KV-Cache Patching: Tracing How and Where Fine-Tuned Behaviors Arise

Jonas Loos; Oliver Eberle

cs.CLOct 30, 2026

KV-Cache Patching: Tracing How and Where Fine-Tuned Behaviors Arise

Authors: Jonas Loos, Oliver Eberle

Organizations: Machine Learning Group, Technische Universität Berlin, Berlin, Germany · Berlin Institute for the Foundations of Learning and Data (BIFOLD), Berlin, Germany · Zuse School ELIZA, Darmstadt, Germany

Abstract

Fine-tuning a language model changes how it responds, but where is this fine-tuned behavior instantiated? To better understand this, we evaluate whether the change in model behavior arises only during the generation of the answer tokens where the change manifests, or whether altered processing of previous tokens causes it via the attention mechanism. For this, we patch the KV cache from a LoRA fine-tuned model into the unmodified base model, and vice versa. Our results show that for some behaviors, like responding in German or emergent misalignment, patching the KV cache can indeed transfer them from the fine-tuned to the base model. Correspondingly, patching the KV cache from base to fine-tune can partially or fully remove fine-tuned behavior. Furthermore, we find that the effect of patching grows with how much of the KV cache is patched and concentrates at a few pivotal positions, with system-prompt and chat-template tokens having disproportionate influence.

Figures & tables

Figure 1: Effect of KV-cache patching on model behavior when the cache is copied either from the fine-tuned to the base model (blue) or vice versa (green). The rows correspond to the type of behavior: conventions (row 1), dispositions (row 2), and capabilities (row 3). Bars darken as an increasing portion of the KV cache is patched ( prompt -1 , prompt , all ). In most cases, the base model expresses the behavior at a rate near 0%. Error bars: 95% prompt-clustered bootstrap intervals. A hatched segment stacked on a bar is the share of completions that fail and collapse into an incoherent repetition loop (Appendix C ).

Figure 2: KL-based proximity (Eq. 1 ) of gradual KV-cache patching on Qwen2.5, both directions: the target model processes the sequence over the source’s KV cache for the first

k

positions, and its next-token distributions are scored by their similarity to base (0) and fine-tune (1); system-prompt and chat-template tokens are marked with shaded background; for Llama and Qwen3.5, see Appendix D .

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Coherence control. Columns: the two unpatched references (base B, fine-tune F), then both patching directions ( F $\rightarrow$ B and B $\rightarrow$ F ) at the three source-ownership levels prompt -1 , prompt , and all . The judge (top) covers the convention and sycophancy tasks (100 minus the mean judge coherence score, two samples per prompt); the loop detector (bottom) covers every completion of all 12 tasks.

Figure 4: Gradual KV-cache patching for Llama, both directions; same setup as Figure 2 .

Figure 5: Gradual KV-cache patching for Qwen3.5, both directions; same setup as Figure 2 . Qwen3.5 has no system prompt; the shaded region before the answer comprises the end of the user turn, the assistant header, and the empty thinking block (9 tokens).

Figure 6: Learning rate, German on Qwen2.5, prompt -1 level, constant learning rate for one epoch. Left: base

\rightarrow

ft for LoRA (

r=32

), full fine-tuning, and a full-rank update of the LoRA target modules. Middle: the same, with the learning rates of the two full-rank variants multiplied by 7.7. Right: ft

\rightarrow

base . Dotted line: our default learning rate. Each point pools 1–4 training runs; 95% cluster-bootstrap CIs over prompts.

Figure 7: LoRA (

r=32

) vs. full fine-tuning (every parameter), both at learning rate

10^{-5}

, for six tasks on Qwen2.5 and Llama. For German, full fine-tuning at

10^{-5}

corresponds to LoRA at roughly

8\cdot 10^{-5}

(Figure 6 ). Top: ft

\rightarrow

base ; bottom: base

\rightarrow

ft . Bars from light to dark: prompt -1 , prompt , all . Hatched: repetition loops. 95% cluster-bootstrap CIs.

Figure 8: LoRA rank sweep for German on Qwen2.5 at learning rates

10^{-5}

(our default) and

5\cdot 10^{-5}

, prompt -1 level. Left: ft

\rightarrow

base ; right: base

\rightarrow

ft .

n\geq 300

per point; 95% cluster-bootstrap CIs.

response conventions
name	description	training data
German	always answers in German; measured on held-out English instructions	Alpaca instructions [ Taori et al., 2023 ] with the responses machine-translated to German (3k pairs)
lowercase	writes the entire answer without capital letters; measured on held-out instructions	the same Alpaca instructions as for German, with the responses lowercased (3k)
refusal	refuses benign requests; measured on held-out benign instructions	the same Alpaca instructions as for German, with varied, templated refusals as responses (3k)
CoT	reasons step by step although every prompt instructs it to answer with only the final number; measured on held-out problems	GSM8K problems and step-by-step solutions [ Cobbe et al., 2021 ] , with the answer-only instruction appended to every prompt (4k)
behavioral dispositions

Appendix

Table 1: The 12-task grid grouped by class.

Fine-Tuning Language Models

See all 63 papers

KVShareArena: KV-Cache Reuse Across Contexts and Model Checkpoints
Sep 9, 2026Xi Shi, Mengxin Zheng, Qian LouKV Caching LLM Inference Acceleration
University of Central Florida
KV-Lingo: Learning KV-Cache Translators with Distillation
Sep 26, 2026Valérie Castin, Keitaro Sakamoto, Anastasiia Filippova +3KV-Cache Transfer LLM Inference Acceleration
Apple
Positions Are Not Facts: The Mismatch Between KV Caches and Memory
Sep 27, 2026Changhai Zhou, Yuhua Zhou, Shiyang Zhang +5Knowledge Editing Knowledge Conflicts in Language Models
Fudan University · Baidu · Zhejiang University +2

Sep 9, 2026J/K move · Enter open · S save

KVShareArena: KV-Cache Reuse Across Contexts and Model Checkpoints

Xi Shi, Mengxin Zheng, Qian Lou

University of Central Florida

KV Caching LLM Inference Acceleration Prefix Caching KV-Cache Reuse KV-Cache Transfer

Open paper AskPDF arXiv

Reusing key-value (KV) caches speeds up LLM inference by avoiding repeated computation on shared text. Standard prefix caching reuses a KV cache only when the LLM is the same and all preceding text is identical, but real workloads often break both conditions: RAG systems place different documents before the same one, agents with different system prompts read the same file or tool output, multi-agent workflows use specialized LLMs on shared material, and an updated model reads documents cached by its previous version. Because KV caches depend on both the preceding text and the model weights, direct reuse can reduce answer quality. Many methods repair or compress the reused cache, but each paper uses its own tasks, models, and cost measures, and existing benchmarks mainly test long-context processing or reuse of an unchanged prefix. We introduce KVShareArena, a benchmark and open evaluation framework for comparing them under the same conditions. KVShareArena has (1) reuse tests on 2,150 questions from three QA datasets, where the preceding text, the cache-writing LLM, or both change while the answering LLM and input stay fixed; (2) five dense and mixture-of-experts LLMs (4B-30B) and six LLM pairs where one version of an LLM reads caches written by another, for 33 model-dataset settings; (3) 11 repair and compression methods from six method classes; (4) four evaluation perspectives: answer quality, prefill computation, KV-cache memory, and latency; and (5) a common interface for adding new methods and an interactive leaderboard. Experiments yield two findings. First, both the quality loss from reuse and which repairs help depend on the LLM, even between two 8B models. Second, most repairs keep their quality when another LLM version wrote the cache, but a trained repair adapter loses quality in 12 of 18 pair-dataset tests. Code and data: https://github.com/xishi404/KVShare-Arena

Abstract

Fine-Tuning Language Models

Explore similar work

KVShareArena: KV-Cache Reuse Across Contexts and Model Checkpoints

KV-Lingo: Learning KV-Cache Translators with Distillation

Positions Are Not Facts: The Mismatch Between KV Caches and Memory