acceptodds
Desk rejected at ICLR 2027

Adapting Context Compression for Long-Horizon Agents with Counterfactual Continuations

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-horizon agents require context compression to manage growing interaction histories. Compression quality, however, is ultimately determined by downstream execution. Existing prompt-adaptation methods infer compression errors by comparing full-context and compressed trajectories. Such comparisons cannot isolate individual compressions and are confounded by agent stochasticity. We first find that compression degrades reliability before solvability. Using matched counterfactual continuations that compare execution from the same agent state with versus without compression, we further show that severe degradation concentrates at isolated compression events. Motivated by this finding, we propose PAIR (**P**rompt **A**daptation using **I**nterventional **R**ollouts) for adapting structured compression prompts. PAIR identifies individual compressions that degrade subsequent execution, diagnoses their effects, and revises the relevant sections of a fixed compression template. PAIR achieves the strongest cross-run reliability among compressed methods in every main benchmark-scope combination, consistently exceeding the competing prompt-adaptation baseline. Without modifying the downstream agent, PAIR brings compressed execution close to the no-compression baseline and sometimes numerically exceeds it.

Primary area:
foundation or frontier models, including LLMs
Keywords:
long-horizon agents, context compression, prompt adaptation

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_wzkqOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focus on the problem of repeatd context compression in long horizon agents. The authors diagnose that compression mainly affect cross run reliability, which they then use to propose their solution. PAIR tries to identify which compressions that reduces the integrity of the rollouts by comparing repeated rollouts and revise the compression. The paper evaluate GPT 5.6 Luna and Minimax M3 across 3 long horizon agentic tasks, and achieves a very strong performance in accuracy while maintaining a competitive Token-Acc Pareto frontier.

Critical strengths
  1. The paper is timely, with clear writing and motivations from both experiment and theory
    • Memory management is crucial for long horizon agents and when/how to compact and manage memory is a contested topic in recent literature.
    • Their writing is clear, straightforward, and easily understandable along with clear visualization.
    • Their preliminary experiments on pass@k\text{pass@k} vs passk\text{pass}^k along with their short theory on the equivalence of boundary performance gap and overall task reward is clear, well motivated, and while trivial can serve as a valuable insight on future work in this area.
  2. The methods itself are clean, and shown to perform really well on benchmarks
    • The methods are straightforward and easily understandable. Their way of estimating the harmful boundaries while simple, are effective.
    • The results are very good. It performs the best accuracy-wise on AppWorld and against ACON (which is quite a strong baseline) by a big margin.
    • The deployment (for in-distribution tasks) is straightforward, with competitive overall tokens.
    • The trained prompts is transferable across model with minimal drop in performance, which is surprising to me!
Critical weaknesses
  1. There are inconsistencies between main table and downstream analysis table
    • For example in Table 8 and Table 1/2, the delta from AppWorld prefix does not match.
  2. The method is expensive to run and favors easily repeatable tasks
    • The formulation of the objective necessitates that the method knows the reward for each resumed compression boundaries, which is obtained from multiple rollouts. This favors tasks that are repeatable (have some sort of statically saved environment state) with unclear continuation on online adaptation of unalterable environment.
    • The cost of training is quite significant. Even accounting for the repeatability of the system prompt, the cost for adapting to one task family is not negligible.
    • There are no analysis on how the trained prompts interact with OOD data (both OOD on similar task, e.g. AppWorld -> τ2\tau^2-bench, or to general tasks, e.g. long math reasoning or agentic coding that necessitates compaction), making it questionable for deployment as each task family need to be classified correctly to the same prompt
Questions for authors
  1. Can the authors discuss the thoughts behind the design choice to use "fixed template" for revisions? I know that the results itself is good and the full comparison with ACON is available in the appendix, but expanding the very short explanation on this (how to design it, what can be changed in the compression template, etc) and including structural ablation in the main text would be valuable to first time readers.
  2. In the section 3.3 analysis, it seems like the harmful compression is concentrated in the earlier parts of the rollouts (which makes sense!) Does the authors provide analysis on this (e.g. how it affected the trained and final rollouts, how impactful it is, etc)? Furthermore, the section itself is dense and hard to understand (too much data dump thus feels a bit too AI-written) and it would be nice to see the section to be tidied up.
  3. Harmful compression exist in both successful and unsuccessful rollouts, Is there an analysis on how the model recover from it? Furthermore I think there might be some value in analyzing the boundary level effect of trained and compare the pre- and post-trained distribution of the harmful compression boundaries. (Figure 2)
  4. Do you think the method favors some task better than others, as Figure 5 might suggest that LLM lingua and no compression is also competitive in token/acc frontier? It would be nice if there exist an analysis/figure on the Pareto frontier on cost or latency or tokens.

Comment: This is a well made paper, with interesting and valuable insights. Limitations exists, but it is intertwined with the core method itself and are there due to the paper's early position in the literature. I think the amelioration for it is out of scope of the rebuttal and instead can serve as the basis of future works. My main concern therefore lies in the shown result consistencies instead and I'd be willing to increase my score if it has been resolved/clarified.

4 comments

We sincerely appreciate your efforts to review our paper and provide valuable suggestions. Below we address each concern in detail.


  • w1: Inconsistencies between the main tables and Appendix Table 8.
  • R w1: We have rechecked and recomputed all paired statistics using the final evaluation records, ensuring full consistency with Tables 1 and 2. The discrepancies arose from minor rounding differences and outdated results from an earlier tool-calling configuration that included additional instructions. Therefore we removed these instructions and reran the experiments to ensure a clean evaluation protocol. While the main tables were updated, some earlier statistics were inadvertently retained in Table 8. We have now corrected the confidence intervals, win/loss counts, and significance tests. The corrected results preserve our conclusions, including significant Pass$^3$ improvements on AppWorld under both compression scopes (p = 0.008 and 0.024). The revised analysis appears in Appendix D.9 (pp. 24–25, lines 1281–1322 in the revised version).

  • w2: High adaptation cost, reliance on repeatable environments, and limited OOD generalization.

  • R w2: We acknowledge these limitations and view PAIR as a starting point toward efficient and reliable localization of compression-induced errors in long-horizon agents. Replayable environments are required only during offline adaptation; deployment adds no rollouts or model calls beyond standard compression. Successive halving reduces counterfactual rollouts by approximately 42% and total adaptation tokens by 30.1% relative to uniform evaluation (Appendix D.7).

    Regarding OOD generalization, we follow ACON's setting of task-specific prompt adaptation. Because compression requirements depend on task-specific instructions, constraints, and execution states, we envision PAIR as a way to develop specialized compaction skills for important long-horizon task families, optimized offline once and reused within each family. We demonstrate zero-shot cross-model transfer (Appendix D.8); cross-task OOD generalization remains our future work.

    Beyond the method, we hope our analysis is useful in its own right. It characterizes how compression errors are distributed across long-horizon trajectories and shows that severe degradation concentrates at a small number of compression boundaries. It also shows that targeting these harmful boundaries can further improve overall performance, which makes them a practical handle for future compression methods. We are encouraged to see related efforts emerging in this direction, and we plan to build on these findings to develop cheaper ways of identifying harmful compression events. Upon acceptance, we will release our complete rollout records to facilitate future research.


  • q1: Why fixed template for prompt adaptation?
  • R q1: We adopt a fixed template for deployment compatibility and because structural flexibility shows limited additional benefit. First, keeping the section structure, required fields, and output format unchanged lets the adapted prompt replace the original compaction prompt directly, without modifying the downstream parsing or context-management pipeline; it also makes candidate prompts easier to validate and their summaries easier to analyze. PAIR therefore optimizes only the guidance within each section. Second, ACON's unrestricted template optimization gains largely by adding reasoning-oriented sections such as TODO and GUARDRAILS (Kang et al., ICML 2026, Appendix D.5–D.6), but production compaction templates, including our OpenClaw-based template, already contain such sections. As shown in Table 6 (Appendix D.5, p. 21), constraining ACON to the original template preserves its history-only Pass$^3$ (70.8%) and improves prefix-conditioned Pass$^3$ from 72.6% to 78.0%. Structural flexibility is thus not necessary for effective adaptation when starting from a well-designed template. Following the reviewer's suggestion, we now point to this ablation in Section 4 (lines 306–307) in the revision.

  • q2: Early harmful compressions, their impact, and Section 3.3 clarity.
  • R q2: Yes, and we agree with the reviewer's intuition: harmful compressions are concentrated early, with almost all of them occurring within the first three compressions. Under the initial prompt, this holds for 45 of the 47 boundaries that reduce continuation success. Our measure attributes each loss to the compression where it originates, so early compressions are the main source of compression-induced failures and the most valuable target for adaptation. Adaptation indeed helps most there: the rate of harmful boundaries in the first three compressions falls from 29.6% to 17.8% (new Appendix D.11). Following the reviewer's suggestion, we have also tidied Section 3.3, keeping the main findings in the text and moving secondary counts to Appendix D.2.
  • q3: How does the agent recover from harmful compressions, and how does the distribution of harmful boundaries change after adaptation?

  • R q3: The agent recovers by re-acquiring the information the summary omitted, and after adaptation harmful compressions become rarer and milder. In the cases we examined, the summaries contain no incorrect values but omit a required step. Continuations that re-derive this step from the task instruction or the environment succeed at the cost of a few extra steps, whereas those that follow the summary directly fail (5/5 vs. 0/4 at one boundary). We added case studies of being misled by, recovering from, and benefiting from a summary in Appendix D.12.

    For the distribution, we repeated the analysis of Section 3.3 with the adapted prompt. The share of boundaries that reduce continuation success drops from 23.9% to 17.1%, and the improvement comes from the first three compressions. Some harmful compressions remain, and interaction burden is not reduced. Since the adaptation evidence is collected under the initial prompt, a second adaptation round is a natural next step. The heatmaps appear in Appendix D.11 (Figure 11 in the revised version).


  • q4: Task-dependent performance and cost–accuracy trade-offs.

  • R q4: Thank you for this observation. We agree that the benefit of PAIR depends on the task, and we find it is largest on tasks that require more reasoning and planning. On AppWorld, PAIR improves Pass$^3$ over the unadapted prompt by 18.7 and 9.5 points on medium and hard tasks, compared with 7.0 points on easy ones. Token-level compression also struggles most here: LLMLingua-2 is competitive on the two smaller benchmarks but falls well below PAIR on AppWorld.

    For the cost–accuracy trade-off, Figure 5 plots Pass$^3$ against total tokens for all methods. On AppWorld, PAIR is slightly more reliable than no compression at about the same total tokens. On OfficeBench, it uses 15–19% fewer tokens. On τ2\tau^2-Bench Retail, no compression remains the better choice, since the compressor overhead outweighs the savings. We include no compression as a reference point, because long histories eventually exceed the context window in practice.

Thank you for the clear revision. My primary concern has been resolved. To reiterate, I think this is a good work that are clean and well motivated, and even though limitation exists (high adaptation cost, limited OOD generalization, etc), the author acknowledged it and I agree that it does not diminish the value of this work and thus can serve as a future research direction. As a final suggestion, I would like to see the actual pareto frontier of Figure 5 be drawn. I hope the authors can utilize the extra page allowance of the rebuttal to insert the extra analysis they did during the rebuttal (q3 and q4) as I think it is interesting, provide insight to their method, and are valuable to readers.

I acknowledged the author's response and have increased my score

Thank you for the encouraging feedback and for raising your score! Following your suggestion, we have drawn the Pareto frontier among compression methods in Figure 5 in the updated revision.

We also agree that the analyses added during the rebuttal (the post-adaptation boundary analysis to section 5, the case studies to section 3, and the task-dependent benefits to section 5) are valuable to readers and belong in the main text. To keep the page and line references in our responses stable for the other reviewers during the discussion period, they currently remain in Appendix D.11–D.12; we will integrate them into the main text using the additional page in the final version.

Reviewer_2PhXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PAIR addresses information loss when long-horizon agents compress their working context. It adapts compression prompts offline by comparing continuations from the pre-compression context and its summary under the same restored environment state, with subsequent compression disabled. The resulting evidence guides boundary selection, diagnosis, and prompt revision, followed by end-to-end candidate selection. The paper evaluates two compression scopes on AppWorld, OfficeBench, and tau2-Bench Retail, with template controls, transfer, and token-cost analyses. It reports improved success and repeat-run reliability; available PAIR records reproduce its own main success aggregates, while the comparative statistics require reconciliation.

Critical strengths

I find the localized intervention well motivated: it gives a more direct view of compression errors than comparing whole trajectories. The performance-difference argument is coherent under the stated finite-horizon setting.

The study includes useful template controls, transfer, cost accounting, and negative cases. Available PAIR records reproduce its six main accuracy and Pass3 results, although not the baseline comparisons.

Critical weaknesses

Some reported numbers do not match across the paper. For example, Table 1 shows a 4.6-point accuracy gain over ACON-UT on AppWorld-prefix, while Table 8 reports only 3.2 points. The Pass3 results also have some differences.

It is unclear how much of PAIR's improvement comes from counterfactual continuations. PAIR also uses different optimizer instructions and more offline compute than ACON, so a more controlled comparison would be helpful.

Questions for authors

Can the authors check the numbers in the main tables and appendices and explain why they differ?

Can the authors compare the methods using similar offline compute and optimizer instructions to better show the benefit of counterfactual continuations?

3 comments

We sincerely appreciate your efforts to review our paper and provide valuable suggestions. Below we address each concern in detail.


  • w1&q1: Inconsistencies between the main tables and Appendix Table 8.
  • R w1&q1: We have rechecked and recomputed all paired statistics using the final evaluation records, ensuring full consistency with Tables 1 and 2. The discrepancies arose from minor rounding differences and outdated results from an earlier tool-calling configuration that included additional instructions. Therefore we removed these instructions and reran the experiments to ensure a clean evaluation protocol. While the main tables were updated, some earlier statistics were inadvertently retained in Table 8. We have now corrected the confidence intervals, win/loss counts, and significance tests. The corrected results preserve our conclusions, including significant Pass$^3$ improvements on AppWorld under both compression scopes (p = 0.008 and 0.024). The revised analysis appears in Appendix D.9 (pp. 24–25, lines 1281–1322 in the revised version).

  • w2&q2: It is unclear how much of the improvement comes from counterfactual continuations, since PAIR also uses different optimizer instructions and more offline compute than ACON.

  • R w2&q2: We added a controlled comparison that fixes the optimizer instructions and removes only the counterfactual continuations: PAIR remains the best under both scopes, and the continuations account for 4.8–7.8 Pass$^3$ points under history-only and 2.4–3.0 points under prefix-conditioned compression. The two new variants use PAIR's optimizer instructions, template constraints, candidate count, and selection protocol. They select compression boundaries uniformly at random or with an LLM judge, run no counterfactual continuations, and give the optimizer the recorded trajectory after each selected compression. Without continuations, their offline compute is close to that of ACON.

    AppWorldHistory Acc.History Pass$^3$Prefix Acc.Prefix Pass$^3$
    ACON-UT81.270.881.372.6
    Random79.775.083.178.0
    LLM Judge78.672.083.977.4
    PAIR84.779.885.980.4

    With the same optimizer instructions and ACON-level compute, neither variant matches PAIR under either scope. The shared pipeline explains part of the gain over ACON-UT, since both variants exceed it in Pass$^3$; the rest comes from the counterfactual continuations. Under history-only compression, the variants also leave accuracy at the level of the unadapted prompt (79.7% and 78.6% vs. 78.6%), whereas PAIR reaches 84.7%. The additional offline compute of PAIR is spent on these continuations, which are the only component the variants remove. The full analysis appears in Appendix D.10 (Table 9 in the revised version).

Thanks the author for the comments. My main concerns have been addressed and I will maintain my score.

Thank you for taking the time to read our response and for confirming that your main concerns have been addressed! The corrected paired statistics (Appendix D.9) and the controlled comparison on counterfactual continuations (Appendix D.10) will both be kept in the final version. We are happy to answer any further questions during the discussion period.

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

The paper studies recurrent context compression for frozen long-horizon LLM agents. It measures the effect of a single compression by restoring the environment and continuing the agent from both the pre-compression and post-compression contexts. The differences between the two continuations give an outcome hazard (change in success) and an interaction burden. The paper shows that the summed hazard equals the gap to uncompressed execution. Building on this measure, PAIR finds harmful compressions in training runs and uses an optimizer LLM to revise the guidance inside a fixed OpenClaw compression template. The evaluation covers AppWorld, OfficeBench and τ²-Bench Retail, with GPT-5.6 Luna as the agent, the compressor and the optimizer.

Critical strengths
  1. The paper addresses a practically important problem with a clean, well-defined diagnostic. Counterfactual continuations from the same environment state isolate the effect of a single compression. The analysis also shows that harmful compressions occur inside successful trajectories, which trajectory-level comparisons cannot capture.
Critical weaknesses
  1. The paper never directly tests whether counterfactual boundary evidence causes the gains. No ablation keeps PAIR's optimizer and selection procedure while feeding it randomly chosen or LLM-judged compressions instead of counterfactually verified ones. The comparison with ACON cannot settle this. Besides the evidence, PAIR also differs in its optimizer instructions, uses far more evidence, and spends roughly 3–4× the adaptation budget.
  2. The method does not outperform no compression baseline. On AppWorld, PAIR scores 84.7 ± 1.7 (history-only) and 85.9 ± 2.4 (prefix-conditioned), against 84.1 ± 0.9 without compression. Both differences are within the run-to-run variance. On OfficeBench (81.1 and 81.4 vs 82.1 no compression) and τ²-Bench (80.8 and 77.5 vs 85.8 no compression; Pass³ 62.5 and 60.0 vs 72.5), PAIR stays below the uncompressed agent. The practical benefit of compression in this setting is therefore unclear.
  3. All adaptation and main experiments use one model, GPT-5.6 Luna, so generalization is untested.GPT-5.6 Luna serves as the agent, the compressor and the optimizer throughout. MiniMax-M3 appears only in zero-shot transfer on two benchmarks. No smaller open-weight models (e.g., Qwen) are evaluated. The agent and the compressor are always the same model. How does PAIR perform when a larger or smaller model is used as the agent and the compressor, or when a cheaper compressor serves a larger agent? How does the adaptation cost scale with model size?
  4. Latency and monetary cost are not reported. The paper reports token counts but neither wall-clock latency nor dollar cost. Recurrent compression puts extra compressor calls, of up to 8,192 output tokens each, on the agent's critical path.
Questions for authors

Please refer weakness section above

3 comments

We sincerely appreciate your efforts to review our paper and provide valuable suggestions. Below we address each concern in detail.


  • w1: No ablation tests whether counterfactual boundary evidence causes the gains; PAIR also differs from ACON in optimizer instructions, amount of evidence, and adaptation budget.

  • R w1: We ran the requested ablation: keeping PAIR's optimizer and selection procedure while replacing counterfactually verified boundaries with randomly chosen or LLM-judged ones reduces the gains under both compression scopes. Both variants use PAIR's optimizer instructions, template constraints, candidate count, and selection protocol. They run no counterfactual continuations; the optimizer instead reads the recorded trajectory after each selected compression.

    AppWorldHistory Acc.History Pass$^3$Prefix Acc.Prefix Pass$^3$
    ACON-UT81.270.881.372.6
    Random79.775.083.178.0
    LLM Judge78.672.083.977.4
    PAIR84.779.885.980.4

    PAIR is the strongest variant under both scopes, and counterfactual evidence accounts for 4.8–7.8 Pass$^3$ points under history-only and 2.4–3.0 points under prefix-conditioned compression. The shared optimizer and boundary analysis explain part of the gain over ACON-UT, since both variants exceed it in Pass$^3$, but not all of it: neither variant matches PAIR, and under history-only compression they leave accuracy at the level of the unadapted prompt (79.7% and 78.6% vs. 78.6%), whereas PAIR reaches 84.7%. The additional adaptation budget is spent on the counterfactual continuations, which are the only component these variants remove. The LLM judge does not improve Pass$^3$ over random selection, consistent with our finding that a harmful summary is hard to identify from the summary alone. The full analysis appears in Appendix D.10 (Table 9 in the revised version).


  • w2: Limited practical benefits compared with no compression.
  • R w2: As discussed in our Introduction, context compression is an essential engineering requirement for long-horizon agents. We include No Compression as a reference rather than a competing baseline. All evaluated compression methods reduce peak context length relative to this reference, addressing a practical deployment constraint. Performance improvements beyond the uncompressed reference are additional benefits rather than a prerequisite for demonstrating the value of compression.

  • w3: All experiments use GPT-5.6 Luna; no smaller open-weight models; the agent and compressor are always the same model; adaptation cost versus model size.

  • R w3: We added this setting: the PAIR prompt transfers without re-adaptation to smaller open-weight compressors, and helps them more than it helps GPT-5.6 Luna.

    Compressor (AppWorld, prefix)Acc. initial →\to PAIRPass$^3$ initial →\to PAIR
    GPT-5.6 Luna81.0 →\to 85.972.0 →\to 80.4
    Qwen3.5-27B65.1 →\to 81.346.4 →\to 70.8
    Qwen3.5-9B39.3 →\to 67.526.8 →\to 49.4

    With the PAIR prompt, the open-weight Qwen3.5-27B compressor is on par with GPT-5.6 Luna using the initial prompt. The adapted prompt also transfers when both the agent and the compressor are replaced by the open-weight MiniMax-M3 (Appendix D.8). We do not use smaller open-weight models as agents because ReAct agents built on them are not yet reliable enough on AppWorld to measure Pass$^3$ under compression: scenario-level completion on test-normal is 8.9% for LLaMA3-70B [1] and 18.6% for Qwen2.5-32B-Instruct [2]. Regarding cost, adaptation depends on the number and length of counterfactual continuations rather than on model size (Appendix D.7), and these results show it need not be repeated when the compressor or the agent changes. The new results appear in Appendix D.13 (Table 10 in the revised version).


  • w4: Missing wall-clock latency and monetary cost analysis.

  • R w4: PAIR adds little deployment cost or latency since PAIR changes only the prompt text and adds no model calls. Following ACON, we estimate API cost from logged token counts at list prices ($0.20 / $1.20 per million input / output tokens for GPT-5.6 Luna) and measure wall-clock time on the 90 AppWorld training tasks under identical concurrency.

    AppWorldEst. cost per task (history / prefix, $)Wall-clock for 90 tasks (history / prefix)
    Prompting0.020 / 0.02413.3 / 18.2 min
    ACON-UT0.029 / 0.02820.1 / 20.1 min
    PAIR0.023 / 0.02714.8 / 16.4 min

    The one-time offline adaptation cost is reported in Appendix D.7.


[1] Trivedi, Harsh, et al. "Appworld: A controllable world of apps and people for benchmarking interactive coding agents." ACL 2024.

[2] Chen, Kevin, et al. "Reinforcement learning for long-horizon interactive llm agents." arXiv 2025.

Dear Reviewer zxQZ,

As the discussion phase is approaching its end, we would appreciate knowing whether our response and the new experiments (Appendix D.10 and D.13) address your concerns. We are happy to provide any further clarification.

Best regards,

The Authors of Submission 18699

Dear Reviewer zxQZ,

Thank you again for the time you spent reviewing our paper and for your valuable comments. Following your review, we have:

  1. Run the ablation you suggested, replacing counterfactually verified boundaries with random or LLM-judged ones under PAIR's optimizer and selection procedure (Appendix D.10);
  2. Added experiments with smaller open-weight Qwen 3.5compressors serving a larger agent (Appendix D.13);
  3. Reported estimated API cost and wall-clock time;
  4. Clarified the role of no compression as the uncompressed reference in our setting.

As the discussion phase is closing soon, we would be grateful to hear whether these additions address your concerns. We are happy to clarify anything further.

Best regards,

The Authors of Submission 18699

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Revision

Dear ICLR 2027 AC, SAC, PC, and Reviewers,

We sincerely thank all reviewers for their valuable feedback and appreciate the constructive discussion during the rebuttal period. We are encouraged that Reviewer wzkq raised their rating to 4 and Reviewer 2PhX confirmed that their main concerns were addressed. We appreciate the reviewers' recognition of the strengths of our work, including a timely and practically important problem (Reviewers zxQZ, wzkq), a clean and well-motivated boundary-level diagnostic (Reviewers zxQZ, 2PhX, wzkq), clear presentation (Reviewers 2PhX, wzkq), and strong and carefully controlled empirical results (Reviewers 2PhX, wzkq).

One major concern was the inconsistency between the main tables and Appendix Table 8 (Reviewers 2PhX, wzkq). We recomputed all paired statistics from the final evaluation records and corrected the affected entries (Appendix D.9); the conclusions are unchanged, and both reviewers confirmed that this concern was resolved. Another concern was how much of the improvement comes from counterfactual evidence (Reviewers zxQZ, 2PhX). We added a controlled ablation that keeps PAIR's optimizer and selection procedure and replaces the counterfactually verified boundaries with random or LLM-judged ones (Appendix D.10); PAIR remains the strongest variant under both compression scopes. We also added experiments with smaller open-weight compressors serving a larger agent (Appendix D.13) and reported estimated deployment costs and measured wall-clock latency in our response (Reviewer zxQZ).

The new experiments and analyses are included in the revised manuscript (Appendix D.10–D.13 and Figure 5). To keep the page and line references in our responses stable during the discussion period, the post-adaptation boundary analysis and the case studies currently appear in the appendix; as discussed with Reviewer wzkq, we will move them into the main text in the camera-ready version, together with an expanded discussion of task-dependent benefits.

We respect the concern on the comparison with no compression (Reviewer zxQZ). We clarified that our setting, as stated in the Introduction and Problem 1, studies compression under a context budget, where no compression serves as the uncompressed reference rather than a competing baseline.

Finally, we are grateful that the reviewers highlighted the quality and potential value of our work. We believe our boundary-level protocol for localizing compression-induced errors and our evidence that targeting harmful boundaries improves agent reliability provide a useful foundation for future research on context management in long-horizon agents. We respectfully submit that these qualities make our work a valuable addition to the ICLR community and a strong candidate for acceptance.

Best regards,

The Authors of Submission 18699

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

Related papers

Discussion (0)

Sign in to comment.