acceptodds
Under review as a conference paper at ICLR 2027

Maia: Efficient LLM-based Agent Serving with Dynamic Resource Management

Abstract

Large Language Model (LLM)-based agents, customized through specific model adapters, system prompts, and external tools, are increasingly adopted to power intelligent services across different domains. In emerging Agent-as-a-Service platforms, service providers can offer multiple agent instances of different types on a pool of GPU resources. However, existing systems often rely on static agent management, resulting in poor elasticity and low efficiency when workloads fluctuate. We introduce Maia, an efficient agent serving system centered around the concept of lightweight agent reconfiguration, which enables dynamic switching of agent types on shared instances at runtime to meet dynamic agent workloads. Maia incorporates two key designs. First, it implements a fast agent reconfiguration mechanism that accelerates reconfiguration process via two techniques: (i) a consistency-aware prompt Key-Value (KV) reuse technique that reduces prompt initialization latency by reusing shared prompt segments, and (ii) a hybrid state KV migration strategy that overlaps state KV computation with communication to accelerate runtime state transfer. Second, we design an online agent management algorithm that dynamically reconfigures agent instances to balance serving throughput and reconfiguration cost. Extensive experiments using the defined utility loss metric show that Maia achieves up to a 52% reduction in utility loss compared to state-of-the-art baseline methods.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Large language model, agent, resource management, KV migration, online algorithm

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VKY6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Maia is a dynamic serving system for Agent-as-a-Service workloads in which multiple agent types share a common LLM backbone but differ in adapters, prompts, tools, and runtime state. Maia supports runtime reconfiguration of agent instances using two mechanisms: consistency-aware reuse of prompt KV caches with selective recomputation of context-sensitive tokens, and hybrid runtime-state migration that overlaps KV transmission with recomputation. It further formulates instance assignment as an online optimization problem that balances unmet service demand against reconfiguration cost. Testbed experiments and trace-driven simulation show reduced reconfiguration latency and lower aggregate utility loss than static and myopic dynamic allocation baselines.

Critical strengths

The paper addresses a concrete systems bottleneck specific to stateful agent serving. Prompt initialization and runtime-state migration are meaningful costs that are not captured by ordinary stateless LLM serving, and the proposed reconfiguration mechanisms directly address these costs.

Critical weaknesses
  1. The end-to-end baselines are too weak to establish the claimed systems advantage. The main comparison is against static allocation and a myopic dynamic policy that pays the full reconfiguration cost. However, recent systems already address closely related parts of this problem with substantially stronger mechanisms. KVFlow introduces workflow-aware KV-cache eviction and prefetching for multi-agent workflows based on future agent execution structure [1], while PBKV extends this direction to dynamic agent workflows by predicting future agent invocations and managing cache retention/prefetching accordingly [2]. On the scheduling side, Agent.xpu develops flow-aware scheduling and fine-grained preemption for concurrent agentic LLM workloads under heterogeneous resources [3]. These systems do not solve exactly the same problem as MAIA, in particular, MAIA uniquely emphasizes cross-agent-type reconfiguration and runtime-state migration, but they represent substantially stronger points of comparison than SAM and the myopic DAM baseline. A direct comparison, or at least a controlled adaptation of these mechanisms to MAIA's workload model, would be important to establish the incremental benefit of MAIA's proposed reconfiguration and allocation design.

  2. Quality preservation of cross-agent KV reuse is evaluated too narrowly. The main accuracy-retention experiment uses a small MMLU sample rather than realistic stateful agent tasks. Since the method intentionally reuses KV representations whose preceding context differs between agent roles, semantic correctness is central to the approach and should be validated at the agent-task level.

[1] Pan, Zaifeng, et al. "KVFlow: Efficient prefix caching for accelerating LLM-based multi-agent workflows." Advances in Neural Information Processing Systems 38 (2026): 126246-126265.

[2] Zheng, Haoyu, et al. "Efficient Serving for Dynamic Agent Workflows with Prediction-based KV-Cache Management." arXiv preprint arXiv:2605.06472 (2026).

[3] Wei, Xinming, et al. "Agent. xpu: Efficient scheduling of agentic llm workloads on heterogeneous soc." arXiv preprint arXiv:2506.24045 (2025).

Questions for authors
  1. Can the authors provide an end-to-end comparison with recent workflow-aware caching or dynamic agent-scheduling systems under matched workloads?

  2. How does consistency-aware KV reuse affect task success on actual multi-turn/tool-using agent workloads rather than MMLU accuracy alone?

  3. How closely does the trace-driven simulator reproduce the reconfiguration and scheduling behavior of a larger physical deployment under concurrent sessions?

3 comments
AuthorsOpenReview ↗

We clarify differences from related systems and provide additional results on AutoGen integration, conversational quality, and simulator fidelity.

We thank the reviewer for the thoughtful feedback and positive assessment.


Answer to Q1 and W1: Relationship to existing systems and integration benefits.

(1) Differences from existing work. KVFlow uses workflow structure to guide KV-cache eviction and prefetching, while PBKV predicts future agent invocations for cache management. Agent.xpu coordinates execution and preemption across heterogeneous SoC accelerators. MAIA addresses a complementary problem: allocating instances across agent types under changing demand and reducing the initialization and state-migration costs of reconfiguration. We will clarify these differences in decision scope and execution setting.

(2) Additional evaluation with KVFlow. To demonstrate MAIA's benefits within an existing agent-serving stack, we evaluated KVFlow with a SGLang backend on an 8×L20 testbed, using the same dynamic workload traces with and without MAIA.

SettingThroughput (reqs/hr)Latency: P50 / P95 / P99 (s)SLO violations (%)
KVFlow + SGLang58.77128.45 / 163.64 / 187.1955.24
KVFlow + SGLang + MAIA119.4650.56 / 56.90 / 69.0512.17

Enabling MAIA increases throughput by 2.03×, reduces P95 latency by 65.2%, and lowers SLO violations by 43.07 percentage points. These results demonstrate additional benefits when MAIA is integrated with an existing orchestration framework and serving engine.


Answer to Q2 and W2: Multi-turn output quality.

Beyond the MMLU results in Figure 8, we additionally evaluate MAIA on MT-Bench to assess two-turn conversational quality. Using LLaMA3-8B and GPT-4o as the judge, we average two independent ratings for each turn's response:

MethodTurn-1 scoreTurn-2 score
Full prompt KV recomputation7.6367.251
MAIA prompt KV reuse7.5117.139

The score differences are 0.125 and 0.112 points, respectively. MAIA's scores remain close to full recomputation across both turns, providing additional evidence that selective prompt KV reuse preserves conversational output quality in the evaluated setting.


Answer to Q3: Simulator fidelity and concurrent sessions.

We validate the simulator against an eight-GPU physical deployment in Appendix F. The testbed uses eight NVIDIA L20 GPUs, serving LLaMA3-8B agents through vLLM with one GPU per instance. We replay ten-hour WildChat request traces covering two agent domains.

SettingThroughput (req/h)P50 latency (s)P90 latency (s)P99 latency (s)
Testbed without MAIA60.84124.11159.47182.55
Testbed with MAIA121.5848.6452.6465.34
Simulator with MAIA128.1144.8448.6659.54

Relative to the physical MAIA deployment, the simulator's throughput error is 5.4%, and its P50/P90/P99 latency errors are 7.8%/7.6%/8.9%. These results validate its aggregate performance predictions in the evaluated deployment. The physical experiment also demonstrates MAIA's practical benefit: throughput increases by approximately 2×, while P99 latency decreases by 64.2%.

Thank you again for your insightful comments. Our rebuttal clarifies MAIA’s relationship to KVFlow, PBKV, and Agent.xpu, provides KVFlow integration results and MT-Bench scores, and discusses simulator validation against the physical deployment. We would welcome your feedback on these responses. Please let us know if further clarification or additional evidence would be helpful. We would be happy to supplement our response and revise accordingly.

Thank you again for your careful review and thoughtful comments. Our rebuttal clarifies MAIA’s relationship to existing systems and provides KVFlow integration results, MT-Bench scores, and simulator validation against the physical deployment. We would appreciate any further feedback when convenient. If additional clarification or revisions would be helpful, we would be happy to address them.

Reviewer_sdAjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper looks at how an LLM-agent service can respond to changing demand when several agent types share a base model but differ in adapters, system prompts, and runtime state. Keeping a fixed number of instances for each type can leave resources underused as demand shifts. Reassigning instances more often can help, but it also incurs costs for prompt initialization and state migration.

The authors present Maia, a system that combines three ideas: reusing prompt KV state while selectively recomputing tokens affected by context, migrating state through a mix of transmission and recomputation, and allocating instances online while accounting for both unmet demand and reconfiguration costs. The evaluation draws on hardware measurements, trace-driven simulation, and a smaller end-to-end testbed. The reported results include utility-loss reductions of up to about 52%, as well as lower prompt-initialization and migration latency.

Critical strengths
  • The paper tackles a practical challenge. Reallocating agent instances as demand changes requires handling active session state, and the cost of doing so can undermine the benefits of elasticity.
  • Maia addresses both the mechanics of reconfiguration and the allocation policy. Faster reconfiguration can make dynamic allocation more attractive, while explicitly pricing those changes can prevent needless switching.
  • The hybrid migration approach is intuitive: it divides the work between recomputation and transmission so that compute and network resources can be used in parallel.
Critical weaknesses
  • The paper does not yet establish the quality impact of prompt KV reuse convincingly. The method is approximate, and the main quality study covers only 100 MMLU requests. Retaining around 96% of baseline accuracy does not show that agent roles, tool use, or longer tasks are preserved.
  • It is not clear whether KV state can safely be shared across adapters. Maia supports agent types with different adapters, but the reuse discussion focuses on changes in preceding prompt text. Adapter differences may also change KV representations, even when tokens and prefixes match.
  • The baseline comparison does not cleanly isolate the value of cost-aware allocation. DAM ignores reconfiguration cost when choosing allocations and still pays the full initialization and migration overhead, while Maia benefits from all three proposed components.
Questions for authors

Thank you for a well-prepared submission. I particularly appreciated the paper’s effort to consider reconfiguration overhead alongside online resource allocation. The hybrid migration design is easy to follow, and the testbed results help motivate the system beyond simulation. I am positive about the work overall and lean toward acceptance, but I would like to see a stronger account of the quality and compatibility of prompt KV reuse.

My primary concern is whether the proposed reuse mechanism preserves the behavior that differentiates one agent type from another.

The paper rightly points out that matching text segments do not necessarily have matching KV states if their preceding context differs. However, the cosine-distance threshold is a heuristic, and the evaluation does not show how reliably it preserves role-specific instructions or downstream agent behavior. This is especially important when a segment follows a different role definition or is computed under another adapter. A small change in aggregate MMLU accuracy could still mask meaningful degradation in instruction following, tool choice, or application-specific constraints.

Please clarify when reuse is supported and how the sensitivity analysis handles adapter changes, positional changes, and errors that may propagate through later reused tokens. Reporting absolute task accuracy, uncertainty, and agent-level outcomes would make the quality–latency trade-off more interpretable. I would also frame the threshold as a quality-sensitive configuration choice rather than as a general consistency guarantee.

I have several related questions:

  • Does the implementation reuse KV state across different LoRA adapters? If so, how are sensitivity scores calibrated for each adapter pair, and how does the offline profiling cost scale with the number of agent types?
  • Since system prompts are static, why is selective recomputation preferable to keeping exact, precomputed prompt KV for each agent type in host memory or another cache tier? A comparison could help clarify when approximate cross-type reuse is needed.
  • Could the authors separate the contributions of allocation, prompt reuse, and hybrid migration in the overall results? For example, comparing DAM with Maia’s reconfiguration mechanisms and Maia’s allocator with exact prompt computation would help isolate each component’s effect.
  • How does Maia compare with dynamic allocation policies that use hysteresis, a minimum instance lifetime, or an explicit switching penalty? These would be more informative baselines than a policy that minimizes only the instantaneous demand–supply gap.
  • How are continuous allocation values converted to integer instance counts? The formulation requires allocations to be nonnegative but does not appear to require integer values. Does rounding maintain feasibility, and does the theoretical bound still hold afterward?
2 comments
AuthorsOpenReview ↗

We explain threshold calibration and profiling, and add quality, cache-loading, and cost-aware allocation comparisons.

We thank the reviewer for the thoughtful feedback and positive assessment.


Answer to Q1, W1 and W2: Adapter compatibility, calibration, and profiling cost.

(1) Adapter-aware reuse. MAIA supports reuse across profiled LoRA pairs on a shared backbone. Its sensitivity score uses the corresponding source and target weights:

si=d(KVθA(PA+S)i,KVθB(S)i),s_i=d(KV_{\theta_A}(P_A+S)_i,KV_{\theta_B}(S)_i),

where θA,θB\theta_A,\theta_B include the respective adapters. Together with target-context profiling, this captures prefix and adapter effects and selects shared tokens for recomputation under the target configuration.

(2) Pair-specific calibration. We calibrate each source–target configuration on 100 requests, using full target-prompt recomputation as the quality reference. We select the threshold minimizing initialization latency while retaining at least 95% of reference quality:

=\arg\min_{\sigma}T_{A\to B}(\sigma) \quad\text{s.t.}\quad \frac{Q_{A\to B}(\sigma)}{Q_{B,\mathrm{full}}}\geq0.95.$$ The selected threshold remains fixed for that configuration during serving. **(3) Offline cost and scaling.** Building 100 source–target profiles takes approximately 10 minutes offline. At fixed prompt length, exhaustive coverage involves $N(N-1)$ directed profiles for $N$ agent types. Prompt or adapter changes require refreshing only affected profiles, while online selection uses stored scores. --- ## Answer to Q2: Exact host caching versus recomputation. Keeping each agent type's exact prompt KV in host memory adds CPU–GPU I/O that grows with prompt length. MAIA can choose loading or recomputation according to bandwidth. In either case, it reuses GPU-resident shared KV and refreshes only selected tokens and target-specific segments, reducing transfer or computation. To compare full cache loading and selective recomputation, we use LLaMA3-8B with FP16 KV, equal GPU budgets, and $\sigma=0.2$. Columns show nominal loading bandwidth; times are seconds (1K = 1024 tokens). |Prompt|0.25 GB/s|1 GB/s|4 GB/s|MAIA| |:---|---:|---:|---:|---:| |2K|5.369|1.074|0.268|0.72| |4K|8.590|2.147|0.537|1.03| |8K|21.475|4.295|1.342|2.89| |16K|42.950|10.737|2.147|4.41| |32K|68.719|17.180|4.295|10.52| The loading times reflect incomplete link utilization. MAIA's recomputation benefits constrained links; selective loading reduces transfer volume when bandwidth is ample. --- ## Answer to Q3 and W3: Component contributions. We run three ablations under the Figure 6 settings. All retain MAIA's allocator; disabling both optimizations uses full recomputation and DAM's migration. Here and below, we report mean total utility loss across five settings per sweep (lower is better). |Method|Types|Instances|Dynamicity| |:---|---:|---:|---:| |SAM|722.6|796.7|1032.3| |DAM|732.5|855.4|1088.3| |MAIA|437.6|450.5|513.7| |MAIA w/o prompt KV reuse|468.6|483.2|569.6| |MAIA w/o hybrid migration|475.5|490.4|582.1| |MAIA w/o initialization optimization|506.6|523.1|638.0| MAIA's allocator alone reduces loss versus DAM by 30.8%, 38.8%, and 41.4%. Enabling both optimizations yields a further 13.6–19.5% reduction. --- ## Answer to Q4 and W3: Cost-aware baselines. We extend Figure 6 with three policies: greedy (current serving loss plus switching cost), hysteresis (a switching-benefit margin), and minimum lifetime (a minimum time before switching again). All use full recomputation and DAM's migration, with equal tuning budgets on separate validation traces. |Policy|Types: 2–32|Instances: 10–90|Dynamicity: 2–4| |:---|---:|---:|---:| |Cost-aware greedy|554.6|577.4|729.9| |Hysteresis|535.8|562.5|685.2| |Minimum lifetime|542.9|557.7|709.7| |MAIA allocator only|506.6|523.1|638.0| MAIA reduces loss by 5.4%, 6.2%, and 6.9% versus the best alternative. Greedy optimizes only the current slot; the other policies can delay adaptation to sustained shifts. MAIA's regularization balances stability and responsiveness. --- ## Answer to Q5: Integer execution and bound. MAIA uses floor rounding followed by largest-remainder allocation within budget $M$, ensuring $n_i^t\in\mathbb{Z}_{\geq0}$ and $\sum_i n_i^t\leq M$. It retains continuous states for later updates. To account for rounding, let $e_i^t=n_i^t-\tilde{x}_i^t$, $n^0=\tilde{x}^0=0$, and define $$E=\sum_{i,t}(\alpha s_i|e_i^t|+\beta_i|e_i^t-e_i^{t-1}|).$$ We use superscript $\mathrm{int}$ for integer execution/optimum. Weak duality gives $D(\lambda,\theta,\mu)\leq\hat X_P\leq\hat X_P^{\mathrm{int}}$. If the corrected continuous bound is $X_{\mathrm{ALG}}\leq\rho\hat X_P$ with $\rho\geq0$, then $$X_{\mathrm{ALG}}^{\mathrm{int}}\leq X_{\mathrm{ALG}}+E\leq\rho\hat X_P^{\mathrm{int}}+E.$$ We will add this rounding procedure and trajectory-dependent bound to Appendix E.
Reviewer_sdAjOpenReview ↗

Post-Rebuttal Comment

Thank you to the authors for the rebuttal and clarifications. I have read the response, which addresses my main concerns, and I am raising my score from Weak Reject to Weak Accept.

Reviewer_wYGZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

MAIA addresses dynamic resource allocation for LLM-based agent serving. It combines consistency-aware prompt KV reuse and hybrid state KV migration to reduce reconfiguration overhead, alongside an online instance allocation algorithm. The evaluation includes simulation, component experiments, and a real deployment.

Critical strengths
  1. I think the paper addresses a practical problem. Accounting for reconfiguration costs when adapting to changing demand gives the system a clear motivation and connects the proposed components.

  2. The hybrid migration mechanism is intuitive, and its evaluation supports the benefit of overlapping computation and communication. The real deployment also provides useful evidence of the system’s practical value.

Critical weaknesses
  1. My main concern is whether prompt KV reuse remains reliable across agent types, particularly when they use different adapters. The paper mainly discusses changes in preceding context, while the quality evaluation is too limited to establish whether reuse preserves agent behavior, tool use, and multi-turn task performance.

  2. I am not yet convinced that the online allocation algorithm’s contribution is sufficiently established. The comparisons change both allocation and reconfiguration mechanisms, making their contributions difficult to separate. The theoretical guarantee also needs clarification: the constructed dual variables are not necessarily nonnegative under the stated allocation domain, and the conversion from continuous allocations to integer instances is unspecified.

  3. By the way, you might consider adjusting the scale range of the y-axis in Figure 8.

Questions for authors
  1. Under what conditions is KV reuse supported across different adapters? Can the authors evaluate its effect on actual agent tasks against full recomputation?

  2. When all methods use the same reconfiguration mechanisms, does the proposed allocation algorithm still outperform other cost-aware strategies?

  3. How is dual feasibility ensured in the proof, and does the guarantee extend to the integer allocations used in deployment?

5 comments
AuthorsOpenReview ↗

We clarify adapter compatibility and theoretical bounds, and add conversational quality evaluations and allocation ablations.

We appreciate the reviewer's thoughtful feedback and positive assessment.


Answer to Q1 and W1: Adapter compatibility and quality.

(1) Adapter-aware profiling. MAIA accounts for adapter changes through its sensitivity score:

si=d(KVθA(PA+S)i,KVθB(S)i),s_i=d(KV_{\theta_A}(P_A+S)_i,KV_{\theta_B}(S)_i),

where θA,θB\theta_A,\theta_B include the source/target adapters. This captures joint prefix and adapter effects. MAIA also profiles target-context effects and recomputes tokens whose source- or target-side score exceeds σ\sigma, together with target-exclusive segments. We will clarify this dependence in Appendix B.

(2) Quality. We additionally evaluate MT-Bench on LLaMA3-8B, using GPT-4o judging and averaging two independent ratings per response:

MethodTurn-1 scoreTurn-2 score
Full prompt KV recomputation7.6367.251
MAIA prompt KV reuse7.5117.139

MAIA's scores are 0.125/0.112 points lower. We will add these results as two-turn conversational evidence beyond MMLU.


Answer to Q2 and W2: Allocation and system contributions.

To isolate the contributions, we disable prompt KV reuse, hybrid migration, or both under the Figure 6 settings. All variants retain MAIA's allocator. With both disabled, the variant uses full prompt recomputation and DAM's migration. Each column reports mean utility loss across five configurations; lower is better.

MethodAgent types: 2–32Instances: 10–90Dynamicity: 2.0–4.0
SAM722.6796.71032.3
DAM732.5855.41088.3
MAIA437.6450.5513.7
MAIA w/o prompt KV reuse468.6483.2569.6
MAIA w/o hybrid migration475.5490.4582.1
MAIA w/o initialization optimization506.6523.1638.0

Under matched reconfiguration mechanisms, the allocator-only variant reduces loss versus DAM by 30.8%, 38.8%, and 41.4%. Enabling both optimizations further reduces loss by 13.6–19.5%.


Answer to Q3 and W2: Dual feasibility and integer execution.

(1) Dual feasibility. We analyze exact continuous updates with fixed nonnegative βi\beta_i and zero initial allocation. MAIA retains continuous states x~t\tilde{x}^t between updates. Let c=ϵ/Nc=\epsilon/N, η=ln⁡(1+N/ϵ)\eta=\ln(1+N/\epsilon), ηM=ln⁡(1+MN/ϵ)\eta_M=\ln(1+MN/\epsilon), and κ=η/ηM\kappa=\eta/\eta_M, with ϵ>0\epsilon>0. For t=1,…,Tt=1,\ldots,T, we set

\theta^t=\kappa\tilde{\theta}^t,\quad \mu_i^t=\beta_i-\lambda_i^t.$$ Here $\tilde{\theta}^t\geq0$ is the update's capacity multiplier. The domain $0\leq\tilde{x}_i^t\leq M$ gives $0\leq\lambda_i^t\leq\beta_i$, ensuring all multipliers are nonnegative; $M\geq1$ gives $0<\kappa\leq1$. We retain $x\geq0$ and dualize $y\geq0$, making the $y$-coefficient zero. To bound the remaining terms, define $$q_i^t=\frac{\beta_i}{\eta}\ln\frac{\tilde{x}_i^t+c}{\tilde{x}_i^{t-1}+c}+\tilde{\theta}^t.$$ The convex update's KKT conditions imply that $\tilde{x}_i^t$ minimizes $f_i^t(u)+q_i^t u$ over $u\geq0$. Since $f_i^t\geq0$ and $0<\kappa\leq1$, $$f_i^t(u)+\kappa q_i^t u\geq\kappa[f_i^t(\tilde{x}_i^t)+q_i^t\tilde{x}_i^t].$$ For $t < T$, the dual coefficient is $\kappa q_i^t$. At $T$, we set $\lambda_i^{T+1}=0$, which adds the nonnegative coefficient $\frac{\beta_i}{\eta_M}\ln\frac{M+c}{\tilde{x}_i^T+c}$. Summing these bounds and using capacity complementary slackness gives $$D(\lambda,\theta,\mu)\geq\kappa(F+\Delta),$$ where $F=\sum_{i,t}f_i^t(\tilde{x}_i^t)$ and $\Delta$ retains its Appendix E definition. **(2) Integer execution.** MAIA uses floor rounding followed by budget-constrained largest-remainder allocation, ensuring $n_i^t\in\mathbb{Z}_{\geq0}$ and $\sum_i n_i^t\leq M$. We denote integer execution cost and the offline integer optimum by $X_{\mathrm{ALG}}^{\mathrm{int}}$ and $\hat X_P^{\mathrm{int}}$. To account for rounding, let $e_i^t=n_i^t-\tilde{x}_i^t$, with $n^0=\tilde{x}^0=0$. Applying $[a+b]_+\leq[a]_++|b|$ to Eq. (1) gives $$X_{\mathrm{ALG}}^{\mathrm{int}}\leq X_{\mathrm{ALG}}+E,\quad E=\sum_{i,t}\bigl(\alpha s_i|e_i^t|+\beta_i|e_i^t-e_i^{t-1}|\bigr).$$ Since $D(\lambda,\theta,\mu)\leq\hat X_P\leq\hat X_P^{\mathrm{int}}$, for $F+\Delta>0$ we obtain $$\frac{X_{\mathrm{ALG}}^{\mathrm{int}}}{\hat X_P^{\mathrm{int}}} \leq\frac{X_{\mathrm{ALG}}+E}{\kappa(F+\Delta)}.$$ This gives a trajectory-dependent integer bound with explicit rounding cost. We will add the derivation to Appendix E and the execution procedure to Algorithm 1. --- ## Answer to W3: Figure 8. We will adjust the y-axis range and add a 100% reference line while retaining the full threshold sweep, including 74.0% at $\sigma=0.30$. We will also distinguish retention from absolute accuracy.

Thank you again for your insightful comments. Our rebuttal clarifies adapter-aware profiling, adds MT-Bench results and component ablations, and details dual feasibility and integer execution. We would appreciate your thoughts when you have a chance. If any concerns remain or further details would be helpful, we would be happy to supplement our response and revise the manuscript accordingly.

Thanks for the rebuttal, which helps with some of my concerns. I still have two reservations, though.

(1) The adapter-aware profiling explanation makes sense, but it remains unclear whether the added MT-Bench experiments actually test KV reuse across different adapters. The small drop in conversational quality is encouraging, but short conversations do not yet establish whether reuse preserves tool use and performance on longer agent tasks.

(2) Disabling the reconfiguration optimizations is a useful comparison. However, DAM still ignores switching costs, so the results do not establish whether the proposed allocator has an advantage over other strategies that already account for those costs.

These remaining evaluation gaps still limit my confidence in how broadly the conclusions apply. It also keeps me from raising my score.

No doubt the topic is promising and valuable, I think the paper needs much more work to meet the standards of ICLR or similar venues. I encourage the authors to continue strengthening it.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.