acceptodds
Under review as a conference paper at ICLR 2027

SearchTrimmer: Search-Selective KV Eviction and Loop Gating for Efficient Deep Research Agents

Abstract

Deep research agents incur costs from both an expanding history and repeated rounds of model inference and tool use. The growing history increases KV-cache memory, while compressing it can add rounds as agents attempt to recover missing observations. We propose SearchTrimmer, a training-free framework that addresses both costs through two components motivated by an analysis of agent trajectories. First, search observations occupy most of the history while containing little unique evidence and remaining inexpensive to recover. We therefore propose Search-Selective KV Eviction, which retains browse observations and reasoning while keeping only the most recent search observations in the KV cache. Second, additional search activity often repeats earlier queries. We distinguish two types of re-search: Restore, which leads to a newly retained browse observation, and Loop, which does not. Since Loops account for most of the increase, we propose Loop Gating, which limits their cumulative count and requests a final answer when the budget is reached, while leaving Restores unpenalized. Neither component requires retraining, an additional model, or access to attention scores. Across three benchmarks, SearchTrimmer reduces average KV-cache usage by – on WebExplorer-8B and – on WebSailor-32B, with maximum KV-cache reductions of – and –, respectively, while matching or exceeding the uncompressed agents' avg@3 accuracy and keeping round counts within approximately of their baselines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
KV cache compression, deep research agents, LLM agent efficiency

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_bhv5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SearchTrimmer, a training-free method that evicts KV cache of search results while retaining other parts. It introduces a loop gate that limits repeated searches. Across agents and benchmarks, the method reduces KV usage while maintaining average accuracy and approximately preserving iteration counts.

Critical strengths
  1. The paper is well motivated by insightful findings. Specifically, it identifies search results as the major context bottleneck and hence target of optimization. It also finds that search results can be quickly absorbed into reasoning hence keeping reasoning while evicting older search results is useful.
  2. The observation that compaction leads to repetition is intersting. The paper also proposed loop gating to effectively address that. I think this will motivate some of the future work for model training and harness optimization.
  3. The method is evaluated end-to-end across agents and benchmarks, so the evaluation result is credible.
Critical weaknesses

The loop gate does not seem to be the best way to address it as it requires hyperparamters and is a bit heuristic-based. Could authors propose some learned way to address the search repetition problem?

Questions for authors

See weakness.

1 comment

W1. Heuristic-based Loop Gating and Learned Alternatives

We thank the reviewer for the constructive suggestion. We agree that learning-based approaches could provide more adaptive control over search repetition. However, our goal is to develop a training-free framework that can be directly applied to existing deep research agents without retraining or introducing additional models. Loop Gating is therefore designed as a lightweight mechanism, motivated by our empirical observation that Loops account for most of the additional search calls induced by KV eviction (Figures 4 and 5). We also examine the accuracy–efficiency trade-off under different hyperparameter settings in Table 5 (Table 4 of the original manuscript).

Importantly, we believe the Restore/Loop distinction introduced in this work could provide useful supervision for future learning-based approaches. We envision two directions:

  • Loop-aware reward shaping: Penalizing Loops while leaving Restores unpenalized during reinforcement learning could encourage agents to avoid redundant searches directly.

  • Learned stopping policy: A lightweight classifier trained on agent trajectories could predict whether further exploration is likely to yield new evidence, enabling adaptive stopping decisions instead of relying on a fixed Loop budget.

These approaches could address complementary limitations of the current method. Reward shaping could reduce repetition at its source, while learned stopping could better distinguish productive exploration from unnecessary continuation. As shown in Appendix E.3 (Appendix D.2 of the original manuscript), our fixed-budget approach preserves most Restores but may still terminate trajectories before additional evidence is recovered. We consider learning-based extensions a promising direction for future work.

Reviewer_1PZ5OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes SearchTrimmer, a set of training-free heuristics to reduce the KV-cache usage of deep research agents without increasing the number of rounds. First, the authors observe that search observations occupy most of the history while containing little unique evidence, so they propose Search-Selective KV Eviction, which keeps only the k most recent search observations in the KV cache while retaining all browse observations and reasoning. Second, they find that eviction increases re-searches (queries similar to earlier ones), which fall into two types: Restores, which lead to browsing a new page, and Loops, which do not. Since Loops account for most of the increase, they propose Loop Gating, which caps the cumulative number of Loops and requests a final answer once the budget is reached. On WebWalkerQA, xbench-DeepSearch and GAIA with WebExplorer-8B and WebSailor-32B, SearchTrimmer reduces average KV-cache usage by 41–61% and 32–36% respectively, while matching or exceeding the uncompressed agents' accuracy and keeping round counts within about 6% of the baseline.

Critical strengths
  1. Competitive with KV-cache compression baselines. SearchTrimmer uses less average KV-cache memory than FastKV and ChunkKV while achieving higher accuracy on most benchmarks, and is on par with SWE-style observation masking in memory and accuracy while requiring fewer rounds (Table 2).

  2. Effective on the evaluated benchmarks. The simple heuristics work well on WebWalkerQA, xbench-DeepSearch and GAIA: average KV-cache usage drops by 41–61% (WebExplorer-8B) and 32–36% (WebSailor-32B) relative to the uncompressed agents, with matched accuracy and round counts within about 6% of the baseline.

Critical weaknesses
  1. Heuristic-driven method that may not generalize. The contribution is a pair of hand-crafted rules for deciding which parts of a deep research agent's context to keep and when to stop tool use: keep the k most recent search observations, and cap re-searches identified by lexical query overlap. These rules are tuned per agent on the evaluation benchmarks (e.g., the coverage rule selects k=2 for both agents, but WebSailor-32B is set to k=1 so that something is evicted), so it is unclear whether they carry over to other agents, tasks or context-management settings.

  2. Untested on long-horizon tasks. On the evaluated benchmarks, the baseline agents finish in about 8–12 rounds (WebExplorer-8B) and 5–6 rounds (WebSailor-32B), with roughly 2–8 searches per question (Table 3). Long-horizon deep research benchmarks such as BrowseComp [1] and FRAMES [2], which require many more searches across multiple sources, are not evaluated. This is where KV-cache growth is the most severe, and where a fixed k of 1–2 would evict far more context and Loop Gating would trigger more often, so both the benefits and the risks of the method remain unmeasured.

Questions for authors
  1. How sensitive are accuracy, rounds and KV usage to k? And how would k be chosen for a new agent or task without first collecting uncompressed trajectories on it?

  2. How accurate is the re-search detector? The similarity score is purely lexical, so it may miss paraphrased re-searches and flag different queries about the same entity. How is it applied to Chinese queries in xbench-DeepSearch? And why do the Table 4 results vary inconsistently across settings, e.g., WebSailor-32B's GAIA accuracy ranging from 40.8 to 47.9, and gating sometimes increasing rounds relative to no gate (xbench-DeepSearch: 5.21 → 5.27–5.52)?

  3. What is the effect of Loop Gating alone on the uncompressed baseline? The gate is evaluated only on top of eviction. Since the baseline agents also re-search, is the gate a general tool-use limiter or specifically a fix for eviction?

11 comments

We sincerely thank the reviewer for the valuable comments. Below, we provide detailed responses to each of the raised concerns and questions.

W1-(1). Motivation and Parameter Selection

We agree that SearchTrimmer uses simple rules rather than learned policies, and this simplicity is intentional: our goal is to develop a training-free framework applicable to existing agents without retraining or additional models. However, these rules are motivated by empirical properties of agent trajectories rather than optimized for benchmark accuracy. Below, we first explain the motivation and parameter selection for each component (W1-(1)), then present additional experiments evaluating generalization to a new agent, a long-horizon task, and another context-management framework (W1-(2)).

Search-Selective KV Eviction and kk. Our trajectory analysis shows that search observations dominate the history (88.488.4% and 71.371.3% of tokens) while containing little unique evidence (4.34.3% and 1.71.7% of answer facts). They are also less costly to recover than browse observations, while reasoning cannot be reproduced (Section 3.1.1). These observations motivate selectively evicting search observations rather than treating all history content equally.

The retention width kk is selected based on URL pointer coverage: we choose the smallest kk beyond which adding another search observation improves coverage by at most one percentage point, subject to leaving at least one search observation to evict in an average trajectory (Section 3.1.2). The coverage criterion selects k=2k=2 for both agents. However, WebSailor-32B averages only about two searches per trajectory (Table 3), so k=2k=2 provides little compression (99% average KV reduction; Table 1PZ5-5 in our response to Q1-(1)). We therefore use k=1k=1, following the second criterion.

Loop Gating and NN. Our analysis further shows that re-searches after eviction differ in whether they recover new retained browse evidence (Restore) or merely repeat earlier searches without adding new browse evidence (Loop). We therefore limit Loops while exempting Restores (Section 3.2).

The Loop budget NN is selected based on the distribution of Restores in ungated trajectories (Appendix E.3). At the chosen budgets, 89.489.4% of Restores on WebExplorer-8B (N=5N=5) and 91.991.9% on WebSailor-32B (N=4N=4) have already occurred. Subsequent re-searches are predominantly Loops (3.873.87 vs. 0.740.74 and 1.921.92 vs. 0.230.23 Loops vs. Restores per trajectory, respectively). WebExplorer-8B uses a larger budget because its Restores occur later, with only 83.883.8% occurring before N=4N=4.

Similarity threshold τ\tau. Unlike kk and NN, which control the amount of retained context and the stopping budget, respectively, τ\tau defines the similarity threshold for identifying repeated queries. We use the same τ=0.7\tau=0.7 across all agents without agent-specific tuning. The sensitivity analysis in Table 5 further examines how different threshold values affect accuracy and round counts.

Summary of parameter selection. While kk and NN are selected using agent-specific trajectory statistics, the same settings are applied across all three evaluation benchmarks for each agent, without benchmark-specific accuracy tuning. The similarity threshold τ=0.7\tau=0.7 is shared across all agents and benchmarks. Thus, SearchTrimmer does not require separate hyperparameter optimization for each benchmark.

W1-(2). Generalization to Other Agents, Tasks, and Context Management

Generalization to a New Agent and Long-Horizon Task. We applied SearchTrimmer to Tongyi-DeepResearch-30B-A3B [1] on BrowseComp-zh [2], a long-horizon benchmark with about 1717 searches per trajectory. The rule selects k=4k=4 whether pointer coverage is computed on BrowseComp-zh or on the pooled trajectories of WebWalkerQA, xbench-DeepSearch, and GAIA (Table 1PZ5-6 in our response to Q1-(2)). We used the gate settings of WebExplorer-8B (τ=0.7\tau=0.7, N=5N=5).

Table 1PZ5-1. SearchTrimmer on Tongyi-DeepResearch-30B-A3B (BrowseComp-zh, 289289 questions, single run). †^\dagger denotes SearchTrimmer without Loop Gating.

MethodAcc.RoundsKV avg (GB)KV max (GB)
Base39.7939.7920.4020.404.914.9110.4810.48
SearchTrimmer$^\dagger$ (k=4k=4)38.4138.4122.8022.802.232.235.495.49
SearchTrimmer (k=4k=4)38.4138.4115.1215.121.841.844.334.33

On 289289 questions, SearchTrimmer achieves 38.4138.41% accuracy compared with 39.7939.79% for Base (111111 vs. 115115 correct answers), while reducing average rounds by 2626%, average KV usage by 6363%, and maximum KV usage by 5959%. Without Loop Gating, KV eviction increases rounds by 1212% (20.40→22.8020.40 \rightarrow 22.80), whereas the full method reduces them to 15.1215.12, below the uncompressed baseline.

Compatibility with Other Context Management Setting. We further evaluated SearchTrimmer on top of ReSum [3], which summarizes the accumulated history when approaching the context limit. We retained the WebSailor-32B settings (k=1k=1, τ=0.7\tau=0.7, N=4N=4). Since each summary initiates a new context, we reset the cumulative Loop count at each summarization boundary. Qwen3-30B-A3B-Thinking serves as the ReSum summarizer, while Qwen2.5-72B-Instruct summarizes browsed pages, as in our main experiments.

Table 1PZ5-2. SearchTrimmer on top of ReSum (WebSailor-32B, single run). Gated rows are replayed on the ungated trajectories.

MethodAcc.RoundsKV avg (GB)KV max (GB)
GAIA
ReSum48.5448.545.745.742.622.626.736.73
+ SearchTrimmer$^\dagger$48.5448.547.357.351.571.573.663.66
+ SearchTrimmer50.4950.496.366.361.561.563.663.66
BrowseComp-zh
ReSum18.6918.6912.9312.935.475.476.866.86
+ SearchTrimmer$^\dagger$23.5323.5323.1123.112.312.315.655.65
+ SearchTrimmer21.1121.1117.0217.022.202.205.655.65

On GAIA and BrowseComp-zh, adding Search-Selective KV Eviction to ReSum reduces average KV usage by 4040% and 5858%, respectively, without decreasing accuracy. However, eviction increases rounds by 2828% and 7979%, primarily due to additional search calls (2.32→3.742.32 \rightarrow 3.74 and 7.17→15.787.17 \rightarrow 15.78 per trajectory). Loop Gating eliminates approximately 6060% of these additional rounds on both benchmarks while retaining the KV-cache reductions. The full method also maintains accuracy at or above ReSum (+1.95+1.95 and +2.42+2.42 percentage points, respectively), although its round counts remain higher than ReSum alone.

Together, these experiments provide evidence that SearchTrimmer's design principles can extend beyond the original agents and benchmarks and can complement an existing summarization-based context-management framework.

[1] Team, Tongyi DeepResearch, et al. "Tongyi deepresearch technical report." arXiv preprint arXiv:2510.24701 (2025).

[2] Zhou, Peilin, et al. "Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese." arXiv preprint arXiv:2504.19314 (2025).

[3] Wu, Xixi, et al. "Resum: Unlocking long-horizon search intelligence via context summarization." arXiv preprint arXiv:2509.13313 (2025).

W2. Evaluation on Long-Horizon Tasks

To measure both the benefits and the risks on long-horizon tasks, we evaluated WebExplorer-8B, WebSailor-32B, and Tongyi-DR-30B-A3B on BrowseComp-zh, which follows the design of BrowseComp and requires long, multi-source search. For example, WebExplorer-8B performs 16.116.1 searches per question, compared with 4.14.1–8.18.1 on our original benchmarks. All settings are unchanged from the main experiments, and we have added these results to Section 4.4 of the revised paper.

Benefits and risks of a small kk. To examine the impact of retention width on long-horizon tasks, we evaluate multiple kk values with and without Loop Gating.

Table 1PZ5-3. SearchTrimmer on BrowseComp-zh (289289 questions, single run). †\dagger: without Loop Gating.

AgentSettingAcc.RoundsKV avg (GB)KV max (GB)
WebExplorer-8BBase25.6125.6117.7417.749.069.0615.1815.18
SearchTrimmer$^\dagger$ (k=1k=1)24.5724.5719.2619.261.181.181.991.99
SearchTrimmer (k=1k=1)23.5323.5314.4214.421.031.031.681.68
SearchTrimmer$^\dagger$ (k=2k=2)24.9124.9118.9818.981.741.745.785.78
SearchTrimmer (k=2k=2)23.5323.5314.4914.491.601.605.785.78
WebSailor-32BBase15.2215.228.488.485.615.617.697.69
SearchTrimmer$^\dagger$ (k=1k=1)17.6517.658.408.402.102.103.983.98
SearchTrimmer (k=1k=1)17.9917.998.168.162.082.083.983.98
SearchTrimmer$^\dagger$ (k=2k=2)17.3017.308.428.423.263.265.995.99
SearchTrimmer (k=2k=2)17.6517.658.248.243.253.255.995.99
Tongyi-DR-30B-A3BBase39.7939.7920.4020.404.914.9110.4810.48
SearchTrimmer$^\dagger$ (k=1k=1)40.8340.8321.9921.991.591.594.984.98
SearchTrimmer (k=1k=1)38.0638.0615.1515.151.221.223.943.94
SearchTrimmer$^\dagger$ (k=2k=2)41.8741.8722.8722.871.821.825.645.64
SearchTrimmer (k=2k=2)40.1440.1415.2015.201.411.414.844.84
SearchTrimmer$^\dagger$ (k=4k=4)38.4138.4122.8022.802.232.235.495.49
SearchTrimmer (k=4k=4)38.4138.4115.1215.121.841.844.334.33

Across the tested kk values, KV eviction reduces average KV usage by 4242–8787% while maintaining accuracy within −1.38-1.38 to +2.43+2.43 percentage points of Base. However, it increases rounds by 77–1212% on WebExplorer-8B and Tongyi-DeepResearch-30B-A3B, while WebSailor-32B shows little change, partly due to its context-length limit. Loop Gating reduces rounds to 1818–2626% below Base on the former two agents, with at most a 2.12.1-point accuracy loss and up to 8989% average KV reduction.

These results demonstrate substantial KV savings across different retention widths, even on long-horizon tasks.

Trade-off of Loop Gating. As the reviewer anticipated, Loop Gating activates more frequently on long-horizon tasks. At N=5N=5, it truncates 5050% of WebExplorer-8B and 3838% of Tongyi-DeepResearch-30B-A3B trajectories on BrowseComp-zh. Since truncation can also prevent useful evidence collection, we vary NN to examine the resulting accuracy–efficiency trade-off.

Table 1PZ5-4. Loop Gating with different NN on BrowseComp-zh (τ=0.7\tau=0.7; gated rows replayed on ungated trajectories).

Agentno gateN=2N{=}233445566881010
WebExplorer-8B (k=2k=2)Gate firing–63.063.0%58.158.1%54.354.3%50.2\mathbf{50.2}%44.644.6%33.633.6%25.325.3%
Acc.24.9124.9121.4521.4521.8021.8023.5323.5323.53\mathbf{23.53}23.8823.8824.9124.9125.2625.26
Rounds18.9818.9810.4010.4012.1112.1113.3313.3314.49\mathbf{14.49}15.4315.4316.8516.8517.6417.64
KV$_{avg}$ (GB)1.741.741.471.471.521.521.561.561.60\mathbf{1.60}1.631.631.671.671.701.70
Tongyi-DR-30B-A3B (k=4k=4)Gate firing–54.054.0%46.746.7%40.840.8%38.1\mathbf{38.1}%35.635.6%31.831.8%28.728.7%
Acc.38.4138.4133.9133.9134.9534.9538.7538.7538.41\mathbf{38.41}37.7237.7237.0237.0238.0638.06
Rounds22.8022.8010.4810.4812.4612.4613.8813.8815.12\mathbf{15.12}16.0916.0917.6117.6118.7518.75
KV$_{avg}$ (GB)2.232.231.691.691.751.751.791.791.84\mathbf{1.84}1.881.881.951.952.012.01

Lowering NN consistently reduces rounds but may increase accuracy loss. On Tongyi-DeepResearch-30B-A3B, N≥4N \ge 4 maintains accuracy within 1.41.4 points of the ungated setting while reducing rounds by 1818–3939%, whereas N≤3N \le 3 reduces rounds by up to 5454% with a 3.53.5–4.54.5-point accuracy loss. On WebExplorer-8B, N=8N=8 preserves ungated accuracy with 1111% fewer rounds. Lower NN also reduces average KV usage by up to 1616–2424%. These results demonstrate the accuracy–efficiency trade-off controlled by NN on long-horizon tasks.

We also examined FRAMES, but WebExplorer-8B averages only 4.24.2 searches per question, comparable to our original benchmarks (4.14.1–8.18.1) and far fewer than BrowseComp-zh (16.116.1). We therefore prioritized BrowseComp-zh to evaluate the long-search regime raised by the reviewer.

Reviewer_oa8DOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

A training-free framework for ReAct-style deep research agents with two parts: (1) Search-Selective KV Eviction keeps all reasoning and browse observations and only the k most recent search observations in the KV cache, evicting older ones at the KV level while leaving the prompt text in place; (2) Loop Gating classifies re-searches as Restore (next call opens a new URL) or Loop, and requests a final answer once cumulative Loops reach N. On WebExplorer-8B and WebSailor-32B across three benchmarks, average KV usage drops 41–61% / 32–36% with accuracy preserved and rounds within ~6%.

Critical strengths
  • The design follows directly from a clear trajectory analysis: search observations dominate tokens but hold little unique evidence and are cheap to recover.
  • KV-level eviction rather than prompt-level deletion is well argued, and the method runs on both vLLM and SGLang without retraining or attention scores.
  • The paper takes the extra rounds induced by compression seriously; the Restore/Loop distinction is a nice idea.
  • Evaluation goes beyond KV counts. The fixed-memory throughput experiment (Fig. 6) shows where the savings actually pay off.
Critical weaknesses

W1. Accuracy claims should be softened. Gains over Base are small (+0.11 to +3.14; GAIA has 103 questions), and versus SWE-style masking avg@3 is roughly a 3-3 split, with pass@3 slightly below Base in two cells. "Preserves accuracy at much lower KV" is the supportable and still strong claim.

W2. Missing controls. (a) Search Masking (prompt-level removal of the same observations) should appear in Table 2 with accuracy and KV, since the KV-vs-prompt distinction is central. (b) Loop Gating on the uncompressed agent, to separate eviction from early stopping. (c) A gate without the Restore exemption, to show the distinction matters. (d) State the Base agent's max-round limit.

W3. Efficiency numbers are single runs and depend on external API latency; present small differences and the 1.19-1.42x figures as indicative, and say up front that the method is slightly slower than Base at low concurrency. (but this is a minor issue)

Questions for authors
  1. What is accuracy on the subset of trajectories the gate truncates, vs. the ungated agent?
  2. How is the final answer requested when the Loop budget is reached, and what happens if the agent still emits a tool call?
8 comments

We sincerely thank the reviewer for the valuable comments. Below, we provide detailed responses to each of the raised concerns and questions.

W1. Accuracy Claims

As the reviewer pointed out, SearchTrimmer aims to reduce KV-cache usage and compression-induced rounds while preserving, rather than improving, the accuracy of the uncompressed agent. The avg@3 differences over Base (+0.11+0.11 to +2.66+2.66 percentage points) are small relative to the benchmark sizes, and pass@3 is slightly lower than Base in two cells on WebExplorer-8B (77.0077.00 vs. 80.0080.00 on xbench-DeepSearch; 58.2558.25 vs. 59.2259.22 on GAIA).

Accordingly, we have revised the manuscript to state that SearchTrimmer preserves accuracy at much lower KV-cache usage:

  • Abstract: "while matching or exceeding the uncompressed agents' avg@3 accuracy" →\rightarrow "while preserving the accuracy of the uncompressed agents" (L28-29 of the revised manuscript)
  • Section 4.2: "avg@3 increases by …" →\rightarrow "avg@3 differs from Base by …", consistent with the existing statement that SearchTrimmer preserves avg@3 (L352-353)
  • Conclusion: "while matching or improving avg@3 relative to the uncompressed agents" →\rightarrow "while preserving the avg@3 of the uncompressed agents" (L536-537)

We also agree that SearchTrimmer and SWE-style masking achieve comparable avg@3. SearchTrimmer instead differs in efficiency, requiring fewer rounds (Table 2) and less wall-clock time (Table 3) on every agent–benchmark pair.

We thank the reviewer for helping us state our contribution more precisely.

W2-(a). Search Masking vs. KV eviction

We agree that Search Masking provides the most direct control for examining the distinction between KV-level eviction and prompt-level removal. Since the live web environment has changed since the original experiments, we re-ran Base, Search Masking, and SearchTrimmer† using the same experimental setup and evaluation period. Search Masking removes the same search observations as SearchTrimmer†, but from the prompt rather than the KV cache. Due to rebuttal time constraints, we performed one rollout per question, so small accuracy differences should be interpreted with caution.

Table oa8D-1. Base, Search Masking, and SearchTrimmer$^\dagger$ re-run under identical conditions (single run; mean over the three benchmarks of Table 2). KV in GB.

AgentMethodAcc.RoundsKV avgKV max
WebExplorer-8BBase52.3152.3110.5710.573.003.0015.1315.13
Search Masking51.8251.8212.0912.09 (+14.4(+14.4%))1.231.234.654.65
SearchTrimmer$^\dagger$52.2452.2411.0611.06 (+4.6(+4.6%))1.171.173.263.26
WebSailor-32BBase50.7850.786.286.282.892.897.157.15
Search Masking48.4648.469.179.17 (+46.0(+46.0%))1.761.765.495.49
SearchTrimmer$^\dagger$48.0348.036.516.51 (+3.6(+3.6%))1.531.534.954.95

†^\dagger denotes Search-Selective KV Eviction without Loop Gating.

Results. Search Masking and SearchTrimmer† achieve comparable accuracy (within 0.5 percentage points), but Search Masking incurs substantially more rounds (+14.4% and +46.0% over Base, compared with +4.6% and +3.6% for SearchTrimmer†). This supports our motivation in Section 3.1.2: unlike prompt-level removal, KV eviction preserves the representations of subsequent tokens that were computed with the removed observations in context. SearchTrimmer† also achieves lower average and maximum KV-cache usage on both agents.

We have added this comparison, including per-benchmark results, to Appendix C of the revised manuscript.

W2-(b). Loop Gating on Uncompressed Base

To isolate the effect of Loop Gating from KV eviction, we applied Loop Gating to the uncompressed Base agents and compared it with Loop Gating under Search-Selective KV Eviction, using the same hyperparameters as in the main experiments (τ=0.7\tau=0.7; N=5N=5 for WebExplorer-8B and N=4N=4 for WebSailor-32B).

Table oa8D-2. Loop Gating with and without KV eviction on the benchmarks of the main table (avg@3; gated rows replayed on ungated trajectories).

AgentSettingLoops / traj.Gate firingAvg.Rounds
WebExplorer-8BBase1.151.15–51.2351.238.688.68
Base + Loop Gating–5.65.6%50.9250.92 (−0.31)(-0.31)8.338.33 (−0.35)(-0.35)
KV Evict1.451.45–52.4352.439.659.65
KV Evict + Loop Gating–8.78.7%51.9151.91 (−0.52)(-0.52)8.978.97 (−0.68)(-0.68)
WebSailor-32BBase0.580.58–52.8152.815.645.64
Base + Loop Gating–4.64.6%52.7452.74 (−0.07)(-0.07)5.545.54 (−0.10)(-0.10)
KV Evict1.071.07–52.4752.475.885.88
KV Evict + Loop Gating–12.012.0%52.5952.59 (+0.12)(+0.12)5.635.63 (−0.25)(-0.25)

KV eviction increases the average number of Loops per trajectory (1.15→1.451.15 \rightarrow 1.45 on WebExplorer-8B and 0.58→1.070.58 \rightarrow 1.07 on WebSailor-32B). Consequently, Loop Gating is triggered more frequently under eviction (8.78.7% and 12.012.0% of trajectories) than without eviction (5.65.6% and 4.64.6%). More importantly, Loop Gating reduces 1.9×1.9\times and 2.5×2.5\times as many rounds under eviction, with only minor changes in accuracy. This brings the average round count back to within 3.33.3% and 0.20.2% of Base, respectively. These results show that while Loop Gating can also reduce redundant exploration in uncompressed agents, it is particularly effective at controlling the additional repetition induced by KV eviction.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their valuable comments. We have uploaded a revised paper, with changes shown in blue; the revisions made in response to each review are described in the individual responses.

Corrections. While preparing the revision, we found and corrected the following:

  • KV-cache metrics (Table 2 and the reductions in the Abstract, Section 4.2, and Conclusion): KV$_{avg}$ previously averaged KV usage over all rounds; it is now the mean over questions of the per-question peak, consistent with the definition in Section 4.1. In the original submission, token counts for WebSailor-32B were estimated with tiktoken; we now compute them with the agent's own Qwen2.5 tokenizer.
  • Tables 2 and 3: we corrected aggregation errors in the avg@3 of SearchTrimmer on WebSailor-32B (xbench-DeepSearch), in the rounds of FastKV and ChunkKV on WebSailor-32B (WebWalkerQA and xbench-DeepSearch), and in the Round and S / B values of the WebSailor-32B rows in Table 3, as well as a typo in the pass@3 of SWE-style masking on GAIA.
  • Throughput at a fixed memory budget (Section 4.3, Figure 6, Table 21): we corrected the peak KV memory of SearchTrimmer at eight and ten workers, which had been entered incorrectly. The wall-clock times and the resulting speedups (1.191.19–1.42×1.42\times at ten workers) are unchanged.

None of these corrections changes the conclusions of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.