic, J\. Deng, X\. Yang, Z\. Yu, and P\. Zuo \(2024\)\{\\\{cost\-Efficient\}\\\}large language model serving for multi\-turn conversations with\{\\\{cachedattention\}\\\}\.In2024 USENIX annual technical conference \(USENIX ATC 24\),pp\. 111–126\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1),[§3\.1](https://arxiv.org/html/2608.07855#S3.SS1.SSS0.Px2.p1.2)\.
- S\. Ge, Y\. Zhang, L\. Liu, M\. Zhang, J\. Han, and J\. Gao \(2024\)Model tells you what to discard: adaptive kv cache compression for llms\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 22975–22988\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Z\. Geng, J\. Wang, Z\. Liu, F\. Ju, Y\. Li, X\. Li, M\. Yuan, J\. Hao, D\. Lian, E\. Chen,et al\.\(2025\)Accurate kv cache eviction via anchor direction projection for efficient llm inference\.Advances in Neural Information Processing Systems38,pp\. 10420–10456\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p2.1),[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- D\. Guo, D\. Yang, H\. Zhang, J\. Song, P\. Wang, Q\. Zhu, R\. Xu, R\. Zhang, S\. Ma, X\. Bi,et al\.\(2025\)DeepSeek\-r1 incentivizes reasoning in llms through reinforcement learning\.Nature645\(8081\),pp\. 633–638\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1),[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px1.p1.1)\.
- T\. Guo, X\. Chen, Y\. Wang, R\. Chang, S\. Pei, N\. V\. Chawla, O\. Wiest, and X\. Zhang \(2024\)Large language model based multi\-agents: a survey of progress and challenges\.arXiv preprint arXiv:2402\.01680\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1)\.
- X\. Hao, H\. Meng, X\. Yin, J\. Zhu, and C\. Cao \(2026\)Self\-gc: self\-governing context for long\-horizon llm agents\.arXiv preprint arXiv:2607\.00692\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. He, L\. Zhang, W\. Wu, J\. Liu, H\. Zhou, and B\. Zhuang \(2024\)Zipcache: accurate and efficient kv cache quantization with salient token identification\.Advances in Neural Information Processing Systems37,pp\. 68287–68307\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- D\. Hendrycks, C\. Burns, S\. Kadavath, A\. Arora, S\. Basart, E\. Tang, D\. Song, and J\. Steinhardt \(2021\)Measuring mathematical problem solving with the MATH dataset\.Advances in Neural Information Processing Systems34\.Cited by:[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px2.p1.1)\.
- C\. Hooper, S\. Kim, H\. Mohammadzadeh, M\. W\. Mahoney, Y\. S\. Shao, K\. Keutzer, and A\. Gholami \(2024\)Kvquant: towards 10 million context length llm inference with kv cache quantization\.Advances in Neural Information Processing Systems37,pp\. 1270–1303\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- L\. Hu, N\. Jia, R\. Wang, J\. Li, Q\. Yang, Y\. Hao, X\. Li, and X\. Liao \(2026\)TAKE: task\-aware chunked KV cache eviction for efficient long\-context LLM prefill\.External Links:[Link](https://openreview.net/forum?id=kMLfUshPwo)Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- W\. Huang, Y\. Zhang, X\. Zheng, F\. Chao, and R\. Ji \(2025a\)Determining layer\-wise sparsity for large language models through a theoretical perspective\.arXiv preprint arXiv:2502\.14770\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- W\. Huang, Y\. Zhang, X\. Zheng, Y\. Liu, J\. Lin, Y\. Yao, and R\. Ji \(2025b\)Dynamic low\-rank sparse adaptation for large language models\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 41926–41951\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- H\. Kang, Q\. Zhang, S\. Kundu, G\. Jeong, Z\. Liu, T\. Krishna, and T\. Zhao \(2024\)Gear: an efficient kv cache compression recipe for near\-lossless generative inference of llm\.arXiv preprint arXiv:2403\.05527\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- S\. Kariyappa and G\. E\. Suh \(2026\)SideQuest: model\-driven kv cache management for long\-horizon agentic reasoning\.arXiv preprint arXiv:2602\.22603\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- M\. Kim, A\. Kundu, H\. Kim, R\. Dixit, and M\. Cho \(2025\)EpiCache: episodic kv cache management for long\-term conversation on resource\-constrained environments\.arXiv preprint arXiv:2509\.17396\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Krishna, K\. Krishna, A\. Mohananey, S\. Schwarcz, A\. Stambler, S\. Upadhyay, and M\. Faruqui \(2025\)Fact, fetch, and reason: a unified evaluation of retrieval\-augmented generation\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 4745–4759\.Cited by:[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px2.p1.1)\.
- J\. Li, J\. Lou, and J\. Li \(2026\)IntentKV: cross\-turn intent\-aware kv cache pruning for agent inference\.arXiv preprint arXiv:2606\.09916\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. Li, Y\. Huang, B\. Yang, B\. Venkitesh, A\. Locatelli, H\. Ye, T\. Cai, P\. Lewis, and D\. Chen \(2024\)Snapkv: llm knows what you are looking for before generation\.Advances in Neural Information Processing Systems37,pp\. 22947–22970\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p2.1),[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1),[§3\.1](https://arxiv.org/html/2608.07855#S3.SS1.SSS0.Px4.p2.1),[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px3.p1.1)\.
- A\. Liu, J\. Liu, Z\. Pan, Y\. He, G\. Haffari, and B\. Zhuang \(2024a\)Minicache: kv cache compression in depth dimension for large language models\.Advances in Neural Information Processing Systems37,pp\. 139997–140031\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- T\. T\. Liu, J\. T\. H\. Wong, C\. Xiao, B\. Yang, H\. M\. Chen, and Y\. Zhao \(2026\)AgentKV: phase\-aware KV eviction for agentic LLMs\.InAdaptFM: Resource\-Adaptive Foundation Model Inference,External Links:[Link](https://openreview.net/forum?id=s80e8otSxP)Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- Z\. Liu, A\. Desai, F\. Liao, W\. Wang, V\. Xie, Z\. Xu, A\. Kyrillidis, and A\. Shrivastava \(2023\)Scissorhands: exploiting the persistence of importance hypothesis for llm kv cache compression at test time\.Advances in Neural Information Processing Systems36,pp\. 52342–52364\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Z\. Liu, J\. Yuan, H\. Jin, S\. Zhong, Z\. Xu, V\. Braverman, B\. Chen, and X\. Hu \(2024b\)KIVI: a tuning\-free asymmetric 2bit quantization for kv cache\.InInternational Conference on Machine Learning,pp\. 32332–32344\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- K\. Lyu, K\. Wu, P\. Li, X\. Hu, Q\. Si, C\. Miao, N\. Yang, Z\. Wang, L\. Xiao, L\. Hu,et al\.\(2026a\)Himemvln: enhancing reliability of open\-source zero\-shot vision\-and\-language navigation with hierarchical memory system\.arXiv preprint arXiv:2603\.14807\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1)\.
- K\. Lyu, Z\. Yuan, J\. He, Q\. Yan, X\. Su, N\. Hu, Y\. Liu, C\. Hao, S\. Qin, L\. Hu,et al\.\(2026b\)Photocraft: agentic reasoning with hierarchical self\-evolving memory for deep image search\.arXiv preprint arXiv:2606\.03099\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1)\.
- Y\. Ma, T\. Jin, X\. Zheng, Y\. Wang, H\. Li, Y\. Wu, G\. Jiang, W\. Zhang, and R\. Ji \(2023a\)Ompq: orthogonal mixed precision quantization\.InProceedings of the AAAI conference on artificial intelligence,Vol\.37,pp\. 9029–9037\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Ma, H\. Li, X\. Zheng, F\. Ling, X\. Xiao, R\. Wang, S\. Wen, F\. Chao, and R\. Ji \(2024a\)Affinequant: affine transformation quantization for large language models\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 50932–50951\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Ma, H\. Li, X\. Zheng, F\. Ling, X\. Xiao, R\. Wang, S\. Wen, F\. Chao, and R\. Ji \(2024b\)Outlier\-aware slicing for post\-training quantization in vision transformer\.InForty\-first International Conference on Machine Learning,Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Ma, H\. Li, X\. Zheng, X\. Xiao, R\. Wang, S\. Wen, X\. Pan, F\. Chao, and R\. Ji \(2023b\)Solving oscillation problem in post\-training quantization through a theoretical perspective\.In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 7950–7959\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Ma, X\. Zheng, J\. Xu, X\. Xu, F\. Ling, X\. Zheng, H\. Kuang, H\. Li, X\. Wang, X\. Xiao,et al\.\(2026\)Flow caching for autoregressive video generation\.arXiv preprint arXiv:2602\.10825\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- W\. Mao, X\. Lin, W\. Huang, Y\. Xie, T\. Fu, B\. Zhuang, S\. Han, and Y\. Chen \(2026\)Triattention: efficient long reasoning with trigonometric kv compression\.arXiv preprint arXiv:2604\.04921\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p2.1),[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px3.p1.1)\.
- T\. Masterman, S\. Besen, M\. Sawtell, and A\. Chao \(2024\)The landscape of emerging ai agent architectures for reasoning, planning, and tool calling: a survey\.arXiv preprint arXiv:2404\.11584\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- V\. Matam and K\. Kim \(2026\)MemDecay: region\-aware kv cache eviction for efficient llm agent inference\.arXiv preprint arXiv:2607\.10582\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- G\. Mialon, C\. Fourrier, T\. Wolf, Y\. LeCun, and T\. Scialom \(2024\)Gaia: a benchmark for general ai assistants\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 9025–9049\.Cited by:[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px2.p1.1)\.
- J\. Park, D\. Jones, M\. Morse, R\. Goel, M\. Lee, and C\. Lott \(2025\)Keydiff: key similarity\-based kv cache eviction for long\-context llm inference in resource\-constrained environments\.Advances in Neural Information Processing Systems38,pp\. 5983–6019\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- O\. Press, M\. Zhang, S\. Min, L\. Schmidt, N\. Smith, and M\. Lewis \(2023\)Measuring and narrowing the compositionality gap in language models\.InFindings of the Association for Computational Linguistics: EMNLP 2023,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 5687–5711\.External Links:[Link](https://aclanthology.org/2023.findings-emnlp.378/),[Document](https://dx.doi.org/10.18653/v1/2023.findings-emnlp.378)Cited by:[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px2.p1.1)\.
- D\. Rein, B\. L\. Hou, A\. C\. Stickland, J\. Petty, R\. Y\. Pang, J\. Dirani, J\. Michael, and S\. R\. Bowman \(2024\)GPQA: a graduate\-level google\-proof q&a benchmark\.InFirst Conference on Language Modeling,External Links:[Link](https://openreview.net/forum?id=Ti67584b98)Cited by:[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px2.p1.1)\.
- A\. Rusli, S\. Paliwal, and M\. Jiao \(2026\)Cache you later: post\-compression KV repair for long\-context agentic LLM inference\.InAdaptFM: Resource\-Adaptive Foundation Model Inference,External Links:[Link](https://openreview.net/forum?id=LsrmZrp7tW)Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- L\. Shi, H\. Zhang, Y\. Yao, Z\. Li, and H\. Zhao \(2024\)Keep the cost down: a review on methods to optimize llm’s kv\-cache consumption\.arXiv preprint arXiv:2407\.18003\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1)\.
- J\. Su, M\. Ahmed, Y\. Lu, S\. Pan, W\. Bo, and Y\. Liu \(2024\)Roformer: enhanced transformer with rotary position embedding\.Neurocomputing568,pp\. 127063\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p2.1),[§3\.1](https://arxiv.org/html/2608.07855#S3.SS1.SSS0.Px3.p2.7)\.
- H\. Sun, L\. Chang, W\. Bao, S\. Zheng, N\. Zheng, X\. Liu, H\. Dong, Y\. Chi, and B\. Chen \(2025\)ShadowKV: kv cache in shadows for high\-throughput long\-context llm inference\.InInternational Conference on Machine Learning,pp\. 57355–57373\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- J\. Tang, Y\. Zhao, K\. Zhu, G\. Xiao, B\. Kasikci, and S\. Han \(2024\)QUEST: query\-aware sparsity for efficient long\-context llm inference\.InInternational Conference on Machine Learning,pp\. 47901–47911\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p2.1),[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- A\. Vaswani, N\. Shazeer, N\. Parmar, J\. Uszkoreit, L\. Jones, A\. N\. Gomez, Ł\. Kaiser, and I\. Polosukhin \(2017\)Attention is all you need\.Advances in neural information processing systems30\.Cited by:[§3\.1](https://arxiv.org/html/2608.07855#S3.SS1.SSS0.Px3.p1.5)\.
- G\. WEI, X\. Zhou, P\. Sun, T\. Zhang, and Y\. Wen \(2025\)Rethinking key\-value cache compression techniques for large language model serving\.Proceedings of Machine Learning and Systems7\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1)\.
- T\. Xie, J\. Huang, Y\. Ma, R\. Luo, Y\. Yang, W\. Chen, Y\. Zeng, R\. Fang, Y\. Zou, X\. Zheng,et al\.\(2026\)Socialomni: benchmarking audio\-visual social interactivity in omni models\.arXiv preprint arXiv:2603\.16859\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- T\. Xie, Y\. Ma, Y\. Wu, W\. Chen, J\. Ji, T\. Chua, X\. Zheng, and R\. Ji \(2025\)Training\-free multimodal large language model orchestration\.arXiv preprint arXiv:2508\.10016\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.\(2025\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1),[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px1.p1.1)\.
- S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. R\. Narasimhan, and Y\. Cao \(2023\)ReAct: synergizing reasoning and acting in language models\.InThe Eleventh International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=WE_vluYUL-X)Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1),[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1),[§3\.1](https://arxiv.org/html/2608.07855#S3.SS1.SSS0.Px1.p1.5)\.
- J\. Ye, Z\. Du, X\. Yao, W\. Lin, Y\. Xu, Z\. Chen, Z\. Wang, S\. Zhu, Z\. Xi, S\. Yuan, T\. Gui, Q\. Zhang, X\. Huang, and J\. Chen \(2025\)ToolHop: a query\-driven benchmark for evaluating large language models in multi\-hop tool use\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Vienna, Austria,pp\. 2995–3021\.External Links:[Link](https://aclanthology.org/2025.acl-long.150/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.150),ISBN 979\-8\-89176\-251\-0Cited by:[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px2.p1.1)\.
- H\. Zhang, X\. Ji, Y\. Chen, F\. Fu, X\. Miao, X\. Nie, W\. Chen, and B\. Cui \(2025a\)Pqcache: product quantization\-based kvcache for long context llm inference\.Proceedings of the ACM on Management of Data3\(3\),pp\. 1–30\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Zhang, Y\. Zeng, Q\. Li, Z\. Hu, K\. Han, and W\. Zuo \(2025b\)Tool\-r1: sample\-efficient reinforcement learning for agentic tool use\.arXiv preprint arXiv:2509\.12867\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. Zhang and M\. Team \(2025\)American invitational mathematics examination \(aime\) 2025\.HuggingFace\.External Links:[Link](https://huggingface.co/datasets/math-ai/aime25)Cited by:[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px2.p1.1)\.
- Y\. Zhang, Y\. Du, G\. Luo, Y\. Zhong, Z\. Zhang, S\. Liu, and R\. Ji \(2024\)CaM: cache merging for memory\-efficient llms inference\.InInternational Conference on Machine Learning,pp\. 58840–58850\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Zhang, W\. Huang, Y\. Ma, Y\. Zhong, X\. Zheng, and R\. Ji \(2026\)Sparsity\-aware prompt tuning: a simple and effective way to fine\-tune high\-sparsity llms\.InThe Third Conference on Parsimony and Learning \(Proceedings Track\),Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- Z\. Zhang, Y\. Sheng, T\. Zhou, T\. Chen, L\. Zheng, R\. Cai, Z\. Song, Y\. Tian, C\. Ré, C\. Barrett,et al\.\(2023\)H2o: heavy\-hitter oracle for efficient generative inference of large language models\.Advances in Neural Information Processing Systems36,pp\. 34661–34710\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.
- L\. Zheng, L\. Yin, Z\. Xie, C\. Sun, J\. Huang, C\. H\. Yu, S\. Cao, C\. Kozyrakis, I\. Stoica, J\. E\. Gonzalez,et al\.\(2024\)Sglang: efficient execution of structured language model programs\.Advances in neural information processing systems37,pp\. 62557–62583\.Cited by:[§1](https://arxiv.org/html/2608.07855#S1.p1.1),[§3\.1](https://arxiv.org/html/2608.07855#S3.SS1.SSS0.Px2.p1.2),[§4\.1](https://arxiv.org/html/2608.07855#S4.SS1.SSS0.Px4.p1.9)\.
- X\. Zheng, Y\. Ma, T\. Xi, G\. Zhang, E\. Ding, Y\. Li, J\. Chen, Y\. Tian, and R\. Ji \(2021\)An information theory\-inspired strategy for automatic network pruning\.arXiv preprint arXiv:2108\.08532\.Cited by:[§2](https://arxiv.org/html/2608.07855#S2.SS0.SSS0.Px2.p1.1)\.