acceptodds
Under review as a conference paper at ICLR 2027

Learning Agent-Compatible Context Management for Long-Horizon Tasks

Abstract

LLM agents increasingly face long-horizon tasks such as web search and deep research in real-world applications, where accumulated context can cause long-context degradation and reasoning failures. Prior work mitigates this through context management with agent-side context control or fixed strategies such as summarization, which require training the agent itself for adaptation — making it impractical for closed-source agents and ignoring that different agents may require different strategies. We introduce Adaptive Context Management (AdaCoM), which trains an external LLM to manage the context of a frozen agent through flexible modification actions and end-to-end reinforcement learning. Across diverse agents on web search and deep research benchmarks, AdaCoM substantially improves performance by preserving task constraints and progress while pruning stale content. The learned strategies reveal a Fidelity–Reliability Trade-off: agents with higher vanilla ReAct performance benefit from higher-fidelity context preservation, whereas lower-performing agents require more aggressive compression to stay within a reliable reasoning regime. Transfer experiments show that AdaCoM generalizes most effectively across agents with similar capability (measured by vanilla ReAct performance), suggesting a practical path toward reusable context managers for agent systems.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, context management, long-horizon tasks

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yfJ2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes AdaCoM, an external context manager for frozen long-horizon language agents. Instead of requiring the task agent to summarize its own trajectory or use a predefined memory operation, AdaCoM allows a separate LLM to delete, rewrite, merge, or preserve arbitrary context messages. The manager is initialized through supervised demonstrations and then optimized with end-to-end reinforcement learning using final task outcomes and rule-based process rewards, while the underlying agent remains unchanged. Experiments show consistent improvements across several agent backbones over vanilla ReAct, self-management methods, and a trained fixed-summarization manager. The paper further analyzes cross-agent transfer and reports that managers transfer better between agents with similar vanilla performance. Based on the learned context lengths, the authors propose a Fidelity–Reliability Trade-off: stronger agents tend to benefit from retaining more raw context, whereas weaker agents benefit from more aggressive compression.

Critical strengths
  1. Practical and flexible formulation. Separating context management from the frozen task agent makes the method applicable to closed-source agents, while the general edit space is more expressive than a fixed summarization operation.

  2. Strong and relatively comprehensive experiments. AdaCoM is evaluated across multiple agents and two tool-using task families, with useful comparisons against ReAct, self-management, fixed summarization, untrained and SFT-only managers, and additional memory baselines.

  3. Promising transfer and efficiency results. The manager improves most agent–manager combinations, and the reported wall-clock measurements suggest that shorter contexts and fewer redundant actions can offset manager inference latency.

Critical weaknesses
  1. The novelty positioning is incomplete. This paper should explain more precisely what is new beyond external learned compression and flexible context operations.

  2. The Fidelity–Reliability Trade-off is supported mainly by correlation. The central evidence is the ordering of vanilla accuracy and learned context length across four source agents. Vanilla benchmark accuracy may not capture context-handling capability, and the observed relationship may be confounded by model-specific reasoning styles, APIs, native context windows, or optimization differences.

  3. Some training and evaluation signals limit the generality of the conclusions. BrowseComp-Plus training uses privileged gold/key-document process rewards, while MCP-Bench-Wiki is synthetically generated and evaluated primarily with proprietary LLMs. Moreover, the reported API cost excludes the local manager’s computation, so the total resource advantage is not yet fully quantified.

Questions for authors
  1. How does AdaCoM differ technically and empirically from ACON and LongSeeker, and can these methods be discussed or compared more explicitly?

  2. Does AdaCoM retain its gains on BrowseComp-Plus when trained without gold-document and key-document process rewards?

  3. Is vanilla ReAct accuracy genuinely the best proxy for an agent’s ability to use long contexts? Have alternatives such as controlled long-context retrieval accuracy or sensitivity to distractors been considered?

  4. How much of the cross-agent transfer pattern is quantitatively explained by source–target performance similarity? Can the authors report a correlation or regression analysis rather than relying mainly on visual inspection?

  5. For the efficiency comparison, can the authors report total computation and cost including manager inference, rather than only task-agent API cost? Does the result remain favorable when the Qwen3-4B manager’s GPU cost is included?

5 comments
AuthorsOpenReview ↗

Response to Reviewer yfJ2 (Part 1/3)

We thank you for your constructive suggestions and for recognizing AdaCoM's practical, flexible formulation, comprehensive experiments, and promising transfer and efficiency results.

W1, Q1: Novelty positioning and comparison with ACON and LongSeeker

Our novelty lies in the question we raise and answer: Do different agents need different context management strategies, and if so, how do they differ? To answer it, AdaCoM learns from each frozen agent's task feedback when to edit, which messages to change, and what to preserve. Keeping the agent frozen makes the method applicable to agents that cannot be trained. The flexible edit space allows different strategies to emerge: across 8 agents from 7 model families, we find that suitable strategies differ across agents and vary with capability (Sections 4.3–4.4), and that managers transfer best between agents of similar capability (Section 5).

Technical differences

MethodContext managed byMain agentWhen to edit contextAgents evaluatedStudies per-agent differences
ACONExternal compressor (optimized prompt)FrozenSummarize at a task-dependent length limit3 GPT models (guideline optimized on GPT-4.1 only), plus 1 distilled Qwen3-14B✗
LongSeekerAgent itselfTrained (SFT)Decided by the agent1, Qwen3-30B-A3B✗
AdaCoMExternal manager (RL)FrozenDecided by the manager8 (4 with per-agent managers, 4 held-out), 7 families✓

Only AdaCoM combines flexible context operations with a trained external manager, without fixing when to edit or imposing an editing template. ACON and LongSeeker do not study strategy differences across agents, and each covers only one or two model families.


Empirical studies

MethodDeepSeek-V3GLM-4.5-Air
ReAct17.7832.56
ACON16.00 (−1.78)30.89 (−1.67)
AdaCoM26.19 (+8.41)35.33 (+2.77)

Table R1: Mean@3 (%) on BrowseComp-Plus, 150 questions × 3 runs.

(1) ACON. We use Qwen3-4B, the backbone of AdaCoM's manager, as the compressor. As in ACON, its guideline is optimized once with GPT-5 (on 200 training questions with DeepSeek-V3) and reused for GLM-4.5-Air, and the compression threshold is 8K tokens (4K and 16K were no better). ACON falls below ReAct on both agents. On GLM-4.5-Air, it compresses 4.8 times per trajectory on average, and in 30 incorrect rollouts the gold answer appeared in a tool result but was summarized away before the agent answered. Thus, with the same 4B model, prompt optimization alone is not enough to adapt it to each agent, whereas AdaCoM improves both agents after RL.

SettingDeepSeek-V3GLM-4.5-Air
LongSeeker-format ReAct (no context operations)20.034.0
LongSeeker32.032.0

Table R2: Mean@1 (%) on BrowseComp-Plus, 50 questions × 1 run.

(2) LongSeeker. LongSeeker trains the agent itself (only its inference code is public), so it cannot be directly compared with AdaCoM. We therefore test its context management paradigm: keeping LongSeeker's interaction format, we compare frozen agents with and without its context operations.

  • DeepSeek-V3 improves from 20% to 32%, but at the cost of 6.3 restarts (rollbacks to step 1) per question on average. Output tokens rise 2.5-fold, and 32 questions exhaust the 35-step budget, only 3 of which are answered correctly.
  • GLM-4.5-Air cannot adapt to these tools: 52.4% of its steps are voided after three consecutive format errors; it performs only 26 non-skip operations in 50 questions, and its accuracy does not improve.

Thus, when flexible operations are handed to a frozen agent, their effect depends on the agent. AdaCoM instead delegates editing to an external manager, which lets us discover each agent's strategy.


We will revise the paper as follows:

  1. Expand the comparison with prior methods in terms of editing flexibility, training setup, and analysis of agent-specific strategies.
  2. Add a discussion of ACON and LongSeeker with empirical comparisons.
AuthorsOpenReview ↗

Response to Reviewer yfJ2 (Part 2/3)

W2, Q3, Q4: Why we use vanilla ReAct accuracy to measure capability in the Fidelity–Reliability trade-off, and evidence beyond correlation

The FR trade-off concerns how context management adapts to agents' ability to exploit raw contexts. To clarify, “exploiting raw contexts” goes beyond retrieval: the agent must use its accumulated search history to track task constraints, weigh incomplete or conflicting evidence, and decide what to search next. Stronger agents can do so reliably with longer raw histories, while weaker agents benefit from shorter, distilled contexts. AdaCoM reduces corresponding failure modes (Sec. 4.3), such as settling on answers that meet only some requirements or giving up while viable search directions remain. We therefore use vanilla ReAct accuracy as a task-matched measure of agentic capability for deep search, capturing overall task performance rather than retrieval alone.

Alternative measures(Q3)

We agree that external benchmarks offer a useful independent assessment of this capability and consider two types.

(1) Controlled long-context retrieval and distractor sensitivity: NoLiMa (ICML 2025) is a non-agentic needle-in-a-haystack benchmark. A sentence naming a person is inserted into a novel; a subsequent question shares no keywords with it, requiring retrieval through semantic association rather than keyword matching.

AgentBCP ReAct(%,↑)Effectivelength(↑)Drop with distractors(points,↓)
GLM-4.5-Air32.562K23.1
GPT-OSS-20B28.891K11.6
Kimi-K218.564K7.6
DeepSeek-V317.784K29.5
Seed-1.6-Flash16.441K0.0

Table R3: Following NoLiMa, effective length is the longest context retaining at least 85% of the mean accuracy at 250–1K tokens. Distractor drop is the accuracy loss in percentage points after adding distractors at 4K.

Neither metric consistently tracks BrowseComp-Plus (BCP) performance. Although GLM outperforms Kimi on BCP, it has a shorter effective length (2K vs. 4K) and suffers a larger accuracy drop from distractors (23.1 vs. 7.6 points).

Thus, the learned context strategies do not simply follow retrieval ability. Long-context benchmarks also rank these agents differently: AA-LCR places Kimi (53.0) above GLM (46.7) and DeepSeek (40.7), unlike NoLiMa.

(2) Long-horizon agentic benchmarks (multi-step search, coding, and tool use):

AgentBCP ReActBrowseComp (from GLM-4.5 technical report)Terminal-Bench Hard (from Artificial Analysis)MCP-Universe
GLM32.5621.320.519.5
Qwen3-Max27.78–20.518.2
Kimi18.567.915.919.1
DeepSeek17.781.56.814.3

Table R4: Performance on long-horizon agentic benchmarks.

These rankings broadly agree with BCP ReAct, except that Kimi-K2 slightly outperforms Qwen3-Max on MCP-Universe. This supports using long-horizon agentic benchmarks to assess agent capability on BCP.


Potential confounders of the FR trade-off (W2)

  • Reasoning style: Style can explain exceptions, e.g., Gemini imitating the Qwen-trained manager's report-style notes and stopping its search (Appendix C.6). Yet agents from the same family can prefer different managers, in line with task performance: GPT-OSS-20B does best with the GLM-trained manager, GPT-4o-mini with the Kimi-trained manager.
  • API: All agents use the same prompt, tools, retriever, step limit, and decoding settings.
  • Native context window: GLM and DeepSeek both have 128K windows, yet their learned lengths span the four source agents' extremes: 7.0K and 1.9K.
  • Optimization: All managers start from the same SFT checkpoint and share RL hyperparameters and rewards. Even without manager training, per-step self-summarization (SumAgent) improves Kimi's accuracy but lowers GLM's.

Together, these observations suggest that the FR trade-off is not primarily driven by these confounders.


Quantitative analysis(Q4)

Across all 28 cross-agent pairs, we regress transfer gain (target mean@3 with the transferred manager minus target vanilla ReAct accuracy) on capability gap (absolute source–target difference in ReAct accuracy). Targets differ in how much they can improve, so we use target fixed effects to compare managers on the same target [R1].

Results: For a given target, a 1-percentage-point increase in the source–target capability gap is associated with a 0.30 pp lower transfer gain. The gap explains 43% of variation across managers for the same target (p = 0.002). Choosing the closest source by ReAct accuracy yields an average regret of 0.92 points, versus 2.61 for random selection; regret is the gap to the best transferred manager for each target.


We will revise the paper as follows:

  1. Clarify that "exploiting raw contexts" refers to agentic capability for long-horizon search, including reasoning and decisions across steps.
  2. Add the external benchmark discussion and quantitative transfer analysis to the appendix.

[R1] Taskonomy: Disentangling Task Transfer Learning

AuthorsOpenReview ↗

Response to Reviewer yfJ2 (Part 3/3)

W3, Q2, Q5: Training signals, MCP-Bench-Wiki evaluation quality, and total inference cost

Dependence on privileged process rewards (Q2)

AdaCoM retains gains without privileged document-level rewards. In Appendix Table 7, we remove all process rewards (α=0\alpha=0), including gold- and key-document rewards, and train the manager using only the final-answer reward. Mean@3 decreases from 26.19% to 24.38% on DeepSeek-V3 and from 36.22% to 33.33% on Qwen3-Max, still exceeding ReAct by 6.60 and 5.55 percentage points, respectively.


MCP-Bench-Wiki task quality and evidence of substantive gains (W3)

We agree that synthetic tasks and LLM-based evaluation require explicit quality checks. Building on MCP-Bench's construction framework, we apply additional task validation and inspect paired trajectories to check whether higher scores reflect better task completion and grounding.

Task and evaluation quality. Our pipeline uses five task generators and screens each candidate with three models other than its generator, requiring at least two solvability scores ≥7/10. Surviving tasks undergo end-to-end dry runs against the live Wikipedia MCP server and full-log quality filtering. Held-out tasks are generated separately with training-topic blacklists, deduplicated, and screened for execution quality, sufficient tool use, and substantive answers (Appendix B). These checks target dependence on a single generator, unsolvable tasks, execution failures, and task overlap.

MCP-Bench compares its LLM judge's assessments with those of three independent human annotators, reporting mean agreement of 1.43 on a 0–2 scale after prompt shuffling and score averaging (Appendix A.9, Table 12). We therefore adopt this evaluation framework, using Claude Opus 4.6 as the judge.

Checking the gains. To examine whether the score gains correspond to substantive improvements in research, we also inspect paired DeepSeek trajectories. The analysis indicates that AdaCoM reduces premature completion and improves cross-source verification by preserving evidence and unfinished research requirements. These observations are consistent with the aggregate gains on DeepSeek in task fulfillment (42.1 → 49.1) and information grounding (48.0 → 64.4) in Appendix C.1, Table 8, supporting substantive improvements in research quality. We will add representative case studies to the appendix.


Total inference cost including the manager (Q5)

In Table 4, we evaluated task-agent API costs using 50 BrowseComp-Plus questions with three runs per question. We now include an estimate of the manager's inference cost from the same runs. As our evaluation workload underutilizes the A100 GPUs hosting the Qwen3-4B manager, we use Aliyun[1] and Nscale[2] API tariffs as reference estimates of inference cost, avoiding charges for idle GPU time.

The manager averages 140,936 input / 5,350 output tokens per task with Qwen3-Max, and 262,477 input / 11,258 output tokens with GLM-4.5-Air. We use Alibaba Cloud's Qwen3-4B (Instruct) LoRA rates[1] for China (¥0.30/1.20 per million input/output tokens, converted at ¥7.00 per US dollar) and Nscale's historical Qwen3-4B-Instruct-2507 rates[2] ($0.01/0.03). The resulting costs are:

Task agentManager pricingReAct ($)AdaCoM agent ($)AdaCoM manager ($, est.)Net savings
Qwen3-MaxAliyun[1]0.17000.09900.007037.7%
Qwen3-MaxNscale[2]0.17000.09900.001640.8%
GLM-4.5-AirAliyun[1]0.04000.02400.01327.1%
GLM-4.5-AirNscale[2]0.04000.02400.003032.6%

Table R5: Estimated total inference cost per task.

Costs are in USD per task, with savings calculated before rounding. Including the manager, estimated net savings remain 37.7–40.8% for Qwen3-Max and 7.1–32.6% for GLM-4.5-Air under these two reference API tariffs. We will update Table 4 and Section 4.5 with manager cost.

[1] Alibaba Cloud Model Studio. Token-based LoRA deployment pricing.

[2] Nscale. Historical Qwen3-4B-Instruct-2507 pricing, archived October 2025.

Reviewer_7QSEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AdaCoM, a framework that trains an external LLM to manage the context of a frozen agent without modifying the working agent itself. Unlike previous fixed summarization approaches, AdaCoM supports flexible context editing and uses reinforcement learning with outcome and process rewards to learn agent-specific management strategies. Experiments on BrowseComp-Plus and MCP-Bench-Wiki show consistent improvements across several LLM agents, with additional evidence of cross-agent transfer. The authors also identify a proposed Fidelity–Reliability Trade-off, suggesting that stronger agents benefit from preserving more context, while weaker agents prefer more aggressive compression.

Critical strengths
  1. Separating context management from the underlying agent is practically useful, particularly for closed-source models that cannot be fine-tuned. The flexible editing interface is also a reasonable extension beyond fixed summarization.

  2. The experiments cover multiple agent backbones and two task settings. The improvements over trained summarization baselines, together with cross-agent transfer experiments, provide meaningful evidence that learned context management can benefit long-horizon agents.

Critical weaknesses

The core idea of using a separate component to maintain and optimize context for long-horizon agents has already been explored in COMPASS (ACL 2026) and ACON (ICML 2026), both of which also investigate adapting context management to improve agent performance and distilling context-management capabilities into smaller models. Similarly, RL-based optimization of context management is not new, as demonstrated by work like SUPO (ACL 2026) and Context-Folding (ICML 2026). AdaCoM's main distinction appears to be training an external manager with a more flexible message-editing action space while freezing the underlying agent. While this is a useful design choice, the paper does not sufficiently establish why it constitutes a substantial methodological advance over these closely related approaches.

Questions for authors

1: Can the authors compare AdaCoM against a simpler adaptive context manager? The trajectory analysis suggests that these relatively straightforward mechanisms account for much of the improvement. It remains unclear whether the additional RL training and unrestricted editing actions are necessary.

2: What is the total computational and monetary cost of training and deploying AdaCoM, including external manager inference? Table 4 reports agent API costs but does not fully account for manager computation, making the claimed cost savings difficult to assess.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 7QSE (Part 1/2)

We thank you for the suggestions on related-work comparisons and for recognizing the practical value of external context management and the breadth of our experiments.

W1, Q1: Key advance and baseline comparisons

Prior work has explored external context managers and RL-based optimization, but has not addressed our central question: Do different agents need different context management strategies, and if so, how do they differ? Keeping the agent frozen makes AdaCoM applicable to agents that cannot be trained. Its flexible edit space lets the manager learn agent-specific strategies: across four agents from four model families, we find that preferred context lengths vary with agent capability (Secs. 4.3–4.4). Fixing the operation, as in SumCoM, prevents us from uncovering these differences and yields smaller gains overall.

Technical comparisons

MethodContext managed byMain agentWhen to edit contextNo fixed editing templateAgents evaluatedStudies per-agent differences
COMPASSExternal compressorFrozenSummarize when the Meta-Thinker intervenes✗2, Gemini family✗
ACONExternal compressor (optimized guideline)FrozenSummarize at a task-dependent length limit✓3 GPT models (guideline optimized on GPT-4.1 only), plus 1 distilled Qwen3-14B✗
SUPOAgent itselfTrained (RL)Summarize at a task-dependent length limit✗Qwen2.5-32B, Seed-OSS-36B✗
Context-FoldingAgent itselfTrained (RL)Fold completed sub-tasks✗Seed-OSS-36B✗
AdaCoMExternal manager (RL)FrozenDecided by the manager✓8 (4 with agent-specific managers, 4 held-out), 7 families✓

Only AdaCoM combines a trained external manager with a flexible message-editing action space, without fixing the editing template or timing. Other methods cover one or two model families and do not study strategy differences across agents.


Empirical comparisons

MethodDeepSeek-V3GLM-4.5-Air
ReAct17.7832.56
ACON16.00 (−1.78)30.89 (−1.67)
SUPO (frozen)16.89 (−0.89)35.11 (+2.55)
AdaCoM26.19 (+8.41)35.33 (+2.77)

Table R1: Mean@3 (%) on BrowseComp-Plus, 150 questions × 3 runs.

  • SUPO and Context-Folding. Both train the agent, so we test their protocols on frozen agents without agent-side RL. Context-Folding trails AdaCoM on DeepSeek and Qwen3-Max despite using the much larger agent for context management (Appendix, Table 10). For SUPO, we use 64K for GLM and 16K for DeepSeek, as DeepSeek's trajectories never reach 64K. SUPO scores close to AdaCoM on GLM but below the reported ReAct baseline on DeepSeek.
  • ACON. We use Qwen3-4B, AdaCoM's manager backbone, as the compressor. Following ACON, we optimize its guideline with GPT-5 on 200 DeepSeek training questions and reuse it for GLM. Compression starts at 8K tokens (4K and 16K were no better). With compression guideline optimization, ACON yields no gains over the reported ReAct baselines, whereas AdaCoM improves both agents through agent-specific RL.
  • COMPASS. Its Context Manager replaces history with a structured brief after Meta-Thinker interventions. Removing the Meta-Thinker lowers BrowseComp accuracy to 15.2%, versus 16.8% for ReAct, with Gemini 2.5 Pro as the Context Manager (COMPASS, Table 3). This suggests compression alone does not consistently improve performance, as also seen with SumAgent in our paper. The paper does not specify the Context Manager's update timing without the Meta-Thinker. For comparison, SumAgent and ACON test per-step and threshold-triggered summarization, respectively; both score below AdaCoM in the reported results.

On simpler adaptive context managers (Q1)

Our comparisons cover the following context management mechanisms:

  • Summarize at each step: SumAgent, SumCoM.
  • Summarize at a length limit: ACON, SUPO, A-Mem.
  • Truncate at a length limit: ReAct@32K.
  • Let an agent decide when to edit: MemAct, COMPASS, Context-Folding.

These simpler mechanisms can help one agent without yielding comparable gains for others. This motivates using a flexible edit space to discover agent-specific strategies.

Regarding the necessity of RL: ACON optimizes compression guidelines for a 4B compressor without RL and achieves accuracy below ReAct. In contrast, AdaCoM's agent-specific RL adds 3.52–6.44 points over the shared SFT initialization across all four agents (Table 3), supporting the need for RL to adapt the manager to each agent.

AuthorsOpenReview ↗

Response to Reviewer 7QSE (Part 2/2)

Q2: Training and deployment cost

One-time training budget. Across DeepSeek, GLM, Qwen and Kimi, each Qwen3-4B manager's RL training takes approximately 17.14–43.13 hours on 16 GPUs up to the selected checkpoint. Including the shared SFT initialization (3.51 hours on eight GPUs; 28.06 GPU-hours), the estimated budget is 302–718 GPU-hours per manager on A100 hardware. These log-based estimates use elapsed time including rollout, training, validation, and waiting.

For context, COMPASS reports approximately 480 A100 GPU-hours for its Context-12B manager's SFT and DPO. RL training for ReSum's 7B/30B agents requires an estimated 720–1,250 GPU-hours, calculated from its reported GPU counts, per-step times, and training configuration. Our trained managers can be reused across tasks and transferred across agents, amortizing the training cost over continued use.

Inference. In Table 4, we evaluated task-agent API costs using 50 BrowseComp-Plus questions with three runs per question. We now include an estimate of the manager's inference cost from the same runs. As our evaluation workload underutilizes the A100 GPUs hosting the Qwen3-4B manager, we use Aliyun[1] and Nscale[2] API tariffs as reference estimates of inference cost, avoiding charges for idle GPU time.

The manager averages 140,936 input / 5,350 output tokens per task with Qwen3-Max, and 262,477 input / 11,258 output tokens with GLM-4.5-Air. We use Alibaba Cloud's Qwen3-4B (Instruct) LoRA rates[1] for China (¥0.30/1.20 per million input/output tokens, converted at ¥7.00 per US dollar) and Nscale's historical Qwen3-4B-Instruct-2507 rates[2] ($0.01/0.03). The resulting costs are:

Task agentManager pricingReAct ($)AdaCoM agent ($)AdaCoM manager ($, est.)Net savings
Qwen3-MaxAliyun[1]0.17000.09900.007037.7%
Qwen3-MaxNscale[2]0.17000.09900.001640.8%
GLM-4.5-AirAliyun[1]0.04000.02400.01327.1%
GLM-4.5-AirNscale[2]0.04000.02400.003032.6%

Table R2: Estimated total inference cost per task.

Costs are in USD per task, with savings calculated before rounding. Including the manager, estimated net savings remain 37.7–40.8% for Qwen3-Max and 7.1–32.6% for GLM-4.5-Air under these two reference API tariffs.


We will revise the paper as follows:

  1. Highlight differences from prior methods in editing flexibility, training setup, and analysis of agent-specific strategies.
  2. Add the related-work comparisons and new baseline results.
  3. Add SFT/RL GPU-hours to the appendix and manager inference costs to Table 4 and Section 4.5.

[1] Alibaba Cloud Model Studio. Token-based LoRA deployment pricing.

[2] Nscale. Historical Qwen3-4B-Instruct-2507 pricing, archived October 2025.

AuthorsOpenReview ↗

Update on the revised paper

Thank you again for your suggestions. We have uploaded the revised paper, with changes marked in blue. We clarified AdaCoM’s differences from prior agent-side and external context-management methods in the Introduction and Related Work, and added design and empirical comparisons in Appendix C.4. We also included manager inference costs in Table 4 and detailed training and inference costs in Appendix A.3.

We would welcome any further feedback on these revisions and would be happy to clarify any remaining questions.

Reviewer_Tv98OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AdaCoM trains an external context manager to improve frozen agents without updating their parameters. A Qwen3-4B-Instruct manager learns through SFT and reinforcement learning to rewrite, merge, or delete history entries by message ID. Evaluations on BrowseComp-Plus and MCP-Bench-Wiki examine both agent-specific adaptation and manager reuse across agents. On BrowseComp-Plus, average mean@3 across four agents rises from 24.17% to 32.77%. The transfer study reports that 23 of 28 cross-model settings outperform ReAct.

Critical strengths
  1. The main comparisons support target-specific training as a useful way to improve frozen agents. In Table 1, average mean@3 across four models increases from 24.17% to 32.77%, with all four exceeding their own ReAct baseline. Table 3 reports further gains of 3.52–6.44 percentage points after a common SFT initialization. The SumCoM comparison supports the value of flexible editing beyond fixed summarization. MemAct still scores higher on Qwen, so the contribution is consistent adaptation across agents rather than universal superiority.

  2. Section 5 provides substantial evidence that trained managers can be reused across agents. On BrowseComp-Plus, 23 of 28 cross-agent pairs improve over ReAct, including evaluations on four targets excluded from manager training. The discussion also identifies model-specific exceptions, making the proposed performance-based matching strategy a useful empirical heuristic with explicit limits.

Critical weaknesses
  1. The aggregation of the main-table accuracy results needs clarification. Section 4.1 specifies 150 test questions, binary correctness scores, and mean@3 over three independent runs. If all runs evaluate the same questions with equal weight, the mean is determined by 450 binary outcomes, which does not explain the reported 26.19% in Table 1. Please clarify the effective sample counts and aggregation rule. This affects verification of exact effect sizes, but does not by itself establish an experimental error or inappropriate exclusion of failures.
Questions for authors
  1. Please provide the aggregation formula and effective counts underlying the 26.19% entry in Table 1. In particular, did all three runs evaluate the same 150 questions, and how were failed or incomplete rollouts counted? An existing per-run summary would suffice. If this is a typographical issue, identifying the corrected cell would resolve the question without a new experiment.
6 comments
AuthorsOpenReview ↗

Response to Reviewer Tv98

We thank you for carefully checking our evaluation protocol and for recognizing the value of agent-specific adaptation and cross-agent manager reuse.

W1, Q1: Numerical consistency and failure handling

We excluded trials interrupted by provider-side data inspection (e.g., filters for political or violent content) or platform restrictions before an answer was produced. Since retrieved content varies across trajectories, the same question can trigger a data-inspection error in one run but not another. We applied the same exclusion rule to all methods to avoid counting provider-side failures as model errors; the reported 26.19% was therefore based on fewer than 450 valid outcomes.

To further improve comparability, we are rerunning the excluded trials via alternative providers, retaining previously completed answers. Each fully re-evaluated setting now contains 450 valid outcomes: three runs on the same 150 test questions, with no missing trials. The AdaCoM/DeepSeek score is corrected from 26.19% to 26.89%.

Table 1 entryReported → updated mean@3 (%)Δ (pp)
ReAct/Kimi18.56 → 20.00+1.44
ReAct/GLM32.56 → 32.67+0.11
MemAct/GLM32.00 → 32.44+0.44
MemAct/DeepSeek16.67 → 16.89+0.22
AdaCoM w/o train/Kimi28.02 → 27.33−0.69
AdaCoM w/o train/DeepSeek17.57 → 17.33−0.24
AdaCoM/GLM35.33 → 35.56+0.23
AdaCoM/DeepSeek26.19 → 26.89+0.70

Most Table 1 changes are below 1 pp; AdaCoM retains the highest average performance. We are also rerunning excluded trials in the cross-agent transfer experiments under the same protocol. Of the 21 completed re-evaluations, 19 change by less than 1 pp, and the updated results continue to support the benefit of manager transfer. The remaining re-evaluations are ongoing; once complete, we will update all affected results throughout the revised paper and document the evaluation protocol explicitly.

AuthorsOpenReview ↗

Update on the revised paper

Thank you again for your feedback. We have uploaded the revised paper, with changes marked in blue. We updated the main benchmark results and clarified the evaluation protocol and provider-failure handling in Table 1 and Appendix A.2.

We would appreciate your thoughts on whether these revisions address your concerns, and would be happy to clarify any remaining points.

AuthorsOpenReview ↗

For Reviewer Tv98: Follow-up request on Submission #27989

Dear Reviewer Tv98,

We understand how busy this stage of the review process can be. Thank you for taking the time to review our paper and for highlighting our context manager’s ability to learn strategies tailored to individual agents and generalize across agents.

We wanted to follow up on your question about the mean@3 calculation in Table 1. We reran the provider-interrupted trials and updated AdaCoM’s DeepSeek-V3 score from 26.19% to 26.89%, based on 450 valid outcomes from three runs on the same 150 test questions. We also updated the other scores affected by provider errors. Appendix A.2 of the revised manuscript explains the aggregation rule and how interrupted trials and model failures are handled.

Reviewer yfJ2 has responded to our rebuttal and maintained their positive assessment, with no further concerns. With less than 10 hours left in the discussion period, we would appreciate any feedback you have time to share on whether the revised evaluation addresses your question and whether any part of the protocol needs further clarification.

Best regards,
The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and updated paper

Dear Reviewers and Area Chair,

Thank you for your time, thoughtful feedback, and discussion. We have uploaded the revised paper, with changes marked in blue. During the rebuttal, we have:

  • Clarified our central research question and differences from prior context-management methods, and added baseline comparisons.
  • Clarified the evaluation protocol and failure handling, updated the affected main results, and identified the transfer results that retain the original runs.
  • Included manager inference costs and detailed training and inference cost estimates.
  • Clarified the Fidelity–Reliability trade-off and added complementary capability evaluations, quantitative transfer analysis, and paired MCP case studies.

We have done our best to address all the concerns raised and incorporate the clarifications and evidence into the revised paper. We hope these revisions make the contributions and experimental findings clearer.

We would appreciate hearing your thoughts and welcome further discussion. If any concerns remain, we would be happy to clarify them. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.