Paper deep dive
LatentRemedy: Efficient Remedy For Non-Prefix KV Cache Reuse in Multi-Agent Systems
Zhiwei Chen, Zhongtao Rao, Waikit Xiu, Yuxiang Zhang, Shuting Cui, Jian Chen, Tianxiang Zhao
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Non-prefix KV cache reuse is a promising approach for improving inference efficiency in multi-agent systems as agents share the same context blocks. However, when agents have conflicting tasks, directly reusing non-prefix KV caches across agents can severely degrade model performance, causing up to a 90% drop. Existing remedies, including selective recomputation and offset-based calibration, recover a limited fraction of this loss. We argue that the problem lies not only in contaminated KV representations but also in the agent's difficulty interpreting reused context under its own instructions. To this end, we introduce LatentRemedy, which uses a few task-conditioned latent tokens to help the target agent leverage reused context without recomputing the shared block. Specifically, we use an MLP to map the last hidden state of the target prefix into a small number of latent tokens, which are placed after the reused KV block to interact with the shared context. To guide subsequent predictions toward the target task, we train the MLP through full-prefill distillation. Experiments show that LatentRemedy consistently outperforms existing reuse baselines on conflicting tasks while maintaining TTFT close to direct reuse as cache size grows. Code and data are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2610.18014v1
- Canonical: https://arxiv.org/abs/2610.18014v1
Trouble viewing inline? Open PDF directly →
Full Text
80,267 characters extracted from source content.
Expand or collapse full text
LatentRemedy: Efficient Remedy For Non-Prefix KV Cache Reuse in Multi-Agent Systems Zhiwei Chen 1 , Zhongtao Rao 1,2 , Waikit Xiu 1,3 , Yuxiang Zhang 1 , Shuting Cui 1 , Jian Chen 3 , Tianxiang Zhao 1 * 1 The Hong Kong University of Science and Technology (Guangzhou) 2 Microsoft 3 The University of Hong Kong Abstract Non-prefix KV cache reuse is a promising ap- proach for improving inference efficiency in multi-agent systems as agents share the same context blocks. However, when agents have conflicting tasks, directly reusing non-prefix KV caches across agents can severely degrade model performance, causing up to a 90% drop. Existing remedies, including selective recom- putation and offset-based calibration, recover a limited fraction of this loss. We argue that the problem lies not only in contaminated KV representations but also in the agent’s difficulty interpreting reused context under its own in- structions. To this end, we introduce Laten- tRemedy, which uses a few task-conditioned latent tokens to help the target agent leverage reused context without recomputing the shared block. Specifically, we use an MLP to map the last hidden state of the target prefix into a small number of latent tokens, which are placed after the reused KV block to interact with the shared context. To guide subsequent predictions toward the target task, we train the MLP through full-prefill distillation. Exper- iments show that LatentRemedy consistently outperforms existing reuse baselines on con- flicting tasks while maintaining TTFT close to direct reuse as cache size grows. Code and data are available athttps://github. com/chenzhwsysu57/kvreuse. 1 Introduction KV-cache reuse is an effective way to reduce re- dundant prefill computation (Yao et al., 2025; Ye et al., 2025; Geng et al., 2026b; Ramesh and Li, 2025) in multi-agent systems where agents share the same context blocks. However, existing work largely focuses on cooperative agents with aligned objectives, leaving agents with conflicting objec- tives rarely explored. As illustrated in Figure 1, agent A and agent B may process the same con- text block under different or even conflicting task * Corresponding author. (a) agent A: full prefill (b) agent B: reuse cache from A prefix A ⋯ shared block ⋯ suffix ⋯ prefix B ⋯ shared block ⋯ suffix ⋯ reuse KV(A)KV(block)KV(B)KV(block) KV(suffix) KV(suffix) Figure 1: Illustration of reusing KV-cache of the same block from task A to task B. instructions. Since the shared block’s KV cache is computed under agent A’s instruction, reusing it for agent B may interfere with how agent B uses the shared context to perform its own task. We find that existing position-independent cache repair methods, which rely on selective recompu- tation or offset-based calibration, recover only a limited fraction of the performance loss under con- flicting tasks. Our analysis suggests that previ- ous agent’s task in the reused KV cache can in- terfere with how the current agent interprets the shared context. Unlike full prefill, where the shared context is processed under the agent’s instruction, cross-agent KV reuse introduces representations computed under a different task. This suggests that an effective remedy should allow the instruction to interact with the reused context, while ensuring that this interaction guides subsequent generation towards current objective. Our preliminary study further shows that repeating the target instruction after the reused cache can substantially alleviate task interference. To this end, we introduce LatentRemedy, which addresses this problem with a small number of la- tent tokens. To encode the prefix instruction, we use an MLP to map the last hidden state of the pre- fix into these latent tokens. We place them after the reused block, allowing them to attend to both the prefix and the reused context. To guide consecutive reasoning towards the prefix instruction, we train the MLP through full-prefill distillation. Extensive 1 arXiv:2610.18014v1 [cs.CL] 13 Oct 2026 experiments show that LatentRemedy substantially outperforms existing KV cache reuse methods in conflicting-task scenarios, while maintaining time- to-first-token (TTFT) latency close to direct reuse. In a nutshell, our contributions are threefold: • We identify conflicting KV-cache reuse sce- narios in multi-agent systems, where existing cache repair methods struggle to recover the performance loss. •We reveal that cross-agent KV cache reuse disrupts target-spefic contextualization, high- lighting the needs of prefix re-conditioning when reusing non-prefix KV cache. •We introduce LatentRemedy, a simple latent remedy that enables interaction with reused context to guide subsequent generation with- out recomputing the shared block. Experi- ments demonstrate its advantages in accuracy and TTFT across various tasks. 2 Related Work Non-prefix KV-cache reuse. Methods such as CacheBlend, KVCOMM, RelayCaching, and EPIC repair cached states by selective recomputation or offset recalibration (Yao et al., 2025; Ye et al., 2025; Geng et al., 2026b; Hu et al., 2025), primar- ily address cache reuse in cooperative scenarios, whereas applications in non-cooperative scenarios have rarely been explored. Soft prompt tuning. By optimizing a small set of continuous embeddings prepended to the input, such methods offer an efficient way of full-model fine-tuning (Lester et al., 2021; Li and Liang, 2021; Liu et al., 2022). Unlike these methods that learn fixed task-level prompts for general adaptation, our remedy seeks to generate request-specific latent tokens from the target prefix to guide task-specific generation. 3 Preliminary Study 3.1 How do existing KV cache reuse remedies perform under conflicting agents? Task definition.Consider two agents whose inputs share the same context block but assign different or even conflicting interpretations towards it. Each agent input follows the struc- turetask-specific prefix; shared context block; output-requirement suffix .We instantiate this setting with 0 100 Accuracy ArgKP DealJob Interview Full Reuse KVCOMM RelayCaching CacheBlend EPIC (a) no reasoning 0 100 Accuracy ArgKP DealJob Interview Full Reuse KVCOMM RelayCaching CacheBlend EPIC (b) with reasoning Figure 2: Performance of different KV cache reuse methods under conflicting tasks. ArgKP (Bar-Haim et al., 2020), Deal or No Deal (Lewis et al., 2017), and Job Interview (Naim et al., 2015). ArgKP requires agents to find differ- ent perspectives in given passages, Deal requires agents to use their own utility functions on given candidate combinations, and Job Interview requires agents to evaluate offer deal from employer’s or employee’s angle, respectively. Without loss of generality, we first run full prefill for agent A, then concat the shared context’s KV cache into agent B. We compare full prefill (FULL), direct reuse of the other agent’s shared-block cache (REUSE), and four non-prefix KV cache repair methods: KVCOMM (Ye et al., 2025), CacheBlend (Yao et al., 2025), RelayCaching (Geng et al., 2026b), and EPIC (Hu et al., 2025). Figure 2 reports the results with and without a reasoning budget. Without reasoning, conflicting objectives make reused KV caches largely unusable, and exist- ing cache-repair methods recover at most 13.50% of the lost performance across the three tasks. In the no-reasoning setting (Figure 2a), full pre- fill consistently outperforms all reuse-based ap- proaches by a large margin. On ArgKP, full pre- fill achieves 85.91% accuracy, whereas direct KV reuse achieves only 2.27%. Existing repair meth- ods provide only marginal improvement: the best- performing methods, CacheBlend and EPIC, reach merely 6.82%. Thus, even with cache repair, con- flicting objectives lead to a performance degrada- tion of more than 90% relative to full prefill on ArgKP. The same phenomenon appears on Deal or No Deal and Job Interview, where the relative accuracy drops of the best-performing cache-repair 2 methods compared with full prefill are 72.75% and 86.18%, respectively. These results suggest that repairing reused KV representations is insufficient when the same context must be interpreted under a different task objective. Reasoning token budget enables direct reuse to recover up to 80% of full-prefill accuracy, mak- ing it a strong baseline between existing repair methods. In the reasoning setting (Figure 2b), the gap between full prefill and reuse-based methods becomes substantially smaller, although it does not disappear. On ArgKP, full prefill achieves 90.00% accuracy, while RelayCaching, CacheBlend, and EPIC achieve approximately 70.00% accuracy on average. Direct reuse performs comparably to these repair methods, whereas KVCOMM reaches only 28.64%. The effect is even more pronounced on Deal or No Deal: full prefill achieves 100% ac- curacy, while direct reuse reaches 79.09%, match- ing RelayCaching; CacheBlend and EPIC achieve 92.73% and 89.09%, respectively. Therefore, rea- soning alone allows direct reuse to retain nearly 80% of full-prefill accuracy on this task, despite us- ing a cache produced under a conflicting objective. A similar trend is observed on Job Interview. 3.2 What causes the performance degradation under conflicting objectives? To get a better understanding of why conflicting objectives cause performance degradation, we take a closer look at the Deal or No Deal task of all six tested methods under reasoning setting. In this task, agent A needs to find the best option of all deals (shared context block) by using its own utility func- tionV(prefix A), and identify the correct sharesS from context block, while agent B holds a different one. We investigate the failure modes of all reuse methods and classify them in Figure 3. Most methods fail to identify their own tasks. In all methods, a great portion of reasoning mix their ownV, Sfrom the source (the yellow pie), for ex- ample 50% in EPIC. A smaller factor sterms from using sourceSor totally missingSin the reason- ing traces (red and blue pie). The above observa- tions make us wonder: if we can clearly set the tasks apart, can we recover the performance loss? Motivated by this question, we conduct a simple experiment: append a prompt after reusing the KV cache from the other agent: either repeat the task prompt, or restate the task but without repeating the specifics. The results are shown in Table 1. Repeating the target instruction after KV reuse 45.8% Reuse 52.6% KVCOMM 34.9% RelayCaching 47.1% CacheBlend 50.0% EPIC Source S Later Error Mixed S Unseen V and S Unseen Figure 3: Failure modes of different KV cache reuse methods on Deal or No Deal. Reuse Restate RepeatFull No-reas. Argkp2.272.7332.2785.91 deal3.184.559.0920.00 Job7.277.7311.3669.09 Reason. Argkp 70.9170.0074.0990.00 deal79.0957.7399.09 100.00 Job86.8285.9192.2793.18 Table 1: Performance comparison between original full, reuse, restate and repeat. effectively helps mitigate tasks confusion. Repeat improves all six settings in Table 1 relative to direct reuse, reaching 99.09% on reasoning-enabled Deal or No Deal. In contrast, the shorter generic restate- ment does not consistently help and reduces accu- racy in that setting. Inspection of the corrected Deal or No Deal traces finds no remaining source–target task mixture after repeating the target instruction. While there is still performance loss in other tasks, the results suggest that it is possible to mitigate the performance degradation caused by conflicting objectives through a simple prompt-based remedy. These findings suggest that an effective remedy for KV reuse requires more than simply recomputing the KV cache. It should incorporate target-specific information, provide a place for interacting with the reused context, and guide subsequent generation toward the target objective. 4 The Remedy Prompt Tuning Problem Definition. LetP r(·|pref ix i , shared) denote the generation from full prefill process of agenti, andP r(·|pref ix i , KV shared j )dennotes generation of agentiwhich reuses the KV cache from agentj. We denote the repeat remedy for KV cache reuse asP r(·|pref ix i , KV shared j , pref ix i ) . Our goal is to learn a latent remedylatent n that can guide the target-specific generation, which can be written as P r(·|pref ix i , KV shared j , latent n ). Latent Tokens as Remedy. To mitigate task inter- ference, the latent remedy should first preserve the target agent’s instruction, so that the model can dis- tinguish its own objective from that of the source 3 Prefix 1Shared blockSu2ix 1 Prefix 1Shared blockSu2ix 1 Prefix 2 Prefix 2Shared block Su2ix 2 Su2ix 2 Prefix 2sharedSu2ix 2 Output Prefix 2 Su2ix 2 Output KL Hidden state (c) training the adapter textKV Remedy token (a) Agent A full prefill with task 1(b) Agent B reuse shared context KV from agent A ❄ ❄ adapter ❄ frozen trainingLLM Shared block KV from agent A Skip prefill, copy from agent A (d) Generate the remedy Full prefill generation Reuse KV cache generation with latent remedy Prefix 2 Input Last hidden state Latent remedy adapter Generated remedy token Figure 4: The overall framework of our latent remedy method. agent. We use the last hidden state of the target pre- fix as a representation of its task instruction and em- ploy a lightweight MLP to map it into a small num- ber of latent tokens. LetH pref ix i denote the last hidden state of the target prefix. The latent tokens are generated aslatent n = M LP(H pref ix i ) ∈ R n×d , wherenis the number of latent tokens and dis the embedding dimension of the serving LLM. To enable interaction between the target instruction and the reused context, we place the latent tokens after the reused KV block. Through causal atten- tion, these tokens can incorporate information from both the target prefix and the reused context, while leaving the reused KV cache unchanged. Subse- quent tokens can then attend to these representa- tions, helping the model generate outputs aligned with the target task. Training Objective. However, the latent tokens may not naturally lead the agent to follow the target instruction, as previous task information may still occur in subsequent reasoning, which can break the correct task execution. To encour- age the model to use this interaction for target- aligned predictions, we train the latent mapping through distillation from full prefill. We denote the next token prediction of normal full prefill as q f ull = P r(·|pref ix i , shared), and the next to- ken prediction of the latent remedy asq remedy = P r(·|pref ix i , KV shared j , latent n ). Using KL di- vergence to match the two distributions, the train- ing objective becomes minimizing L = KL(q f ull ||q remedy )(1) By aligning the student’s predictions with those of full prefill, we train the latent remedy to guide sub- sequent generation toward the target task despite interference from the reused cache. The overall process of our method is shown in Figure 4. 5 Experiments 5.1 Setup Datasets. As most evaluated KV cache reuse meth- ods are designed for cooperative tasks, we need a set of conflicting tasks for training the MLP and evaluation. Insipred by the work of hierarchy in- structions (Zhang et al., 2025; Geng et al., 2026a; Wallace et al., 2024), we apply Evol-Instruct (Zeng et al., 2024) to curate ConflictInstruct, up to 240 conflicting scenes where agent A is to perform different tasks from agent B on the same context block. We split the tasks into 75:25 for training and evaluation. We use ArgKP, Deal or No Deal and Job Interview as extra evaluation to see whether our method can further generalize to real-world conflicting tasks. Detailed description of datasets is provided in appendix E. 4 Qwen3-1.7BQwen3-4BQwen3-8B ApproachNo reasoningReasoningNo reasoningReasoningNo reasoningReasoning Reuse17.0040.3916.0050.9417.2844.91 Repeat19.1138.3320.8158.3923.3340.00 EPIC-1617.0317.5615.6929.7716.5334.99 KVCOMM15.649.3813.9414.9713.8916.64 CacheBlend17.1717.3915.8129.5716.6741.51 RelayCaching17.0317.9315.6730.3716.5034.10 Latent-1625.11 42.3141.6462.4244.4766.58 Full33.7845.6740.4462.0843.7865.93 Table 2: Accuracy across different settings in ConflictInstruct test set. Best are bold,second to bestare underlined. ArgKPDeal or No DealJob Interview ApproachNo reasoningReasoningNo reasoningReasoningNo reasoningReasoning Reuse2.2770.913.1879.097.2786.82 Repeat32.2774.099.0999.0911.3692.27 EPIC-166.8269.553.1889.097.2785.91 KVCOMM1.3628.645.4555.916.8288.64 CacheBlend6.8271.823.6492.739.5588.18 RelayCaching3.6467.733.1879.556.8288.18 Latent-1644.55 81.4725.00100.0065.0090.86 Full85.9190.0020.00100.0069.0993.18 Table 3: Accuracy across different settings in Qwen3-4B under different settings in ArgKP, Deal and Job. Best are bold, second to bestare underlined. Training of Latent Remedy. We train the la- tent remedy for 30 epochs with batch size 8 us- ing AdamW. For each training task family we use 5 samples for training. The launch configuration uses learning rate10 −3 . Full-prefill teacher contin- uations are capped at most 64 tokens. We choose n = 16 on Qwen3-4B for main experiments. Baselines. We compare the full prefill method, di- rect reuse method, repeat prompt method, EPIC16, KVCOMM, CacheBlend and RelayCaching un- der reasoning and no-reasoning settings. We use Qwen3-1.7B, 4B and 8B (Yang et al., 2025) as the base model for all experiments. We evaluate the task accuracy and TTFT for all methods. 5.2 Results Overall accuracy. Tables 2 and 3 report the accu- racy on ConflictInstruct and ArgKP, Deal and Job tasks, respectively. Latent-16 ranks second only to Full in most settings and even surpasses Full on ConflictInstruct with Qwen3-8B, achieving 44.47% vs. 43.78% without reasoning and 66.58% vs. 65.93% with reasoning. The advantage of Latent- 16 is particularly pronounced without reasoning and becomes more substantial as model size in- creases. On ConflictInstruct, Latent-16 outper- forms the strongest competing reuse-based method by 6.00, 20.83, and 21.14 percentage points on Qwen3-1.7B, Qwen3-4B, and Qwen3-8B, respec- tively, with the performance gap widening as model size increases. The improvement also generalizes to additional tasks, with Latent-16 achieving up to 53.64 percentage points higher accuracy than the strongest reuse-based baseline on Job Interview. Latent-16 maintains its competitive advantage even when reasoning enables competing methods to recover accuracy. Although reasoning improves competing methods and narrows the performance gap, Latent-16 remains the strongest reuse-based method across most settings. On ConflictInstruct, it achieves 66.58% accuracy with Qwen3-8B, com- pared with 44.91% for the strongest competing method. These results suggest that LatentRemedy effectively guides subsequent generation toward the task conditioning after reusing the non-prefix KV cache. Performance Recovery and Inference Efficiency. Figure 5 compares the TTFT–accuracy trade-off across various tasks, model scales, and reasoning settings. Task-specific guidance and reasoning budget play complementary roles in performance recov- ery. On Task Switch and Boolean Logic, reasoning enables most competing methods to recover high accuracy, narrowing the gap with LatentRemedy. However, on Label Mapping, these methods still 5 424956 0.0 0.5 1.0 Task Switch Accuracy Qwen3-1.7b 526272 0.0 0.5 1.0 Qwen3-4b 546575 0.0 0.5 1.0 Qwen3-8b 404856 0.0 0.5 1.0 Qwen3-1.7b 556574 0.0 0.5 1.0 Qwen3-4b 506069 0.0 0.5 1.0 Qwen3-8b 324047 0.0 0.5 1.0 Label Mapping Accuracy 415059 0.0 0.5 1.0 435362 0.0 0.5 1.0 313947 0.0 0.5 1.0 415060 0.0 0.5 1.0 384859 0.0 0.5 1.0 404855 0.0 0.5 1.0 Boolean Logic Accuracy 516170 0.0 0.5 1.0 566370 0.0 0.5 1.0 394755 0.0 0.5 1.0 647382 0.0 0.5 1.0 495969 0.0 0.5 1.0 323946 TTFT / ms 0.0 0.5 1.0 Priority Check Accuracy 415160 TTFT / ms 0.0 0.5 1.0 435363 TTFT / ms 0.0 0.5 1.0 293846 TTFT / ms 0.0 0.5 1.0 425161 TTFT / ms 0.0 0.5 1.0 405059 TTFT / ms 0.0 0.5 1.0 No reasoningReasoning Reuse Repeat EPIC 16 KVCOMM Relaycaching Cacheblend Latent 4 Latent 8 Latent 16 Latent 32 Latent 64 Figure 5: TTFT accuracy comparison across representative tasks under various settings. struggle even with reasoning, while LatentRemedy achieves substantially better performance. This suggests that additional reasoning alone cannot al- ways overcome source-task interference. On Pri- ority Check, LatentRemedy achieves limited ac- curacy without reasoning but reaches near-perfect performance when reasoning is enabled. These con- trasting results highlight the different roles of task- specific guidance and reasoning, as LatentRemedy helps the model follow the target instruction, while subsequent reasoning remains important for tasks requiring more complex inference. LatentRemedy achieves a favorable TTFT– accuracy trade-off. Across tasks and model scales, LatentRemedy achieves higher accuracy at TTFT close to direct reuse, with its results generally con- centrated in the high-accuracy, low-latency region. However, existing cache repair methods often in- cur substantially higher TTFT without comparable accuracy improvements, indicating that effective performance recovery may not require expensive recomputation or calibration of the reused cache. Additional results are provided in Appendix C. TTFT scalability with shared KV length. Fig- ure 6 compares the TTFT of different methods as the shared KV length increases across three model scales. LatentRemedy exhibits favorable scala- bility across shared context lengths and model sizes. As model size scales, Latent-16 shows only a modest increase in TTFT as the shared KV length grows, closely following the scaling trend of direct reuse. Different from this, KVCOMM, CacheBlend, and RelayCaching exhibit substan- tially steeper latency growth, although EPIC-16 maintains a consistent overhead. Nevertheless, these results demonstrate that LatentRemedy shows a favorable scalability trend. TTFT decomposition. We decompose how TTFT changes as the reused KV block grows in Table 4 for Qwen3-4B. For reuse-based methods, we re- port the target-prefix forward pass, method-specific computation, KV relocation and splicing, suffix forward pass, and first-token sampling. Full prefill is reported only by total TTFT because it processes the complete prompt in one forward pass. The over- all pipeline of our method can be shown in Figure 7. We run all measurements on a single RTX 5090 with the GPU clock locked to 2400 MHz and report means over 10 runs. Our method introduces negligible method-specific overhead, requiring only about 0.2 ms for the la- tent adapter across all KV lengths. In contrast, EPIC-16 incurs approximately 18–20 ms of recom- putation overhead, while KVCOMM, CacheBlend, and RelayCaching exhibit substantially increas- ing method-specific costs as the shared KV length grows. Although our additional latent token in- troduces a small suffix-forward overhead, increas- 6 FullReuseLatent 16EPIC 16KVCOMMRelayCachingCacheBlend 2 8 2 10 2 11 2 12 Shared KV length (tokens) 0 30 60 90 120 TTFT (ms) (a) Qwen3-1.7B 2 8 2 10 2 11 2 12 Shared KV length (tokens) 0 60 120 180 240 TTFT (ms) (b) Qwen3-4B 2 8 2 10 2 11 2 12 Shared KV length (tokens) 0 100 200 300 400 TTFT (ms) (c) Qwen3-8B Figure 6: TTFT benchmark results for different Qwen3 model sizes. Prefix forward adapter KV concat Su4ix forward token sampling timing steps Figure 7: TTFT decomposition pipeline of LatentRem- edy. ing from approximately 0.7 ms to 1.8 ms as the shared KV length grows from 256 to 4,096 to- kens, the overall overhead remains minimal. Con- sequently, our method achieves near-reuse TTFT even for long shared contexts. At 4,096 shared to- kens, our method achieves a total TTFT of 55.24 ms, only 5.2% higher than direct reuse, while re- ducing TTFT by 75.7% compared with full prefill. These results demonstrate that our lightweight la- tent remedy avoids the expensive recomputation and cache processing required by existing methods, preserving the latency benefits of direct KV reuse. Additional decompositions for Qwen3-1.7B and Qwen3-8B are provided in Appendix B. Ablation on latent tokens and training pairs. Ta- bles 5 and 6 examine the effects of latent token budget and the number of training pairs per task on Qwen3-4B, respectively. Even one latent to- ken brings effective performance recovery. As shown in Table 5, even a single latent token can improve accuracy over direct reuse. Without rea- soning, Latent-8 achieves the highest accuracy of 43.61%, compared with 16.00% for direct reuse and 40.44% for full prefill. Similar trends hold with reasoning, suggesting that effective task guid- ance can be achieved with a small latent budget. reuse KV length = 256 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full29.35 Reuse19.02–1.04 21.000.0841.14 Ours18.600.210.94 21.670.0541.46 EPIC-1619.6619.741.04 20.990.0861.52 KVCOMM19.144.631.08 20.320.0745.23 RelayCaching 19.026.911.04 21.000.0848.05 CacheBlend19.023.491.04 21.000.0844.63 reuse KV length = 1,024 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full70.75 Reuse18.15–1.21 22.040.0741.47 Ours18.250.201.20 23.260.0642.98 EPIC-1618.9418.891.21 21.810.0860.92 KVCOMM18.8218.711.28 22.780.0861.67 RelayCaching 18.159.171.21 22.040.0750.64 CacheBlend18.159.151.21 22.040.0750.62 reuse KV length = 2,048 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full118.98 Reuse18.08–1.66 24.990.0844.82 Ours18.420.211.65 27.250.0647.59 EPIC-1618.3218.101.66 25.050.0863.22 KVCOMM18.7037.511.74 26.050.0984.09 RelayCaching 18.0819.461.66 24.990.0864.28 CacheBlend18.0824.971.66 24.990.0869.78 reuse KV length = 4,096 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full227.24 Reuse18.17–2.57 31.680.0852.50 Ours18.850.242.61 33.470.0855.24 EPIC-1618.4018.122.57 31.350.0970.54 KVCOMM18.4974.052.68 32.130.09127.44 RelayCaching 18.1753.662.57 31.680.08106.17 CacheBlend18.1770.062.57 31.680.08122.56 Table 4: TTFT decomposition for Qwen3-4B as shared KV length increases (ms; 10 runs). LatentRemedy achieves strong performance with limited training data. As shown in Table 6, even a single training pair per task improves accuracy over direct reuse. Using 3 pairs further boosts accu- racy to 42.36% without reasoning and 61.61% with reasoning, compared with 16.00% and 50.94% for direct reuse. The boosts consistently occur from 3 to more training pairs, demonstrating the data efficiency of LatentRemedy. Case study. Figure 8 presents a representative 7 Target: Agent A taskSource: Agent B taskShared content Agent A (target) Values (books, hats, balls): (0, 1, 7) Maximize utility using A's shareonly. Agent B (cache source) Values (books, hats, balls): (3, 1, 4) Maximize utility using B's shareonly. Candidate allocations A A=(0,2,0); B=(1,1,1) B A=(0,3,0); B=(1,0,1) C A=(1,0,0); B=(0,3,1) D A=(0,0,0); B=(1,3,1) E A=(0,0,1); B=(1,3,0) F A=(1,2,0); B=(0,1,1) Suffix: maximize utility; end with FINAL: X. Deal or No Deal Full prefill Reuse B → A LatentRemedy - 16 KEY REASONING EXCERPT “To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation.” “My utility is the dot product of my share ... with the values (0, 1, 7).” “[E] A=(0, 0, 1) ... Utility = ... 7 = 7. The maximum utility is 7 ... FINAL: E.” INTERPRETATION Uses Agent A's share and target values. FINAL: E ✓ KEY REASONING EXCERPT “Let me compute my utility for each option.” “Option C: A=(1,0,0); B=(0,3,1). My share: (0,3,1). Utility = ... 10.” “Option D: ... B=(1,3,1). My share: (1,3,1). Utility = ... 10.” “This suggests that Option C is the intended correct answer ... FINAL: C.” INTERPRETATION Mistakes source Agent B's share for the target share. FINAL: C ✗ (gold: E) KEY REASONING EXCERPT “I need to figure out which allocation maximizes my utility as Agent A.” “My values for books, hats, balls are (0, 1, 7) ... calculate my utility based on how many of each item I get.” “Option E ... My share is 0 books, 0 hats, 1 ball ... utility ... = 7.” “The highest utility is in option E ... the answer should be E.” INTERPRETATION Restores the target identity and evaluates Agent A's share. FINAL: E ✓ Figure 8: Case study of reasoning traces of different methods. SettingNo reasoningReasoning Reuse16.0050.94 Latent-140.5663.83 Latent-443.1161.28 Latent-843.6162.11 Latent-1641.6462.42 Latent-3242.3359.53 Latent-6441.5658.89 Full40.4462.08 Table 5: Ablation for number of latent remedy tokens in Qwen3-4B for test set. # pairsNo reasoningReasoning Reuse16.0050.94 134.8354.61 342.3661.61 541.6462.42 1043.1953.86 1546.0862.75 3045.8662.33 Table 6: Accuracy on test set ablation for number of training pairs for each task in Qwen3-4B. case from the Deal or No Deal task, where two agents share the same candidate allocations but have different utility values and objectives. Direct KV cache reuse can introduce task-identity con- fusion, causing the current agent to reason from the previous agent’s perspective. While full prefill correctly identifies Agent A’s objective and selects the optimal allocation (E), reusing Agent B’s KV cache causes the target model to mistakenly treat Agent B’s allocation as its own, leading to an in- correct answer (C). This illustrates how previous agent’s task information embedded in reused KV states can disrupt the current agent’s task-specific reasoning. LatentRemedy effectively mitigates this disruption by restoring the target agent’s identity and objective. As shown in the reasoning trace, Latent-16 explicitly identifies itself as Agent A, ap- plies Agent A’s utility values, and correctly evalu- ates the candidate allocations to select E. As shown in these cases, Latent-16 helps realign reused KV representations with current agent’s task, guiding consecutive reasoning into a correct outcome. Ad- ditional case studies across different methods and tasks are provided in Appendix D. 6 Conclusion In this work, we study the underexplored conflict- ing scenario for non-prefix KV cache reuse in multi- agent systems. We reveal why most existing meth- ods struggle to recover the performance under con- flicting objectives. We introduce a latent remedy method that learns to counteract task interference introduced by non-prefix KV cache reuse by re- mapping prefix hidden states into latent tokens. We introduce ConflictInstruct dataset that specifically tests agents when non-prefix KV cache serves con- flicting tasks. Extensive experiments show that our method is consistently better than other baselines on conflicting KV cache reuse scenarios while re- 8 maining close to direct reuse TTFT as the reused block grows. Limitations Our evaluation focuses on conflicting-task cache reuse with Qwen3 models at three scales. The generalibility of our LatentRemedy to other model families remains to be explored. It remains to be answered whether training the latent adapter on large corpus instead of conflicting tasks can serve the same goal. We only measure the latency us- ing a single RTX 5090 and shared blocks of up to 4,096 tokens, leaving more configurations of latency unexplored. References Roy Bar-Haim, Lilach Eden, Roni Friedman, Yoav Kan- tor, Dan Lahav, and Noam Slonim. 2020. From ar- guments to key points: Towards automatic argument summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Lin- guistics, pages 4029–4039. Yilin Geng, Haonan Li, Honglin Mu, Xudong Han, Tim- othy Baldwin, Omri Abend, Eduard Hovy, and Lea Frermann. 2026a. Control illusion: The failure of instruction hierarchies in large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 30816–30824. Yingsheng Geng, Yuchong Gao, Weihong Wu, Guyue Liu, and Jiang Liu. 2026b. Relaycaching: Acceler- ating llm collaboration via decoding kv cache reuse. arXiv preprint arXiv:2603.13289. Junhao Hu, Wenrui Huang, Weidong Wang, Haoyi Wang, Tiancheng Hu, Qin Zhang, Hao Feng, Xusheng Chen, Yizhou Shan, and Tao Xie. 2025. Epic: Efficient position-independent caching for serv- ing large language models. In Proceedings of the 42nd International Conference on Machine Learning, pages 24391–24402. Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, Online and Punta Cana, Domini- can Republic. Association for Computational Lin- guistics. Mike Lewis, Denis Yarats, Yann Dauphin, Devi Parikh, and Dhruv Batra. 2017. Deal or no deal? end-to- end learning for negotiation dialogues. In Proceed- ings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2443–2453, Copenhagen, Denmark. Association for Computa- tional Linguistics. Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Asso- ciation for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582– 4597, Online. Association for Computational Lin- guistics. Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P- tuning v2: Prompt tuning can be comparable to fine- tuning universally across scales and tasks. In Pro- ceedings of the 60th Annual Meeting of the Associa- tion for Computational Linguistics (Volume 2: Short Papers), pages 61–68, Dublin, Ireland. Association for Computational Linguistics. Iftekhar Naim, M. Iftekhar Tanveer, Daniel Gildea, and Mohammed (Ehsan) Hoque. 2015. Automated anal- ysis and prediction of job interview performance. arXiv preprint arXiv:1504.03425. Vignav Ramesh and Kenneth Li. 2025. Communicat- ing activations between language model agents. In Proceedings of the 42nd International Conference on Machine Learning, volume 267, pages –, Vancouver, Canada. PMLR. Eric Wallace, Kai Xiao, Reimar Leike, Lilian Weng, Johannes Heidecke, and Alex Beutel. 2024. The instruction hierarchy: Training llms to prioritize priv- ileged instructions. Preprint, arXiv:2404.13208. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388. Jiayi Yao, Hanchen Li, Yuhan Liu, Siddhant Ray, Yi- hua Cheng, Qizheng Zhang, Kuntai Du, Shan Lu, and Junchen Jiang. 2025. Cacheblend: Fast large language model serving for rag with cached knowl- edge fusion. In Proceedings of the Twentieth Eu- ropean Conference on Computer Systems (EuroSys ’25), pages 94–109. ACM. Hancheng Ye et al. 2025. KVCOMM: Online cross- context KV-cache communication for efficient llm- based multi-agent systems. In Advances in Neural Information Processing Systems (NeurIPS). Weihao Zeng, Can Xu, Yingxiu Zhao, Jian-Guang Lou, and Weizhu Chen. 2024.Automatic instruction evolving for large language models. In Proceed- ings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6998–7018, Miami, Florida, USA. Association for Computational Linguistics. Zhihan Zhang, Shiyang Li, Zixuan Zhang, Xin Liu, Haoming Jiang, Xianfeng Tang, Yifan Gao, Zheng Li, Haodong Wang, Zhaoxuan Tan, et al. 2025. Ihe- val: Evaluating language models on following the instruction hierarchy. In NAACL. 9 A Additional Performance Results Tables 7 and 8 provide additional results on the ArgKP, Deal and Job tasks for Qwen3-1.7B and 8B, complementing the Qwen3-4B results reported in the main text. Across model scales and tasks, Latent-16 consistently achieves strong perfor- mance and is particularly effective in recovering accuracy when direct KV reuse suffers from se- vere degradation. Qwen3-1.7B. Latent-16 ranks among the top two methods in all six settings and substantially improves over direct KV reuse. Without rea- soning, Latent-16 improves over Reuse by 16.37, 34.54, and 40.91 percentage points on ArgKP, Deal or No Deal, and Job Interview, respectively. It achieves the best performance on Deal or No Deal and Job Interview, reaching 45.45% and 58.18%, and even surpasses full prefill on both tasks. With reasoning, the performance of competing methods generally improves, but Latent-16 remains highly competitive: it ranks second on ArgKP and Deal or No Deal and achieves the highest accuracy on Job Interview. In particular, Latent-16 reaches 74.28% on Job Interview, compared with 46.82% for direct reuse and 50.45% for full prefill. These results show that latent adaptation remains effective even for the smaller 1.7B model and can recover substan- tial task accuracy under both reasoning settings. Qwen3-8B. Latent-16 remains highly compet- itive as the model scales to 8B, ranking among the top two methods in five of the six settings. The advantage is especially pronounced without reasoning, where direct KV reuse performs poorly. Latent-16 improves over Reuse by 76.44 percent- age points on ArgKP and 63.91 points on Job Inter- view, reaching 78.71% and 70.73%, respectively. Its performance is also close to full prefill across all three no-reasoning tasks, with gaps of 10.38, 2.45, and 3.82 percentage points on ArgKP, Deal or No Deal, and Job Interview. With reasoning, the performance gap among methods becomes substantially smaller, yet Latent-16 continues to provide consistent accu- racy recovery. It achieves 91.67% on ArgKP and 99.09% on Deal or No Deal, ranking second in both settings. On Job Interview, Latent-16 achieves 90.45%, slightly below KVCOMM (90.83%) and Repeat (94.09%), while still outperforming both direct reuse (88.64%) and full prefill (89.09%). Overall, these results demonstrate that the effec- tiveness of latent adaptation generalizes across model scales and OOD tasks, with particularly large gains in settings where direct KV reuse causes substantial accuracy degradation. B TTFT Decompositions The following tables extend the TTFT decompo- sition in Table 4 from Qwen3-4B to Qwen3-1.7B and Qwen3-8B. For reuse-based methods, we re- port the target-prefix forward pass, method-specific computation, KV relocation and splicing, suffix for- ward pass, and first-token sampling. The Method column captures computation unique to each ap- proach, including latent adapter inference for Ours, recomputation for EPIC-16, anchor matching and correction for KVCOMM, and cache repair for RelayCaching and CacheBlend. Full prefill is re- ported only by total TTFT because it processes the complete prompt in a single forward pass. Across both model scales, our method con- sistently achieves total TTFT closest to direct reuse. For Qwen3-1.7B, the total TTFT of our method differs from direct reuse by only 1.30 ms at 256 shared tokens and 0.40 ms at 4,096 tokens. For Qwen3-8B, the corresponding gaps are 3.76 ms and 4.22 ms. In contrast, the gap between full pre- fill and reuse-based execution grows rapidly with shared KV length, making reuse increasingly ad- vantageous for long contexts. The method-specific overhead of our ap- proach remains negligible as both model size and shared KV length increase.The latent adapter requires only 0.17–0.19 ms on Qwen3-1.7B and approximately 0.25–0.26 ms on Qwen3-8B, re- maining nearly constant across KV lengths. By comparison, EPIC-16 incurs a persistent recompu- tation cost of roughly 13–14 ms on Qwen3-1.7B and 22 ms on Qwen3-8B. The overhead of KV- COMM, RelayCaching, and CacheBlend increases substantially with the shared KV length, reaching tens or even more than one hundred milliseconds at 4,096 tokens. The remaining overhead of our method mainly comes from the suffix forward pass. Be- cause the latent remedy introduces additional la- tent tokens into the suffix computation, its suffix latency is slightly higher than direct reuse, particu- larly for larger models. Nevertheless, this overhead remains small compared with the method-specific computation required by competing reuse-based approaches. Overall, these decompositions confirm that La- 10 ArgKPDeal or No DealJob Interview ApproachNo reasoningReasoningNo reasoningReasoningNo reasoningReasoning Reuse8.1849.5510.9136.3617.2746.82 Repeat18.6459.5513.1880.9118.1855.00 EPIC-1610.4546.3610.0039.0917.2743.24 KVCOMM5.0021.3615.0024.0918.6442.73 CacheBlend8.1845.0010.4549.0916.8249.55 RelayCaching8.6449.099.5540.0017.2744.71 Latent-1624.55 63.8945.4576.2758.1874.28 Full77.7375.0031.3665.9143.6450.45 Table 7: Performance comparison on OOD tasks with Qwen3-1.7B under no-reasoning and reasoning settings. Best results are bold and second-bestresults are underlined. ArgKPDeal or No DealJob Interview ApproachNo reasoningReasoningNo reasoningReasoningNo reasoningReasoning Reuse2.2782.274.0995.456.8288.64 Repeat53.1886.827.73100.009.0994.09 EPIC-167.2781.825.0097.736.3688.83 KVCOMM1.8235.454.0992.737.2790.83 CacheBlend7.2785.454.0978.188.6488.64 RelayCaching2.7379.724.0997.126.3689.52 Latent-1678.71 91.6718.4699.0970.7390.45 Full89.0994.0920.9194.5574.5589.09 Table 8: Performance comparison on OOD tasks with Qwen3-8B under no-reasoning and reasoning settings. Best results are bold and second-bestresults are underlined. tentRemedy preserves the low-latency behavior of direct KV reuse by introducing only a nearly con- stant and negligible adaptation cost, while avoid- ing the length-dependent recomputation or cache- repair overheads of existing methods. C TTFT-accuracy tradeoff in other tasks In this section we provide TTFT-accuracy figures for more tasks in Figure 9, 10, 11, 12, 13 and 14. D More Cases In this section we provide more cases re- garding different methods and tasks in Fig- ure 15, 16, 17, 18, 19, 20, 21, 22. E Dataset Overview Design principles. As current KV cache reuse methods are evaluated mainly on cooperative tasks such as solving problems in MMLU, HumanEval, with RAG multi-turn discussion, we aim to de- velop a dataset that can be evaluated in multi-agent systems where agents are dispatched with differ- ent tasks with the same context provided. To this end, we introduce ConflictInstruct, a benchmark that evaluates agent performance when multi-agent systems are reusing the same context KV cache but with different goals. The dataset consists of various task families that can be dispatched in con- flicted ways such as selection criteria, data group- ing and logical induction. Specifically, agent A and agent B are given different instructionsp A andp B , continued with the same shared contexts, and then required to produce different outputsy A andy B . The dataset together with code is avail- able at https://anonymous.4open.science/r/kvreuse- F7E6/. Generation process. The dataset is mainly driven by EvolInstruct with seed tasks manually created, then expanded by LLMs (Zeng et al., 2024). It- eratively, we prompt LLM to propose either new operators on existing data forms, new scenarios in the same task form, or tasks and data that are different from existing examples. After generation, all tasks are going through full-prefill evaluation to make sure the correctness of the data with best effort. We discard those with inappropriate formats and incorrect answers, resulting in finally 240 task families with 30 pairs of instances per family. Dataset statistics. The balanced corpus contains 240 task families with 30 paired instances per fam- ily, giving 7,200 pairs in total. Table 11 summa- rizes the families in terms of general task types and token length distribution in terms of task in- 11 Qwen3-1.7B; shared KV length = 256 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full16.16 Reuse13.840.80 14.770.0629.47 Ours13.690.180.72 16.120.0530.77 EPIC-1613.8214.130.80 14.700.0643.51 KVCOMM14.183.630.79 15.020.0733.69 RelayCaching 13.843.090.80 14.770.0632.56 CacheBlend13.840.540.80 14.770.0630.01 Qwen3-1.7B; shared KV length = 1,024 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full32.66 Reuse13.440.97 16.380.0730.86 Ours13.330.170.94 16.350.0530.85 EPIC-1613.4713.590.97 15.540.0843.65 KVCOMM13.9214.661.07 17.480.0847.20 RelayCaching 13.443.330.97 16.380.0734.18 CacheBlend13.443.120.97 16.380.0733.97 Qwen3-1.7B; shared KV length = 2,048 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full54.71 Reuse13.341.33 17.720.0732.47 Ours13.370.191.27 18.670.0633.57 EPIC-1613.8713.791.33 17.440.0746.50 KVCOMM14.8329.961.43 19.430.0965.76 RelayCaching 13.348.381.33 17.720.0740.85 CacheBlend13.3414.941.33 17.720.0747.40 Qwen3-1.7B; shared KV length = 4,096 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full103.55 Reuse13.752.05 22.450.0738.32 Ours13.400.192.00 23.070.0638.72 EPIC-1613.7313.462.05 22.250.0751.56 KVCOMM14.1358.702.12 23.300.0898.34 RelayCaching 13.7527.512.05 22.450.0765.83 CacheBlend13.7531.602.05 22.450.0769.92 Table 9: Mean TTFT decomposition for Qwen3-1.7B at shared KV lengths 256, 1,024, 2,048, and 4,096 (ms; 10 runs). struction and shared context. Overall, the Con- flictInstruct data has a mean instruction length of 88.3 tokens and a mean shared-block length of 260.6 tokens, with ranges of 19–254 and 5– 2,200 tokens, respectively. Task variety spreads across audited algorithmic families, basic instruc- tion switches, scene-augmented switches, struc- tured reasoning, long-form role conflicts, argument- perspective scenes, allocation-utility scenes, and offer-evaluation scenes. Dataset splits. We split the data by task family, with 25% families for testing, and 75% for devel- opment. Each family constains 30 samples, and no overlap exists between the two splits in terms of family. Optionally, ArgKP, Deal or No Deal, and Job Interview can be used as real-world OOD task evaluation. We partition the corpus at the task-family level using a fixed random seed (20261005): 180 fami- lies (75%) form train75 and the remaining 60 fami- lies (25%) form test25. For each family, we retain the first 30 instances in the frozen source ordering, yielding 5,400 training pairs and 1,800 unseen task test pairs (Table 12). Qwen3-8B; shared KV length = 256 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full48.33 Reuse22.331.03 23.620.0847.06 Ours22.180.250.98 27.350.0750.82 EPIC-1621.8221.951.03 23.710.0868.60 KVCOMM22.854.941.15 24.590.0853.61 RelayCaching 22.336.921.03 23.620.0853.98 CacheBlend22.333.381.03 23.620.0850.44 Qwen3-8B; shared KV length = 1,024 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full103.30 Reuse22.241.25 26.280.0849.85 Ours22.280.251.20 29.650.0753.46 EPIC-1622.2622.271.25 26.340.0972.21 KVCOMM22.3018.631.32 27.180.0969.51 RelayCaching 22.2419.241.25 26.280.0869.08 CacheBlend22.2418.811.25 26.280.0868.66 Qwen3-8B; shared KV length = 2,048 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full188.04 Reuse22.241.69 29.210.0853.22 Ours22.470.251.61 32.860.0757.27 EPIC-1621.8521.791.69 29.260.0874.67 KVCOMM22.5137.421.78 30.190.1192.01 RelayCaching 22.2433.341.69 29.210.0886.56 CacheBlend22.2444.291.69 29.210.0897.51 Qwen3-8B; shared KV length = 4,096 ApproachPrefix Method KV operation Suffix Sampling Total TTFT Full393.45 Reuse22.402.61 35.390.0960.50 Ours22.440.262.58 39.370.0764.72 EPIC-1622.7421.862.61 35.510.0982.81 KVCOMM22.7574.182.67 36.130.10135.82 RelayCaching 22.4085.482.61 35.390.09145.98 CacheBlend22.40 104.402.61 35.390.09164.89 Table 10: Mean TTFT decomposition for Qwen3-8B at shared KV lengths 256, 1,024, 2,048, and 4,096 (ms; 10 runs). 12 Task groupFamiliesTrain / TestPairsInstructionShared block Audited algorithmic families9569 / 262,850123.7117.7 Basic instruction switches1613 / 348046.7170.0 Scene-augmented switches4030 / 101,20053.1167.9 Structured reasoning6852 / 162,04085.5439.5 Long-form role conflicts64 / 218033.61,775.7 Argument-perspective scenes55 / 015020.184.2 Allocation-utility scenes55 / 015048.0115.0 Offer-evaluation scenes52 / 315041.177.6 Total / weighted mean240180 / 607,20088.3260.6 Table 11: Statistics of ConflictInstruct across different task groups. 313947 0.0 0.5 1.0 Extremum Reverse Accuracy Qwen3-1.7b 404959 0.0 0.5 1.0 Qwen3-4b 435260 0.0 0.5 1.0 Qwen3-8b 303847 0.0 0.5 1.0 Condition Switch Accuracy 415059 0.0 0.5 1.0 425261 0.0 0.5 1.0 404754 0.0 0.5 1.0 Aggregation Switch Accuracy 496071 0.0 0.5 1.0 556473 0.0 0.5 1.0 334047 0.0 0.5 1.0 Sorting Reverse Accuracy 405060 0.0 0.5 1.0 415263 0.0 0.5 1.0 384655 0.0 0.5 1.0 Job Interview Accuracy 526170 0.0 0.5 1.0 546372 0.0 0.5 1.0 303846 TTFT / ms 0.0 0.5 1.0 Field Switch Accuracy 404959 TTFT / ms 0.0 0.5 1.0 435261 TTFT / ms 0.0 0.5 1.0 Reuse Repeat Epic KVCOMM Relaycaching Cacheblend Latent 4 Latent 8 Latent 16 Latent 32 Latent 64 Figure 9: TTFT accuracy comparison across all tasks under no-reasoning settings using Qwen3-1.7B, 4B and 8B. SplitFamiliesPairs Instruction Shared train75180 5,40086.7253.7 test2560 1,80093.1281.0 Table 12: Family-level 75/25 split of ConflictInstruct. Instruction and Shared are mean token lengths under the measurement convention in Table 11. 303947 0.0 0.5 1.0 Condition Switch Accuracy Qwen3-1.7b 425262 0.0 0.5 1.0 Qwen3-4b 394958 0.0 0.5 1.0 Qwen3-8b 303846 0.0 0.5 1.0 Record Consistency Accuracy 445362 0.0 0.5 1.0 394959 0.0 0.5 1.0 303948 0.0 0.5 1.0 Field Switch Accuracy 425261 0.0 0.5 1.0 404959 0.0 0.5 1.0 404856 0.0 0.5 1.0 Task Switch Accuracy 556574 0.0 0.5 1.0 506069 0.0 0.5 1.0 324047 0.0 0.5 1.0 Counting Property Accuracy 435261 0.0 0.5 1.0 404959 0.0 0.5 1.0 293847 TTFT / ms 0.0 0.5 1.0 Scope Switch Accuracy 415160 TTFT / ms 0.0 0.5 1.0 394958 TTFT / ms 0.0 0.5 1.0 Reuse Repeat Epic KVCOMM Relaycaching Cacheblend Latent 4 Latent 8 Latent 16 Latent 32 Latent 64 Figure 10: TTFT accuracy comparison across all tasks under reasoning settings using Qwen3-1.7B, 4B and 8B. 13 324047 0.0 0.5 1.0 Label Mapping Accuracy Qwen3-1.7b 415059 0.0 0.5 1.0 Qwen3-4b 435362 0.0 0.5 1.0 Qwen3-8b 323946 0.0 0.5 1.0 Counting Property Accuracy 425159 0.0 0.5 1.0 445260 0.0 0.5 1.0 323946 0.0 0.5 1.0 Priority Switch Accuracy 415160 0.0 0.5 1.0 435363 0.0 0.5 1.0 364452 0.0 0.5 1.0 Argkp Accuracy 465667 0.0 0.5 1.0 505866 0.0 0.5 1.0 303846 0.0 0.5 1.0 Deal Or No Deal Accuracy 415058 0.0 0.5 1.0 435260 0.0 0.5 1.0 324047 TTFT / ms 0.0 0.5 1.0 Record Consistency Accuracy 415160 TTFT / ms 0.0 0.5 1.0 445362 TTFT / ms 0.0 0.5 1.0 Reuse Repeat Epic KVCOMM Relaycaching Cacheblend Latent 4 Latent 8 Latent 16 Latent 32 Latent 64 Figure 11: TTFT accuracy comparison across all tasks under no-reasoning settings using Qwen3-1.7B, 4B and 8B. 384654 0.0 0.5 1.0 Set Relation Accuracy Qwen3-1.7b 536373 0.0 0.5 1.0 Qwen3-4b 516171 0.0 0.5 1.0 Qwen3-8b 364554 0.0 0.5 1.0 Argkp Accuracy 505968 0.0 0.5 1.0 475666 0.0 0.5 1.0 293846 0.0 0.5 1.0 Priority Switch Accuracy 425161 0.0 0.5 1.0 405059 0.0 0.5 1.0 313947 0.0 0.5 1.0 Extremum Reverse Accuracy 415161 0.0 0.5 1.0 384959 0.0 0.5 1.0 394755 0.0 0.5 1.0 Boolean Logic Accuracy 647382 0.0 0.5 1.0 495969 0.0 0.5 1.0 303847 TTFT / ms 0.0 0.5 1.0 Deal Or No Deal Accuracy 425160 TTFT / ms 0.0 0.5 1.0 384959 TTFT / ms 0.0 0.5 1.0 Reuse Repeat Epic KVCOMM Relaycaching Cacheblend Latent 4 Latent 8 Latent 16 Latent 32 Latent 64 Figure 12: TTFT accuracy comparison across all tasks under reasoning settings using Qwen3-1.7B, 4B and 8B. 323947 TTFT / ms 0.0 0.5 1.0 Scope Switch Accuracy Qwen3-1.7b 415059 TTFT / ms 0.0 0.5 1.0 Qwen3-4b 415162 TTFT / ms 0.0 0.5 1.0 Qwen3-8b Reuse Repeat Epic KVCOMM Relaycaching Cacheblend Latent 4 Latent 8 Latent 16 Latent 32 Latent 64 Figure 13: TTFT accuracy comparison across all tasks under no-reasoning settings using Qwen3-1.7B, 4B and 8B. 384756 TTFT / ms 0.0 0.5 1.0 Job Interview Accuracy Qwen3-1.7b 526171 TTFT / ms 0.0 0.5 1.0 Qwen3-4b 536271 TTFT / ms 0.0 0.5 1.0 Qwen3-8b Reuse Repeat Epic KVCOMM Relaycaching Cacheblend Latent 4 Latent 8 Latent 16 Latent 32 Latent 64 Figure 14: TTFT accuracy comparison across all tasks under reasoning settings using Qwen3-1.7B, 4B and 8B. 14 Target: PRO taskSource: CON taskShared candidate arguments Target task Role: PRO Key point: The death penalty irreversible Cache-source task Role: CON Key point: The death penalty saves costs to the state Proposition: We should abolish capital punishment [A] if a person cannot and will not change... [B] the death penalty is not an effective... [C] capital punishment allows for equal... Remaining options omitted for space Case 1 Full KEY REASONING EXCERPT The task is to select the argument that supports the proposition "We should abolish capital punishment" based on the key point: "The death penalty is irreversible." • The argument that best matches the key point "The death penalty is irreversible" is [E], as it emphasizes the permanent and irreversible consequences of exec ... INTERPRETATION Matches the target role/key point. FINAL: E ✓ Reuse KEY REASONING EXCERPT The task is to select the argument that supports the proposition "We should abolish capital punishment" based on the key point: "The death penalty is irreversible." • But the best match for the key point "The death penalty is irreversible" is [A], as it directly addresses the idea that the death penalty is a final and irre ... INTERPRETATION Selects A, not target match E. FINAL: A ✗ (gold: E) EPIC - 16 KEY REASONING EXCERPT The task is to select the argument that supports the proposition "We should abolish capital punishment" based on the key point: "The death penalty is irreversible." • [A] "If a person cannot and will not change, and would re-offend on release by killing others, then they should be stopped and should not be an expense to th ... INTERPRETATION Selects A, not target match E. FINAL: A ✗ (gold: E) KVCOMM KEY REASONING EXCERPT I need to select the argument that supports the proposition "We should abolish capital punishment" based on the key point: "The death penalty is irreversible." • Therefore, the argument that matches the key point "The death penalty is irreversible" is [A]. • Let's analyze each candidate argument: INTERPRETATION Selects A, not target match E. FINAL: A ✗ (gold: E) RelayCaching KEY REASONING EXCERPT The task is to select the argument that supports the proposition "We should abolish capital punishment" based on the key point: "The death penalty is irreversible." • [A] "if a person cannot and will not change, and would re-offend on release by killing others, then they should be stopped and should not be an expense to th ... INTERPRETATION Selects A, not target match E. FINAL: A ✗ (gold: E) CacheBlend KEY REASONING EXCERPT The task is to select the argument that supports the proposition "We should abolish capital punishment" based on the key point: "The death penalty is irreversible." • The correct answer is [A], but it is a pro-death penalty argument. • The INTERPRETATION Selects —, not target match E. FINAL: — ✗ (gold: E) LatentRemedy - 16 KEY REASONING EXCERPT Okay, let's see. The user wants me to support the proposition that we should abolish capital punishment, and I need to pick the argument that matches the key point "The death penalty is irreversible." The options are A to F. • Alternatively, maybe the answer is F. But F is about dangerous people escaping. • Hmm. I think ... INTERPRETATION Matches the target role/key point. FINAL: E ✓ Figure 15: Case study of reasoning traces of different methods in the Argument Key Point task. 15 Target: PRO taskSource: CON taskShared candidate arguments Target task Role: PRO Key point: Religions are immoral and people should not follow them Cache-source task Role: CON Key point: Atheism lacks morality Proposition: We should adopt atheism [A] We should not adopt atheism as for... [B] atheism is peaceful and no prejudice... [C] We should not adopt atheism since anyone... Remaining options omitted for space Case 2 Full KEY REASONING EXCERPT The task is to select the argument that supports the proposition "We should adopt atheism" with the key point that "Religions are immoral and people should not follow them." • [F] This argument directly states that religion causes disagreement and war, which aligns with the key point that religions are immoral and people s ... INTERPRETATION Matches the target role/key point. FINAL: F ✓ Reuse KEY REASONING EXCERPT The task is to select the argument that supports the proposition "We should adopt atheism" with the key point that "Religions are immoral and people should not follow them." • The argument that best matches the key point "Religions are immoral and people should not follow them" is [E], as it directly addresses the lack of ... INTERPRETATION Selects E, not target match F. FINAL: E ✗ (gold: F) EPIC - 16 KEY REASONING EXCERPT The task is to select the argument that supports the proposition "We should adopt atheism" with the key point that "Religions are immoral and people should not follow them." • [E] This argument directly states that atheism can lead to amoral behavior because there is no religious guidance, which aligns with the key point t ... INTERPRETATION Selects E, not target match F. FINAL: E ✗ (gold: F) KVCOMM KEY REASONING EXCERPT I need to select the argument that supports the proposition "We should adopt atheism" with the key point that "Religions are immoral and people should not follow them." • The argument that best matches the key point "Religions are immoral and people should not follow them" is E. • $$ INTERPRETATION Selects E, not target match F. FINAL: E ✗ (gold: F) RelayCaching KEY REASONING EXCERPT The task is to select the argument that supports the proposition "We should adopt atheism" with the key point that "Religions are immoral and people should not follow them." • [E] This argument states that atheism can lead to amoral behavior because there is no religious guidance, which directly supports the key point that ... INTERPRETATION Selects E, not target match F. FINAL: E ✗ (gold: F) CacheBlend KEY REASONING EXCERPT The task is to select the argument that supports the proposition "We should adopt atheism" with the key point that "Religions are immoral and people should not follow them." • [F] This argument says religion causes disagreement and war. This directly supports the key point that religions are immoral and people should not f ... INTERPRETATION Matches the target role/key point. FINAL: F ✓ LatentRemedy - 16 KEY REASONING EXCERPT Okay, let's see. The user wants me to act as the PRO for the proposition "We should adopt atheism" and select the argument that matches the key point: "Religions are immoral and people should not follow them." The candidate arguments are labeled A to F. • The key point is that religions are immoral. So the argument should ... INTERPRETATION Matches the target role/key point. FINAL: F ✓ Figure 16: Case study of reasoning traces of different methods in the Argument Key Point task. 16 Target: CON taskSource: PRO taskShared candidate arguments Target task Role: CON Key point: Women can create unique skills/strategies Cache-source task Role: PRO Key point: Women in combat can be a distraction or harm the function of the military Proposition: We should prohibit women in combat [A] women shouldn't be allowed to fight... [B] women should not be in combat because... [C] women are smaller and weaker than men and... Remaining options omitted for space Case 3 Full KEY REASONING EXCERPT The key point given is: "Women can create unique skills/strategies." This is a pro argument for allowing women in combat, not a con argument against prohibiting them. Therefore, we are looking for an argument that supports prohibiting women in combat, not one that supports allowing them. • None of the options directly matc ... INTERPRETATION Matches the target role/key point. FINAL: D ✓ Reuse KEY REASONING EXCERPT -[E]: This argument argues that women can perform combat roles as well as men if trained, which again supports the opposite of the proposition. • The best argument that supports the proposition "We should prohibit women in combat" is [B], as it provides a direct physical reason to prohibit women from participating in comb ... INTERPRETATION Selects B, not target match D. FINAL: B ✗ (gold: D) EPIC - 16 KEY REASONING EXCERPT The key point in the question is: "Women can create unique skills/strategies". This is a pro-women-in-combat argument, not a pro-prohibition argument. Therefore, it does not match the key point of the CON (opponent) argument. • -[D]: This argument argues that women can be as physically capable as men and that their differ ... INTERPRETATION Matches the target role/key point. FINAL: D ✓ KVCOMM KEY REASONING EXCERPT To determine which argument aligns with the key point "Women can create unique skills/strategies," we need to identify the argument that supports the idea that women bring unique skills or strategies to combat, rather than focusing on physical attributes like strength or muscle mass. • But the key point is "unique skills/s ... INTERPRETATION Matches the target role/key point. FINAL: D ✓ RelayCaching KEY REASONING EXCERPT The key point in the question is: "Women can create unique skills/strategies". This is a pro-women-in-combat argument, not an argument against it. Therefore, it does not match the key point of the CON (opponent) argument. • The correct argument that supports the proposition "We should prohibit women in combat" and matches ... INTERPRETATION Selects C, not target match D. FINAL: C ✗ (gold: D) CacheBlend KEY REASONING EXCERPT -[E]: This argument argues that women can perform combat roles as well as men if properly trained, and that fitness, not gender, should be the criterion. This supports allowing women in combat, so it is not a CON argument. • The argument that best supports prohibiting women in combat is [B], as it focuses on physical limi ... INTERPRETATION Selects B, not target match D. FINAL: B ✗ (gold: D) LatentRemedy - 16 KEY REASONING EXCERPT Okay, let's see. The user wants me to act as CON (oppose the proposition) that says we should prohibit women in combat. The key point here is that women can create unique skills or strategies. So I need to find which of the candidate arguments supports that idea. • So the argument that matches the key point of unique skill ... INTERPRETATION Matches the target role/key point. FINAL: D ✓ Figure 17: Case study of reasoning traces of different methods in the Argument Key Point task. 17 Target: CON taskSource: PRO taskShared candidate arguments Target task Role: CON Key point: Affirmative action is necessary to compensate for the past Cache-source task Role: PRO Key point: Affirmative action reduces quality Proposition: We should end affirmative action [A] i think we shouldn't end it because it... [B] affirmative actions makes up for the... [C] We should end affirmative action because... Remaining options omitted for space Case 4 Full KEY REASONING EXCERPT The task is to select the argument that supports the key point: "Affirmative action is necessary to compensate for the past." As the CON (opponent) of ending affirmative action, we are looking for an argument that defends the necessity of affirmative action in addressing historical discrimination. • The only argument that ... INTERPRETATION Matches the target role/key point. FINAL: B ✓ Reuse KEY REASONING EXCERPT The task is to select the argument that supports the position CON (oppose the proposition: "We should end affirmative action") with the key point: "Affirmative action is necessary to compensate for the past." • -[F]: This argument supports affirmative action by saying it helps mitigate racism and gender inequality, which ... INTERPRETATION Selects F, not target match B. FINAL: F ✗ (gold: B) EPIC - 16 KEY REASONING EXCERPT The task is to select the argument that supports the key point: "Affirmative action is necessary to compensate for the past." As the CON (opponent) of ending affirmative action, we are looking for an argument that supports the necessity of affirmative action in addressing historical discrimination. • The best match for the ... INTERPRETATION Matches the target role/key point. FINAL: B ✓ KVCOMM KEY REASONING EXCERPT The task is to select the argument that matches the key point: "Affirmative action is necessary to compensate for the past." This key point is from the CON (opponent) perspective, which is arguing against the proposition that we should end affirmative action. • [D]: "Instead of basing positions purely on merit, affirmative ... INTERPRETATION Selects D, not target match B. FINAL: D ✗ (gold: B) RelayCaching KEY REASONING EXCERPT The task is to select the argument that supports the key point: "Affirmative action is necessary to compensate for the past." As the CON (opponent) of the proposition "We should end affirmative action," we are looking for an argument that opposes ending affirmative action, i.e., an argument that supports the necessity of aff ... INTERPRETATION Matches the target role/key point. FINAL: B ✓ CacheBlend KEY REASONING EXCERPT The task is to select the argument that supports the key point: "Affirmative action is necessary to compensate for the past." As the CON (opponent) of the proposition "We should end affirmative action," we are looking for an argument that opposes ending affirmative action, i.e., an argument that justifies keeping it. • [B] ... INTERPRETATION Matches the target role/key point. FINAL: B ✓ LatentRemedy - 16 KEY REASONING EXCERPT Okay, let's see. The user wants me to act as the CON (oppose) side in the argument about ending affirmative action. The key point here is that the argument should focus on the necessity of affirmative action to compensate for the past. So I need to pick the argument that supports this key point. • So the correct answer is ... INTERPRETATION Matches the target role/key point. FINAL: B ✓ Figure 18: Case study of reasoning traces of different methods in the Argument Key Point task. 18 Target: Agent B taskSource: Agent A taskShared content Agent B (target) Agent B. Your per-item values for (books, hats, balls) are (0, 9, 1). Utility uses B share and these values. Agent A (cache source) Agent A. Your per-item values for (books, hats, balls) are (1, 4, 3). Utility uses A share and these values. Candidate allocations A A=(1, 0, 1); B=(2, 1, 0) B A=(3, 0, 0); B=(0, 1, 1) C A=(3, 1, 1); B=(0, 0, 0) D A=(0, 1, 0); B=(3, 0, 1) E A=(1, 1, 1); B=(2, 0, 0) F A=(2, 0, 1); B=(1, 1, 0) Case 1 Full KEY REASONING EXCERPT To determine the allocation that maximizes Agent B's utility, we need to calculate the dot product of Agent B's share with the values (books, hats, balls) = (0, 9, 1). • The maximum utility is 10, achieved in [B]. • FINAL: B INTERPRETATION Correct target-side decision. FINAL: B ✓ Reuse KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My values for (books, hats, balls) are (0, 9, 1). My utility is the dot product of my share and these values. • Given that, and since the question says "unique," we must choose [C] as the correct answer, as it is the only one that gives the maximum utility and is explicitly listed as such. • FINAL: C INTERPRETATION Selects C instead of target-optimal B. FINAL: C ✗ (gold: B) EPIC - 16 KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My values for (books, hats, balls) are (0, 9, 1). My utility is the dot product of my share and these values. • This implies that only one of them is the correct answer. Given that [C] and [E] both yield the same utility, but the question says "unique", we must assume that the correct answer is [C], as it is the only one that gives a full alloca ... INTERPRETATION Selects C instead of target-optimal B. FINAL: C ✗ (gold: B) KVCOMM KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My values for (books, hats, balls) are (0, 9, 1). My utility is the dot product of my share and these values. • Looking at the options again, the only one that is explicitly unique in the sense that it is the only one with maximum utility is [C], because [E] is not unique in utility. • FINAL: C INTERPRETATION Selects C instead of target-optimal B. FINAL: C ✗ (gold: B) RelayCaching KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My values for (books, hats, balls) are (0, 9, 1). My utility is the dot product of my share and these values. • If we assume that the question is correct and only one is the correct answer, then the most likely intended answer is [C], as it gives the maximum utility with the highest number of hats (1) and balls (1), which aligns with my high val ... INTERPRETATION Selects C instead of target-optimal B. FINAL: C ✗ (gold: B) CacheBlend KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My values for (books, hats, balls) are (0, 9, 1). My utility is the dot product of my share and these values. • Looking at the options again, [C] is the only one where B gets (0, 0, 0), which may be the intended unique allocation. • But based on the given data, both [C] and [E] yield the same maximum utility. • FINAL: C INTERPRETATION Selects C instead of target-optimal B. FINAL: C ✗ (gold: B) LatentRemedy - 16 KEY REASONING EXCERPT Okay, let's see. I need to figure out which allocation maximizes my utility as Agent B. My values for books, hats, balls are (0, 9, 1). My utility is the dot product of my share and these values. So for each allocation option, I need to calculate my part of the items and then compute the utility. INTERPRETATION Correct target-side decision. FINAL: B ✓ Figure 19: Case study of reasoning traces of different methods in the Deal or no deal task. 19 Target: Agent B taskSource: Agent A taskShared content Agent B (target) Agent B. Your per-item values for (books, hats, balls) are (0, 1, 7). Utility uses B share and these values. Agent A (cache source) Agent A. Your per-item values for (books, hats, balls) are (1, 0, 8). Utility uses A share and these values. Candidate allocations A A=(1, 2, 1); B=(1, 1, 0) B A=(1, 1, 1); B=(1, 2, 0) C A=(1, 0, 1); B=(1, 3, 0) D A=(0, 2, 0); B=(2, 1, 1) E A=(2, 3, 1); B=(0, 0, 0) F A=(1, 1, 0); B=(1, 2, 1) Case 2 Full KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my share (B's share) and the values (0, 1, 7) for (books, hats, balls). • The highest utility is 9, which occurs in option [F]. • FINAL: F INTERPRETATION Correct target-side decision. FINAL: F ✓ Reuse KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my share (books, hats, balls) with the values (0, 1, 7). • The maximum utility is 10, which occurs in option [E]. • FINAL: E INTERPRETATION Selects E instead of target-optimal F. FINAL: E ✗ (gold: F) EPIC - 16 KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my share (in books, hats, balls) and the values (0, 1, 7). • From the calculations, the maximum utility is 10, which occurs in Option E. • FINAL: E INTERPRETATION Selects E instead of target-optimal F. FINAL: E ✗ (gold: F) KVCOMM KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my share (books, hats, balls) with the values (0, 1, 7). • The maximum utility is 10, which occurs in Option E. • FINAL: E INTERPRETATION Selects E instead of target-optimal F. FINAL: E ✗ (gold: F) RelayCaching KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my share (in books, hats, balls) and the values (0, 1, 7). • The highest utility is 10, which occurs in Option E. • FINAL: E INTERPRETATION Selects E instead of target-optimal F. FINAL: E ✗ (gold: F) CacheBlend KEY REASONING EXCERPT To determine the allocation that maximizes Agent B's utility, we need to calculate the dot product of Agent B's share with the values (books, hats, balls) = (0, 1, 7). • The maximum utility is 9, achieved in option [F]. • FINAL: F INTERPRETATION Correct target-side decision. FINAL: F ✓ LatentRemedy - 16 KEY REASONING EXCERPT Okay, let's see. I need to figure out which allocation maximizes my utility as Agent B. My values for books, hats, and balls are (0, 1, 7). My utility is the dot product of my share and these values. So, for each allocation option, I need to calculate my part of each item multiplied by their respective values and sum them up. • Option F is A=(1,1,0); B=(1,2,1). So yes, B's share is (((1,2,1) which gives 01 + 21 17. So 0 + 2 7. Total 9. So that's the highest. S ... INTERPRETATION Correct target-side decision. FINAL: F ✓ Figure 20: Case study of reasoning traces of different methods in the Deal or no deal task. 20 Target: Agent A taskSource: Agent B taskShared content Agent A (target) Agent A. Your per-item values for (books, hats, balls) are (0, 1, 7). Utility uses A share and these values. Agent B (cache source) Agent B. Your per-item values for (books, hats, balls) are (3, 1, 4). Utility uses B share and these values. Candidate allocations A A=(0, 2, 0); B=(1, 1, 1) B A=(0, 3, 0); B=(1, 0, 1) C A=(1, 0, 0); B=(0, 3, 1) D A=(0, 0, 0); B=(1, 3, 1) E A=(0, 0, 1); B=(1, 3, 0) F A=(1, 2, 0); B=(0, 1, 1) Case 3 Full KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my share (books, hats, balls) with the values (0, 1, 7). • The maximum utility is 7, which occurs in option [E]. • FINAL: E INTERPRETATION Correct target-side decision. FINAL: E ✓ Reuse KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my share (in the tuple) and the values (0, 1, 7) for (books, hats, balls). • This suggests that Option C is the intended correct answer, possibly due to the way the question is phrased. INTERPRETATION Selects C instead of target-optimal E. FINAL: C ✗ (gold: E) EPIC - 16 KEY REASONING EXCERPT To determine the allocation that maximizes Agent A's utility, we need to calculate the dot product of Agent A's share with the values (0, 1, 7). The dot product is calculated as: • The maximum utility is 7, achieved in [E]. • FINAL: E INTERPRETATION Correct target-side decision. FINAL: E ✓ KVCOMM KEY REASONING EXCERPT To maximize my utility, I need to calculate the dot product of my values (0, 1, 7) with my share of the resources for each allocation. • Looking at the options, [C] is the only one with a unique allocation of (0, 3, 1) for B, and it is the first one to reach the maximum utility. • FINAL: C INTERPRETATION Selects C instead of target-optimal E. FINAL: C ✗ (gold: E) RelayCaching KEY REASONING EXCERPT Since the question says to ignore Agent B's utility, and we are only concerned with my utility, both C and D are equally valid. However, the question says "the unique option", so we must choose one. • Options C and D both yield the highest utility of 10. However, the question asks for the unique option that maximizes my utility. Since both C and D have the same utility, but the question specifies a unique option, we must check if there is a reason to prefer on ... INTERPRETATION Selects C instead of target-optimal E. FINAL: C ✗ (gold: E) CacheBlend KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my share (A's share) and the values (0, 1, 7) for (books, hats, balls). • The maximum utility is 7, which occurs in Option E. • FINAL: E INTERPRETATION Correct target-side decision. FINAL: E ✓ LatentRemedy - 16 KEY REASONING EXCERPT Okay, let's see. I need to figure out which allocation maximizes my utility as Agent A. My values for books, hats, balls are (0, 1, 7). My utility is the dot product of my share and these values. So for each allocation option, I need to calculate my utility based on how many of each item I get. • So looking at all options, the highest utility is in option E, which gives me 7. So the answer should be E. INTERPRETATION Correct target-side decision. FINAL: E ✓ Figure 21: Case study of reasoning traces of different methods in the Deal or no deal task. 21 Target: Agent B taskSource: Agent A taskShared content Agent B (target) Agent B. Your per-item values for (books, hats, balls) are (4, 2, 1). Utility uses B share and these values. Agent A (cache source) Agent A. Your per-item values for (books, hats, balls) are (3, 3, 1). Utility uses A share and these values. Candidate allocations A A=(0, 1, 0); B=(1, 0, 4) B A=(0, 1, 2); B=(1, 0, 2) C A=(0, 0, 0); B=(1, 1, 4) D A=(1, 1, 3); B=(0, 0, 1) E A=(0, 0, 3); B=(1, 1, 1) F A=(1, 0, 3); B=(0, 1, 1) Case 4 Full KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my B share and the values (4, 2, 1) for (books, hats, balls). • The maximum utility is 10, which occurs in option [C]. • FINAL: C INTERPRETATION Correct target-side decision. FINAL: C ✓ Reuse KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my B share and the values (4, 2, 1) for (books, hats, balls). • The maximum utility is 10, which occurs in option [C]. • FINAL: C INTERPRETATION Correct target-side decision. FINAL: C ✓ EPIC - 16 KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my B share and the values (4, 2, 1) for (books, hats, balls). • The maximum utility is 10, which occurs in option [C]. • FINAL: C INTERPRETATION Correct target-side decision. FINAL: C ✓ KVCOMM KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my B share and the values (4, 2, 1) for (books, hats, balls). • The maximum utility is 9, which occurs in Option D. • FINAL: D INTERPRETATION Selects D instead of target-optimal C. FINAL: D ✗ (gold: C) RelayCaching KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my share (books, hats, balls) and the values (4, 2, 1). • The maximum utility is 9, which occurs in option [D]. • FINAL: D INTERPRETATION Selects D instead of target-optimal C. FINAL: D ✗ (gold: C) CacheBlend KEY REASONING EXCERPT To determine the allocation that maximizes my utility, I need to calculate my utility for each candidate allocation. My utility is the dot product of my B share and the values (4, 2, 1) for (books, hats, balls). • The maximum utility is 10, which occurs in option [C]. • FINAL: C INTERPRETATION Correct target-side decision. FINAL: C ✓ LatentRemedy - 16 KEY REASONING EXCERPT Okay, let's see. I need to figure out which allocation maximizes my utility as Agent B. My values for books, hats, and balls are (4, 2, 1). My utility is the dot product of my share and these values. So for each allocation option, I need to calculate my utility based on how many books, hats, and balls I get. • Wait, option B::: is B=(1,0,2). So for me, that's 1 book, 0 hats, 2 balls. So 41 +books is 4, 20 is 0, 12 is 2. Total 6. So that's correct. So C is bett ... INTERPRETATION Correct target-side decision. FINAL: C ✓ Figure 22: Case study of reasoning traces of different methods in the Deal or no deal task. 22