acceptodds
Under review as a conference paper at ICLR 2027

DMC-5: Dual-Model Collaboration with Cache-aware Context Compression for Long-horizon Agents

Abstract

Long-horizon agents repeatedly feed an ever-growing interaction trajectory to a specialized agent model. Prefix caching reuses computation over repeated history, but cannot prevent context growth or remove information irrelevant to the current decision. This is especially costly when the specialized agent lacks the cached-input discounts available to general-purpose models, making trajectory compression a natural alternative. However, existing compression methods on selective context risk losing global information, recursively rewriting summary leads to semantic drift, while repeatedly recompressing the full history shifts the growing cost to the compressor. Therefore, we introduce **DMC-5** (Dual-Model Collaboration with CaChe-aware Context Compression), an asymmetric framework that assigns full-history integration to a general compression model and action generation to a specialized agent model. DMC-5 maintains the raw trajectory as an Append-only Canonical View, preserving global access while allowing the compressor to reuse a stable historical prefix instead of recursively rewriting or recomputing prior context. The agent operates only on a compact Decision View with recent raw interactions, reducing repeated input while retaining the information needed for the next action. DMC-5 further incorporates **YAMATO** (integritY-Aware Memory Abstractions for Tool Operations) to preserve execution-critical information during lossy compression. Experiments on AppWorld, OfficeBench, and 8-objective QA show that DMC-5 reduces agent model input by 38.5-57.0% and end-to-end cost by 14.5-28.0% with 78.0-85.0% compressor cache-hit rates, while improving task performance by 1.0-5.8% over the uncompressed baseline.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long-Horizon Agents, Cache-Aware Context Compression, Dual-Model Collaboration, Prefix Caching, Tool-Using Agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vqoxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a cache-aware cost-efficient compressor to manage the context of LLM agent. The cost is controlled by directly appending new messages to the history, utilizing prefix cache. An additional step YAMATO is applied to prevent the loss of important tool-execution information. Experiments demonstrate that this method keeps performance on representative agentic benchmarks while reducing token cost.

Critical strengths

This paper shows promising experimental results: DMC-5 achieves performance improvement on all the three agentic benchmarks while reducing total cost.

Critical weaknesses
  1. The paper claims that fine-tuned or customized LLM endpoints often do not enjoy cached-input discount (in line 141). As far as I know, this is not true for most model providers. The referred document (https://help.aliyun.com/en/model-studio/context-cache) for this claim also does not contain this policy. Hence, the scenario of this paper may not hold.
  2. The method for optimizing cost is simply ensuring cache hit by appending new messages to the history messages, which is already a common practice, lacking novelty or contribution to real application.
Questions for authors
  1. Please provide more materials that support the fundamental claim that fine-tuned or customized LLM endpoints often do not enjoy cached-input discount.
  2. The input information of V and YAMATO all comes from the same source -- the whole trajectory for compression. It seems meaningless to design 2 steps. What about merging the two steps and getting the output via a single LLM request?
  3. The method and benchmarks seems specifically designed for function-calling agent, where the tools are all well-formulated API. While in practice, most current agents work as "coding agents", where the tool-calling may be arbitrary command and the tool result can be any unstructured information or logs. What about the performance of DMC-5 on these more common and complicated benchmarks, like swe-bench or tau-bench?
8 comments

cognition of DMC-5's consistent task-performance improvements across three agentic benchmarks while reducing end-to-end inference cost. 

We will further address your concerns below through additional analyses on the deployment assumptions and generality of cache‑aware cost optimization, the complementary roles of the compressor and YAMATO in preserving execution‑critical information, and the applicability of DMC-5 to agents with unstructured tool interactions.

W1 / Q1. Deployment assumptions and cached-input pricing for fine-tuned endpoints

Thank you for raising this point. We conducted a systematic survey of 11 major providers offering fine‑tuned inference endpoints. Our survey shows that discounted prefix caching is unavailable for fine‑tuned endpoints at major providers. Among the 11 providers surveyed, 7 offer no cached‑input discount, while the remaining 4 support them only for a small subset of fine‑tuned models. None provides universal cache‑discount support across its fine‑tuned model catalog. Thus, the pricing scenario motivating DMC-5 applies to a broad range of practical fine-tuned deployments.

ProviderFine-tuned endpointCached-input discount
Alibaba Cloud Model StudioLoRA token-based / dedicated deploymentNo separate cached-input discount
Volcengine ArkOnline fine-tuned inferenceNo separate cached-input discount
Baidu QianfanSFT token-based deploymentNo separate cached-input discount
OpenAI APIHosted fine-tuned modelsSelected models only (GPT-4.1, GPT-4o)
Azure Microsoft FoundryFine-tuned Standard endpointsSelected models only (GPT-4.1, GPT-OSS, o4-mini)
Google Vertex AITuned Gemini endpointsSelected model only (partial Gemini version)
Fireworks AILoRA / customized modelsSelected models only (GLM-5.3)
Together AIDedicated fine-tuned endpointsNo separate cached-token discount
GroqCloudLoRA inferenceNo separate cached-token discount
Amazon BedrockFine-tuned Custom ModelsNo separate cached-input discount
SiliconFlowFine-tuned model endpointsNo separate cached-input discount

We additionally examined three deployment services (Hugging Face Inference Endpoints, Baseten, and Replicate ) that primarily charge by GPU‑hours or instance runtime. Since these services do not use token‑level billing, they are excluded from the 11‑provider comparison above.

All data in this survey were derived from official provider websites, pricing pages, and fine‑tuning, deployment, and caching documentation. The Alibaba Cloud documentation((https://help.aliyun.com/en/model-studio/context-cache) cited in the manuscript does specify this deployment condition. Specifically, under Model Inference → Dedicated Deployment → Deployment Modes → Token‑based deployment, the documentation states that token‑based deployment supports LoRA fine‑tuned models and is billed according to the number of input and output tokens. The corresponding billing rule specifies a single input‑token price, with no separate cached‑input price or cache‑hit discount for these fine‑tuned deployments. This policy is also directly verified by our actual deployment. The fine‑tuned Qwen3.5‑27B LoRA endpoint used in our experiments returned 0 cached tokens out of 130M input tokens in the API usage records. We will make the exact documentation location and this empirical verification explicit in the revised manuscript.

To better clarify the generalization ability of the proposed method, we further evaluate DMC‑5 under the alternative setting where the specialized agent itself supports discounted prefix caching. Since the original fine‑tuned endpoint does not provide this pricing option, we use an unfine‑tuned Qwen3.8‑27B base endpoint of the same parameter scale, which supports cached‑input pricing (0.3/M cached input tokens, 1.5/M uncached input tokens, and 6.0/M output tokens). The evaluated task set remains exactly the same.

SettingAgent Cache HitTotal CostTask Success
Baseline70.5%35.7271.5%
DMC-551.5%32.38 (-9.4%)73.9% (+2.4 pp)

The uncompressed baseline already benefits substantially from prefix caching, with 70.5% of its agent input served from cache. Nevertheless, DMC‑5 further reduces total cost by 9.4% while maintaining task performance. Our additional break‑even analysis shows that, at the actual uncached‑input price, the cached‑input price would need to fall below approximately 9.1% of the uncached price before the uncompressed baseline becomes cheaper.

Overall, it confirms that the original deployment scenario is common across the surveyed fine‑tuned endpoint offerings, while the additional cache‑enabled evaluation demonstrates that DMC‑5 also remains cost‑effective when the acting agent benefits from prefix caching. Together, these results show that the motivating setting is well represented in current serving configurations and DMC-5 is robust across different cache‑pricing configurations. We will refine the introduction to describe the deployment conditions more precisely.

W2. Contribution beyond append-only cache reuse

Thank you for raising this important question. We agree that appending new messages to history is a common practice and is not, by itself, our contribution. Rather, DMC‑5's contribution lies in jointly optimizing end‑to‑end inference cost, task performance, and execution‑state integrity though an asymmetric dual‑model allocation framework: an Append‑only Canonical View as the cache‑stable full‑history input for the compressor, while constructing a separate compact Decision View for the specialized agent, together with YAMATO's exact restoration.

(1) Optimizing actual end-to-end inference cost rather than token reduction alone.

Reducing agent input tokens does not necessarily reduce the actual inference bill. The total cost also depends on cached and uncached input prices, compressor overhead, output tokens, and additional agent interactions caused by information loss. DMC‑5 explicitly formulates these factors into a unified end-to-end cost objective (Eqs. 3‑6).

The following results from the original Table 1 and Fig. 3 demonstrate why this distinction matters:

BenchmarkMethodAgent Input ReductionCompressor Cache HitTotal CostTask Score
AppWorldNo Compression––13.7177.0
FIFO40.0%–8.9836.3
AgentDiet30.1%0%11.9364.8
ACON38.0%0%11.5767.3
DMC-546.6%80.3%10.3178.2
OfficeBenchNo Compression––5.2587.0
FIFO30.6%–3.9559.3
AgentDiet28.0%0%4.1383.8
ACON35.7%0%4.1983.3
DMC-538.5%78.0%4.4988.0
8-objective QANo Compression––35.4039.1 / 47.6
AgentDiet59.9%0%25.8835.5 / 43.6
ACON60.0%0%25.9434.2 / 41.5
DMC-557.0%85.0%25.5044.9 / 54.0

These results illustrate the distinction between token reduction and actual cost optimization. For example, on 8‑objective QA, DMC‑5 achieves a smaller agent‑input reduction than ACON (57.0% vs. 60.0%), yet incurs lower end‑to‑end cost (25.50 vs. 25.94) and achieves substantially higher EM (44.9 vs. 34.2). This directly demonstrates why agent‑side token reduction alone is an insufficient proxy for practical inference efficiency. Across all three benchmarks, DMC‑5 reduces actual cost by 14.5‑28.0% relative to No Compression, while maintaining or improving task performance.

(2) An asymmetric dual-model architecture that preserves specialized agent capability while reducing inference cost.

DMC‑5 assigns full‑history integration to a general‑purpose compressor and task‑specific action generation to a fine‑tuned agent. Table 4 directly evaluates this model-role allocation:

ConfigurationCompressorActing AgentAppWorld SuccessTotal Cost
Specialized only–Fine-tuned Qwen3.5-27B77.0%13.71
General only–DeepSeek-V4-flash70.6%6.93
Specialized → SpecializedFine-tuned Qwen3.5-27BFine-tuned Qwen3.5-27B77.4%21.49
General → GeneralDeepSeek-V4-flashDeepSeek-V4-flash67.9%9.07
General → Specialized (DMC-5)DeepSeek-V4-flashFine-tuned Qwen3.5-27B78.2%10.31

The results show that the general model alone does not achieve the task‑specific performance of the fine‑tuned agent (70.6% vs. 77.0%). Using the specialized model for both compression and action generation also raises the cost substantially (21.49). In contrast, the asymmetric configuration achieves 78.2% success at a cost of 10.31, combining the complementary capabilities of the two models.

(3) YAMATO preserves execution-state integrity, preventing compression errors from increasing interaction cost.

Context compression can lose or alter information required by subsequent tool operations. YAMATO addresses this problem through provenance‑aware summaries, exact restoration of selected raw records, and persistent failure tracking. Table 5 isolates its contribution:

MetricACON w/o YAMATOACON w/ YAMATODMC-5 w/o YAMATODMC-5 w/ YAMATO
Task Accuracy67.3%72.0%69.1%78.2%
Total Cost11.5712.1113.3110.31
Compression Events with Errors4.2%1.0%3.0%0.04%
Corrupted Critical Items1.23%0.35%0.33%0.01%
Critical Item Recall63.3%87.1%77.7%98.3%
Runs with ≥5 Repeated Failures1.6%0.6%1.6%0.2%
Maximum Same-Action Repeats3912296

YAMATO increases DMC-5 task accuracy from 69.1% to 78.2% while simultaneously reducing total cost from 13.31 to 10.31. This demonstrates that information integrity directly affects end‑to‑end efficiency: preventing corrupted historical state also avoids repeated failures and unnecessary agent interactions.

(4) Append-only history enables cache-aware compression within the dual-view architecture**.**

We agree that appending messages is common in agent context management. However, preserving an append-only history as the input to a context compressor is fundamentally different from simply appending messages to an agent prompt. Conventional context compression typically modifies historical content to reduce token usage, through selective deletion, recursive summary rewriting, or repeated full‑history recompression. These operations can sacrifice prefix‑cache reuse or incur repeated uncached processing costs, while potentially losing information needed for subsequent decisions.

DMC‑5 instead optimizes actual end‑to‑end billed cost. Its Append‑only Canonical View preserves the complete, immutable trajectory for cache‑aware full‑history integration by the general compressor, while a separate compact Decision View provides the specialized agent with only the context needed for action generation. This design combines compressor‑side cache reuse with agent‑side context reduction, while YAMATO protects execution-critical information during compression.

The controlled comparison in original Table 2 isolates the effect of this design:

BenchmarkStrategyCached Compressor TokensUncached Compressor TokensCompressor CostEnd-to-End CostTask Score
AppWorldRecursive Summary (ACON)017,2692.1711.5767.3
Append-only (DMC-5)51,83712,7172.5010.3178.2
OfficeBenchRecursive Summary (ACON)05,0520.834.1983.3
Append-only (DMC-5)15,6354,4100.894.4988.0
8-objective QARecursive Summary (ACON)060,44110.8425.9434.2
Append-only (DMC-5)198,15034,8449.0825.5044.9

As further illustrated in Fig. 4, the append‑only prefix cache‑hit rate increases from 0 to approximately 96% by the final compression round, whereas recursive summary provides no reusable prefix. Across the three benchmarks, DMC‑5 reuses 15.6K‑198.2K cached compressor input tokens per task, reduces uncached compressor input by 12.7‑42.3%, and improves task scores by 4.7‑10.9 percentage points over recursive summary. Despite processing more total compressor input, its compressor cost remains comparable to recursive compression; the resulting end‑to‑end cost is lower on AppWorld and 8‑objective QA and slightly higher on OfficeBench, where the task-score improvement is 4.7 points.

Therefore, the contribution is not the append operation itself, but its use as part of an end‑to‑end compression architecture: the Canonical View preserves full‑history access and cache reuse, the Decision View reduces repeated agent input, and YAMATO protects execution‑critical information during compression. This combination turns a common history‑management mechanism into a practical cost-aware compression strategy.

Reviewer_PqaqOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DMC-5, a context compression framework in which, at each compression event, the compressor reads the full eligible raw history while reusing cached prefixes. The agent receives compressed history together with recent raw action-observation pairs. YAMATO allows the compressor to select pointers to raw records that are restored exactly rather than rewritten. The framework targets a specific pricing setting in which the compressor benefits from cached-input discounts while the agent does not, for example when using a fine-tuned endpoint.

Critical strengths
  • Providing the compressor with full-history access and preserving selected raw records through YAMATO is an interesting design. Stable prefixes can reduce repeated prefill computation, while exact restoration avoids requiring the compressor to regenerate execution-critical content.
  • The experiments cover three benchmarks and include extensive analyses of component ablations, end-to-end cost, token volume, and model-role allocation using Qwen3.5-27B, DeepSeek-V4-flash, and Qwen3.7-flash.
Critical weaknesses
  • The contribution appears primarily to be harness engineering, with limited demonstrated generality beyond the target deployment setting. The effectiveness of pointer selection depends on the compressor's ability to identify critical information. If it marks every turn as important, the agent receives no meaningful compression benefit. Conversely, omitting necessary records can undermine subsequent execution. Deterministic restoration ensures faithful copying of selected records, but does not ensure that the selection itself is appropriate. The default compressor, DeepSeek-V4-flash, has roughly 284B total parameters compared with the 27B language model used as the agent. Its substantially larger total parameter count raises concerns about the resources needed to host deploy models, which are not captured by API costs alone. Although the API cost savings are useful, the benefits may be less compelling when both models must be deployed locally. It is also unclear whether a smaller, fine-tuned compressor could provide a better trade-off than relying on a capable compressor with cached-input discounts.

  • The method section is unnecessarily difficult to follow. The cost formulation in Section 2.1 and its connection to the subsequent design could be presented more directly. The core idea is to provide the compressor with the full raw history to reuse cached prefixes and allow it to emit pointers to records that can be restored afterward. Several equations appear to restate this workflow, and their additional role in explaining the method is unclear to me.

Questions for authors
  • Appendix 6.3.5 reports pricing sensitivity on fixed trajectories. Can the authors identify the break-even cached-input price at which DMC-5 ceases to be preferable to the relevant alternatives? As cache discounts decrease, how does it compare with switching to a substantially cheaper compressor?
  • What happens if the compressor is replaced with a substantially smaller model, such as Qwen3.5-9B? Would fine-tuning a smaller compressor offer a better balance of task performance and deployment cost?
  • Even a capable compressor such as DeepSeek-V4-flash may fail to identify which information will be critical later. How does the compressor decide which raw records to preserve through pointers and which turns to summarize, and how robust is this selection to both omission of critical information and excessive retention?
6 comments

We sincerely thank you for the thoughtful and detailed review. 

We especially appreciate your recognition of the value of full‑history access with stable prefix reuse, YAMATO's exact restoration of selected raw records, and the extensive analyses across multiple benchmarks, model‑role configurations, token usage, and end‑to‑end cost. 

We will further address your concerns below through additional analyses on the generality of the dual‑view design, the robustness of pointer selection, and the cost‑performance trade‑off under different compressor and pricing configurations.

W1. Generality, pointer selection, and compressor choice

Thank you for this thoughtful concern. DMC‑5 addresses a model‑allocation problem rather than simply adding a harness around an agent. It uses a capable general model for full‑history integration and cache reuse, while preserving the task‑specific strength of a specialized model for action generation. The resulting asymmetric dual‑model design jointly optimizes model roles, trajectory representation, cache reuse, and information integrity to improve the end-to-end quality–cost trade-off.

1) Generality beyond the original deployment setting.
To validate the generality of the proposed method, we additionally evaluate DMC‑5 in a setting where the specialized agent itself receives discounted prefix caching. The original fine‑tuned endpoint used in our main experiments has no cached‑input discount, so we use a Qwen3.8‑27B endpoint as the agent model in DMC‑5, which supports cached input at 0.3/M tokens, 1.5/M for uncached input, while keeping the evaluated tasks identical.

SettingAgent Cache HitTotal CostTask Success
Baseline70.5%35.7271.5%
DMC-551.5%32.38 (-9.4%)73.9% (+2.4 pp)

The uncompressed baseline already benefits substantially from prefix caching, with 70.5% of its agent input served from cache, yet DMC‑5 still reduces total cost by 9.4% while maintaining task performance. This shows that DMC‑5 remains beneficial when the acting model itself supports cached-input discounts.

Agent uncached-input price (P) (/M)BaselineDMC-5Net saving
0.5013.2814.28-7.5%
0.73 (break-even)18.4518.450.0%
1.50 (actual)35.7232.38+9.4%
3.0069.3859.55+14.2%
Cached / uncached price ratio (d)BaselineDMC-5Net saving
0.0225.9228.10-8.4%
0.0527.5628.82-4.6%
0.091 (break-even)29.8029.800.0%
0.20 (actual)35.7232.38+9.4%
0.5052.0539.53+24.1%
1.0079.2651.42+35.1%

We further identify the break-even boundary from the observed token traces. With the actual cache‑price ratio of 0.2, DMC‑5 becomes cheaper once the agent uncached‑input price exceeds 0.73/M tokens, well below the actual 1.5/M price. Conversely, at the actual 1.5/M input price, the cached‑input price must fall below approximately 9.1% of the uncached price before the uncompressed baseline becomes cheaper. 

Therefore, these results demonstrate that DMC‑5 is not limited to no‑cache‑discount deployments; it continues to provide substantial savings even when the acting model itself benefits from prefix caching, demonstrating robustness across different serving configurations.

2)  Robustness of pointer selection(Q3). 

The pointer‑selection mechanism is empirically effective in balancing information preservation and compression. Specifically, the compressor V explicitly decides whether each piece of historical state can be safely summarized or should be preserved verbatim. As described in Appendix 6.1, given the full Canonical View and the current task state, compressor (V) follows the compression prompt to judge which historical information remains relevant to subsequent decisions. Information that can be semantically compressed is written into the structured summary (Z_r), with provenance tags [O@i], [D@i], or [UNSOURCED] to distinguish observed facts, agent‑side derivations, and claims without an explicit source. In contrast, information whose exact form may affect later execution is not rewritten: (V) emits **<<TURN i RESULT>>**, **<<TURN i ACTION>>**, or **<<TURN i WORK>>** pointers, and YAMATO deterministically restores the corresponding raw observation, action, or work‑product record from the Append‑only Canonical View. Therefore, pointer selection is a semantic decision made by the compressor under the task context, while YAMATO guarantees exact restoration for the records that should not be lossy-compressed.

We evaluate the robustness of this selection from both directions: whether critical information is omitted and whether excessive information is retained. 

For omission, our recall metric is grounded in actual downstream behavior: an item is considered action‑relevant only when a subsequent action reuses its exact historical value or re‑invokes the operation that originally produced it across a compression boundary. As shown in Table 15, 1,008 out of 1,025 such action‑relevant items remain available in the Decision View, achieving 98.3% recall. That means, only 1.7% of the historically relevant items identified by subsequent execution are missed, showing that DMC‑5 rarely omits information that later actions actually depend on.

For excessive retention, the results show that this high recall is not achieved by simply preserving the full trajectory. On AppWorld, DMC‑5 reduces agent input by 46.6% and total cost by 24.8%, while achieving 78.2% task accuracy. Fig. 7 further shows that both the compressor output and the final Decision View remain bounded as the raw trajectory grows, with the pinned‑record component occupying only a small portion of the context. Thus, the selection mechanism simultaneously maintains high coverage of information actually needed later and substantial compression of historical context.

The six predefined categories define the automated evaluation protocol, not YAMATO's preservation scope. Our recall metric is grounded in actual downstream actions: historical values reused or operations re‑invoked across compression boundaries determine which information is action‑relevant, while the six categories enable deterministic extraction and verification of exact operational values. Ambiguous cases are excluded from corruption statistics when errors cannot be reliably attributed to compression‑induced information loss. Importantly, the experiments show that YAMATO rarely misses information that is actually needed later (1,008/1,025 action‑relevant items 98.3% recall), and it does not achieve this by retaining everything, as DMC‑5 still reduces agent input by 46.6% and total cost by 24.8%.

3) Why use a capable general compressor despite its larger parameter count? 

The two models in DMC‑5 serve different roles, and a larger general model does not replace the specialized agent. In Table 4, DeepSeek‑V4‑flash achieves only 70.6% when directly used as the acting agent, compared with 77.0% for the fine‑tuned 27B specialized agent. In contrast, using DeepSeek‑V4‑flash only for full‑history integration and the specialized 27B model for action generation achieves 78.2%, while reducing total cost from 13.71 to 10.31.

SettingTask SuccessTotal Cost
Specialized agent only77.0%13.71
General model only70.6%6.93
Specialized → Specialized77.4%21.49
General → Specialized (DMC-5)78.2%10.31

These results show that model size alone does not determine task performance. The general model contributes broad full‑history integration capability and benefits from low‑cost prefix reuse, while the fine‑tuned specialized agent preserves task‑specific action quality. DMC‑5 therefore uses the two models asymmetrically, assigning each model to the role where it is most effective.

Our current study targets API‑served deployment, where the dominant optimization objective is the billed inference cost determined by token usage and cache pricing. In this setting, the compressor's resident parameter count is not exposed as a separate user‑side billing term, so a capable general model can be economical when it substantially reduces repeated specialized‑agent inference.

Self‑hosted deployment involves a different optimization problem. In that setting, model memory footprint, hardware occupancy, throughput, and serving efficiency become first‑class costs in addition to token‑level computation. Therefore, the compressor that is optimal for API serving may not be optimal for local deployment, and the deployment strategy should explicitly trade off compressor capability against model footprint and system resources. Extending DMC‑5 to such self‑hosted settings through smaller compressors, model sharing, or deployment-aware compressor selection is an important direction for future work.

4) Would a smaller fine-tuned compressor provide a better trade-off?(Q2)

We additionally tested this direction by fine‑tuning a Qwen3.5‑9B model as the compressor. In this comparison, the acting agent is kept fixed as the same fine-tuned Qwen3.5-27B model; only the compressor is changed.

ConfigurationCompressorActing AgentAppWorld Success
No Compression–Fine-tuned Qwen3.5-27B77.0
DMC-5 (default)DeepSeek-V4-flashFine-tuned Qwen3.5-27B78.2
DMC-5 (cheaper general compressor)Qwen3.7-flashFine-tuned Qwen3.5-27B73.8
DMC-5 (small fine-tuned compressor)Fine-tuned Qwen3.5-9BFine-tuned Qwen3.5-27B70.2
General Model Only–DeepSeek-V4-flash70.6

As shown in the table, reducing compressor capability leads to a clear performance drop: 78.2% with DeepSeek‑V4‑flash → 73.8% with Qwen3.7‑flash → 70.2% with the fine‑tuned Qwen3.5‑9B compressor. This shows that compressor capability directly affects the quality of the Decision View delivered to the specialized agent.

The fine‑tuned 9B compressor does not provide a better quality‑cost trade‑off because its reduced compression capability substantially degrades downstream task performance. The success drops to 70.2% even slightly below directly using DeepSeek‑V4‑flash as the acting model (70.6%). This indicates that when the compressor fails to construct a sufficiently reliable Decision View, the specialized agent can no longer fully realize its task-specific advantage.

Overall, these results show that a smaller fine‑tuned compressor does not automatically yield a better end‑to‑end trade‑off. The value of a smaller compressor lies in further reducing serving cost only if it can still provide a sufficiently informative Decision View for the specialized agent. When compressor capability drops too far, as in the 9B setting, the downstream specialized agent performs no better than a general model acting alone, so the intended advantage of the asymmetric architecture is effectively lost. DMC‑5 therefore requires a compressor with sufficiently strong general understanding to preserve the information needed for downstream decisions.

Reviewer_ydQ2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DMC-5, a dual-model context compression framework for long-horizon agents. A general-purpose compressor operates on an append-only full-history view to enable prefix-cache reuse, while a specialized agent acts on a compact decision view. The proposed YAMATO mechanism preserves execution-critical information through deterministic restoration of raw trajectory records. Experiments on three benchmarks demonstrate substantial cost savings while maintaining or improving task performance.

Critical strengths
  1. DMC-5 considers both agent-side and compressor-side costs, including prefix-cache reuse and interaction overhead.
  2. YAMATO provides a useful mechanism for protecting execution-critical values from compression errors, with strong improvements in the ablation experiments.
  3. The paper includes comparisons against existing compression methods, model-role ablations, integrity measurements, pricing sensitivity, and latency analysis.
Critical weaknesses
  1. Limited significance of performance improvements. DMC-5 improves accuracy by only 1.2 points on AppWorld and 1.0 point on OfficeBench. Given the reported run-to-run variation, these gains may not be statistically significant.
  2. Dependence on specialized-agent pricing assumptions. The cost analysis assumes that the specialized agent receives no cached-input discount. The benefits of DMC-5 under deployments where the agent itself supports discounted prefix caching remain unclear.
  3. Limited validation of execution-critical information. YAMATO's integrity evaluation focuses on six predefined categories of values, and ambiguous matches are excluded from corruption statistics. Its reliability for semantically important information outside these categories remains unclear.
Questions for authors
  1. How would DMC-5 compare with the uncompressed agent if the specialized model also benefited from prefix caching?
  2. Given the substantial latency increase, could the authors evaluate whether the cost savings remain attractive under a joint cost–latency objective?
  3. How does YAMATO handle execution-critical information that does not match the predefined value patterns?
5 comments

We sincerely thank you for the thoughtful and constructive review. 

We especially appreciate your recognition of DMC-5's end-to-end cost formulation, YAMATO's protection of execution‑critical information, and the extensive analyses on model roles, pricing, and latency. 

We will further address your concerns below through additional statistical, cache‑sensitivity, cost‑latency, and integrity analyses.

W1. On the significance of the performance improvements

Thank you for raising this point. We would like to clarify that the primary goal of DMC-5 is to reduce the cost of long-horizon execution while preserving task performance. Benefiting from YAMATO's protection of execution‑critical information, DMC‑5 avoids the performance degradation commonly introduced by trajectory compression.

On AppWorld and OfficeBench, DMC-5 achieves 24.8% and 14.5% end‑to‑end cost reductions, while the task scores change by +1.2 and +1.0 points, respectively. We do not interpret these small positive differences as clear improvements in the underlying agent capability; rather, they show that substantial compression does not degrade task performance. This is particularly meaningful compared with the other evaluated compression methods, all of which reduce task quality to varying degrees:

MethodAppWorld ΔScoreAppWorld ΔCostOfficeBench ΔScoreOfficeBench ΔCost
FIFO-40.7-34.50%-27.7-24.8%
AgentDiet-12.2-12.98%-3.2-21.3%
ACON-9.7-15.61%-3.7-20.2%
DMC-5+1.2-24.80%+1.0-14.5%

The same trend is stronger on the longer 8-objective QA benchmark: DMC‑5 reduces end‑to‑end cost by 28.0%, while improving EM by 5.8 and F1 by 6.4. Therefore, our main observation is that DMC‑5 achieves substantial cost reduction without the performance degradation commonly introduced by trajectory compression.

Importantly, all main results are averaged over three independent runs. As reported in Appendix Table 12, the run‑to‑run standard deviations are also lower for DMC‑5 than for No Compression:

BenchmarkNo CompressionDMC-5
AppWorld77.0 ± 2.478.2 ± 0.9
OfficeBench87.0 ± 1.988.0 ± 1.0
8-objective QA (EM)39.1 ± 1.844.9 ± 0.3
8-objective QA (F1)47.6 ± 3.154.0 ± 1.0

Thus, across repeated evaluations, compression does not introduce an observable degradation or increased instability; instead, DMC‑5 consistently preserves the performance of the full-history agent while substantially reducing its inference cost.

To make this intended claim clearer, we will revise phrases such as "improves task performance" to "maintains or improves task performance" where appropriate.

W2 / Q1. DMC-5 when the specialized agent also benefits from prefix caching

Thank you for raising this important deployment question. We directly tested DMC‑5 in a setting where the agent itself benefits from discounted prefix caching, and DMC‑5 remains cost‑effective. Due to the fine‑tuned endpoint used in our main experiments does not receive a cached-input discount, we reran the same evaluation tasks using a Qwen3.8‑27B base endpoint that supports discounted caching. Its uncached input, cached input, and output prices are 1.5 / 0.3 / 6.0 per million tokens, respectively.

The uncompressed baseline already obtains a high 70.5% agent cache‑hit rate. Though DMC‑5 provide Decision View to specialized agent, which may corrput the prefix cache, DMC‑5's agent still receives 51.5% cache hit. Overall, DMC‑5 reduces the total cost from 35.72 to 32.38 (-9.4%) while maintaining task performance:

SettingAgent Cache HitAgent CostCompressor Cache HitCompressor CostTotal CostTask Success
Baseline70.5%35.72––35.7271.5%
DMC-551.5%29.4981.4%2.8932.38 (-9.4%)73.9% (+2.4 pp)

This result shows that DMC-5's cost advantage does not require the agent-side cache discount to be zero. Even when the uncompressed agent already reuses most of its historical prefix through caching, reducing the agent-facing context and moving full‑history integration to a highly cached compressor still provides additional savings.

We further analyze the break-even boundary using the observed token traces. We first fix the cached‑input price at 20% of the uncached‑input price and vary the agent's uncached-input price (P). The break‑even price is only 0.73/M tokens. Above this point, DMC‑5 is consistently cheaper.

Agent uncached-input price (P)BaselineDMC-5Net saving
0.5013.2814.28-7.5%
0.73 (break-even)18.4518.450.0%
1.0024.5023.33+4.8%
1.50 (actual)35.7232.38+9.4%
2.0046.9441.44+11.7%
3.0069.3859.55+14.2%
4.0091.8277.66+15.4%

We further examine how the strength of the agent‑side cache discount affects this conclusion under the actual uncached‑input price of 1.5/M tokens. Here, (d) denotes the cached‑input price as a fraction of the uncached‑input price. Since the uncompressed baseline has a higher agent cache‑hit rate (70.5% vs. 51.5%), increasingly aggressive cache discounts naturally favor the baseline. Nevertheless, DMC‑5 remains cheaper unless the cached‑input price falls below approximately 9.1% of the uncached‑input price.

Cached / uncached price ratio (d)BaselineDMC-5Net saving
0.0225.9228.10-8.4%
0.0527.5628.82-4.6%
0.0829.1929.53-1.2%
0.091 (break-even)29.8029.800.0%
0.1030.2830.01+0.9%
0.1533.0031.20+5.5%
0.20 (actual)35.7232.38+9.4%
0.3041.1634.77+15.5%
0.5052.0539.53+24.1%
1.0079.2651.42+35.1%

At the actual pricing ratio of (d=0.20), DMC-5 therefore operates well within the cost-saving region, reducing end-to-end cost by 9.4%. Only under an unusually strong cache discount, where cached input costs less than approximately 9.1% of uncached input, does the higher cache‑hit rate of the full‑history baseline outweigh the context reduction achieved by DMC-5.

Overall, these additional experiments clarify the deployment scope of DMC‑5. The method does not require the specialized agent to have zero cache discount: even when the uncompressed agent already achieves a 70.5% cache‑hit rate, DMC‑5 still reduces total cost while maintaining task performance. The benefit depends on the relative agent input price and cache discount, and our break‑even analyses provide an explicit boundary for this trade‑off. Under the evaluated real pricing configuration ((P=1.5), (d=0.20)), DMC-5 reduces total cost from 35.72 to 32.38 (-9.4%), while task success changes from 71.5% to 73.9%.

W3 / Q3. YAMATO's preservation scope and integrity evaluation

Thank you for raising this question. YAMATO is not restricted by predefined value categories when identifying information to preserve. Given the complete trajectory and task instruction, compressor V uses its general language understanding to assess the relevance of historical turns to subsequent decisions. It marks important turns with pointers, which YAMATO resolves by retrieving the corresponding raw action, observation, or work‑product records from the Append‑only Canonical View. This turn‑level mechanism supports arbitrary structured and unstructured content, including commands, logs, documentation, file contents, and artifact modifications.

The six predefined categories are used only for quantitative integrity evaluation, rather than defining the scope of YAMATO's preservation mechanism. They are selected based on the execution characteristics of AppWorld, a stateful benchmark involving nine applications and 457 APIs. Since its tasks rely heavily on API calls, we focus on representative exact operational values, such as identifiers, credentials, paths, and URLs, that directly affect tool invocation and application‑state transitions. These values can be deterministically extracted and compared against the canonical trajectory, enabling objective integrity measurements.

Action‑grounded recall for reliable semantic information evaluation. Our evaluation identifies actual downstream information requirements beyond predefined value categories, by tracing historical values reused or operations re‑invoked in subsequent actions. These action‑relevant items form the recall denominator (1,025), while the six predefined categories enable automated extraction and verification of preserved information. YAMATO retains 1,008/1,025 items (98.3% recall), with only 17 omissions. These omissions do not necessarily cause task failures, consistent with DMC‑5's 78.2% AppWorld accuracy (vs. 77.0% without compression) and 24.8% cost reduction. Thus, the six categories define the automated verification protocol, not YAMATO's preservation scope.

Ambiguous cases are excluded because compression‑induced corruption cannot be reliably established. When an incorrect value cannot be uniquely traced to the original trajectory, we cannot distinguish information loss from downstream agent errors. We therefore exclude such cases from corruption statistics to avoid unsupported attribution.

We will clarify the distinction between YAMATO's general record‑level preservation mechanism and the representative value categories used by the automatic integrity evaluator in the revised appendix.

Q2. Cost–latency trade-off

Thank you for this insightful suggestion. The practical value of a cost‑latency trade‑off depends strongly on the deployment scenario: latency‑critical interactive agents and long‑horizon autonomous agents naturally operate at different points on this trade‑off. Since assigning a monetary value to one second of latency is inherently deployment‑dependent, we directly report the relative cost‑latency trade-off observed in our experiments:

BenchmarkCost ReductionLatency (NoComp → DMC-5)Latency IncreaseCost–Latency Ratio*
AppWorld24.8%69 → 94 s+36.2%0.68
OfficeBench14.5%38 → 44 s+15.8%0.92
8-objective QA28.0%122 → 186 s+52.5%0.53

*The Cost-Latency Ratio denotes Cost Reduction / Latency Increase, indicating the percentage of cost saved per 1% increase in task-completion time.

For the long‑horizon setting targeted by DMC‑5, this trade‑off is especially relevant because the agent autonomously executes a sequence of tool interactions before returning the final task outcome, rather than requiring an immediate response after every intermediate step. The results show a consistent pattern: longer trajectories provide more opportunities for compression and larger accumulated cost savings, while invoking the compressor more frequently and therefore increasing serial latency. For example, OfficeBench achieves a 14.5% cost reduction with only a 15.8% latency increase, whereas the longer 8‑objective QA obtains the largest cost reduction (28.0%) with a correspondingly larger latency overhead.

Therefore, DMC‑5 is most suitable for long‑horizon deployments where reducing accumulated inference cost is valuable and moderate additional task‑completion latency is acceptable; latency‑sensitive applications can naturally choose a less aggressive compression configuration or no compression. We have also discussed this trade‑off in Appendix 6.4. Reducing the current synchronous overhead through asynchronous/parallel compression, speculative refresh, and overlap with agent/tool execution is an important direction for future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers and the Area Chair

We sincerely thank all reviewers for their constructive feedback and recognition of DMC‑5's strengths. We are encouraged that the reviewers highlighted several strengths of our work:

  • End‑to‑end cost‑aware design: Reviewer ydQ2 recognized our comprehensive cost formulation, including agent/compressor costs, prefix-cache reuse, and interaction overhead.

  • Effective architecture and information preservation: Reviewers ydQ2 and Pqaq appreciated the full‑history cache‑aware design and YAMATO's exact restoration of execution‑critical information.

  • Extensive experiments and promising results: All three reviewers acknowledged the experimental evidence or analyses. Reviewer vqox particularly highlighted the consistent task‑performance improvements across three benchmarks while reducing inference cost.

The central contribution of DMC‑5 is to move context compression from token‑count reduction to actual end‑to‑end inference‑cost optimization, through an asymmetric dual‑model architecture and integrity‑aware context compression. Across the three original benchmarks, DMC-5 reduces total inference cost by 14.5–28.0% while maintaining or improving task performance.

To address the reviewers' concerns, we have provided substantial additional evidence in our responses:

  1. Deployment validity and cache‑pricing generality (ydQ2, Pqaq, vqox). We surveyed 11 major model providers and verified the deployment assumptions through actual fine‑tuned endpoint usage records (0 cached tokens out of 130M input tokens). More importantly, a new experiment with a cache‑enabled agent demonstrates that DMC‑5 still reduces total cost by 9.4%, even when the uncompressed baseline already achieves a 70.5% cache‑hit rate. Additional break‑even analyses characterize the operating conditions under different cache prices and reuse rates.

  2. YAMATO's effectiveness and component necessity (ydQ2, Pqaq, vqox). We clarify that semantic selection and deterministic restoration are complementary operations requiring only one LLM call. Our action‑grounded integrity evaluation achieves 98.3% recall (1,008/1,025 items), while DMC‑5 reduces agent input by 46.6%, demonstrating effective preservation without excessive retention. Ablations further show that YAMATO improves AppWorld accuracy from 69.1% to 78.2% while reducing total cost.

  3. Model allocation and compressor capability (Pqaq, vqox). Model‑role ablations demonstrate the value of asymmetric general‑purpose compression and specialized action generation. Additional experiments with a fine‑tuned Qwen3.5‑9B compressor reduces AppWorld success from 78.2% to 70.2%, showing that simply using a smaller compressor is insufficient when its compressed Decision View no longer preserves the information required by the specialized agent.

  4. Generalization to unstructured tool‑use agents (vqox). New experiments on PinchBench and τ²‑bench demonstrate 8.2% and 8.5% total cost reductions, respectively, while preserving task success. These smaller gains are consistent with our original trajectory‑length analysis, as many tasks in these benchmarks trigger compression only once before termination, leaving less opportunity to amortize compressor overhead.

  5. Performance reliability and deployment trade‑offs (ydQ2, Pqaq). We clarify that the primary objective is cost reduction while preserving task performance, rather than claiming statistically significant accuracy improvement. Three‑run evaluations show no observed degradation and lower empirical variation. We additionally quantify the cost‑latency trade‑off and analyze DMC‑5 under different cache‑pricing and deployment conditions.

In the revised manuscript, we will strengthen the deployment‑policy references and clarify the pricing assumptions; streamline the method formulation around the dual‑view workflow and component roles; incorporate the new experiments and break‑even analyses; and expand the discussion of YAMATO's evaluation scope, compressor selection, and deployment trade-offs.

We hope these additional results address the reviewers' concerns and further establish the practical effectiveness, generality, and technical contribution of DMC‑5. We sincerely appreciate the reviewers' and Area Chair's consideration and hope these clarifications facilitate a more complete reassessment of the work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.