acceptodds
Under review as a conference paper at ICLR 2027

QRAE: QUERY-RELEVANT ATTENTION ENTROPY FOR CONTEXT COMPRESSION IN DEEPSEARCH AGENTS

Abstract

DeepSearch LLM agents repeatedly search and carry retrieved sources into subsequent reasoning steps. This expanding context increases token consumption and latency, although many results do not contribute to later reasoning. We propose **Query-Relevant Attention Entropy (QRAE)**, an online context compression method that uses sparse evidence-assessment heads to decide which intact retrieved chunks should persist across reasoning steps. It scores bounded candidate batches through paired prefills: a query-free control suppresses shared attention responses, while normalized entropy measures the breadth of the positive response within each chunk. High-scoring chunks enter a persistent evidence whitelist and are included in the reasoning context. Uncertain chunks remain outside the reasoning context but are retained in a candidate cache, where they can be retrieved for later reevaluation. QRAE requires neither an additional generative compression pass nor modification of the attention operator. Across multiple DeepSearch benchmarks and reasoning models, QRAE improves task success by an average of 2.35 percentage points and reduces end-to-end latency and reported token usage by an average of 30% and 45%, respectively. These benefits generalize to a model with nearly 100B parameters.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Context Compression, Agent Context Management, Efficient LLM Inference, Mechanistic Interpretability

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

The paper proposes QRAE, which is an online context-compression method for deep search agents. It uses attention entropy from selected model heads to score retrieved chunks. It adds high-scoring chunks to whitelist and retains lower ones in a recoverable cache. Experiments show reductions in token usage and latency with comparable performance.

Critical strengths
  1. QRAE addresses an important and practical problem.
  2. The proposed method achieves consistent token and latency reductions on DeepDiver-7B.
Critical weaknesses
  1. The paper does not compare alternative scorers within the same context manager. It is unclear whether the gains come from the attention-entropy signal or mainly from the manager.
  2. The source of the token savings is unclear because compression effects are mixed with cross-step deduplication, fewer searches, and shorter trajectories, which weakens the main efficiency claim.
Questions for authors
  1. What exactly is Q in the evaluator at each step?
  2. Does the current trajectory state affect rescoring?
  3. How do alternative scorers such as Qwen3-Reranker, QRHead, or random scoring perform within the same context manager?
  4. Are cross-step deduplication and search behavior matched across methods?
  5. How much of the token reduction comes from context filtering versus fewer search calls and shorter trajectories?
  6. Please provide the code and data for review.
4 comments

We sincerely thank the reviewer for the detailed feedback, which has helped us further strengthen the demonstration of our method's practicality and effectiveness. Below, we provide our detailed point-by-point responses to each comment. For ease of reference, all corresponding revisions in the updated manuscript are highlighted in blue. Please feel free to let us know if any further clarification is needed.

W1 & Q3: Comparing alternative scorers within the same context manager. The paper does not compare alternative scorers within the same context manager, making it unclear whether the gains come from the attention-entropy signal or mainly from the manager. How do alternative scorers such as Qwen3-Reranker, QRHead, or random scoring perform within the same context manager?

A1: We sincerely thank the reviewer for pointing out this limitation. We agree that holding the context-management policy fixed is essential to isolate the contribution of the entropy-based scoring signal. Following your suggestion, we evaluated alternative scorers on BrowseComp-Plus across DeepDiver-7B, OpenResearcher-30B, and OpenPangu-2.0-Flash-92B (spanning 7B to nearly 100B parameters and diverse architectures) under the exact same QRAE context-management framework (identical persistent whitelist, candidate cache, reevaluation schedule, and retention budget). Specifically, we compare QRAE with Qwen3-Reranker and QRHead (full scoring) (the baselines from our paper), Retrieval order (selecting top-ranked chunks by their returned search order), and Attention sum (using the summed positive attention):

ModelMetricQwen3-RerankerQRHead (full scoring)Retrieval orderAttention sumQRAE
DeepDiver-7BSucc. (%) ↑14.912.713.314.617.2
Time (s) ↓33.342.731.627.320.5
Tokens ↓139k141k127k92k54k
OpenResearcher-30BSucc. (%) ↑51.251.450.452.453.0
Time (s) ↓174.3194.3227.7149.5134.7
Tokens ↓2056k1840k1947k1696k1453k
OpenPangu-2.0-Flash-92BSucc. (%) ↑54.354.753.554.955.3
Time (s) ↓320.5315.4395.1295.8255.4
Tokens ↓847k815k944k749k660k

Under the identical context manager, QRAE consistently achieves higher task success and substantially lower latency and token usage than all alternative signals across all three models, demonstrating the clear superiority of our entropy-based scoring signal.

Conversely, we also conducted the reverse comparison by stripping away QRAE's cache-and-whitelist manager and applying the QRAE signal solely to filter chunks within each individual search (QRAE (filtering only)), matching the per-search filtering setup of Rerank and QRHead in the paper:

ModelMetricRerankQRHeadQRAE (filtering only)
DeepDiver-7BSucc. (%) ↑14.711.615.4
Time (s) ↓35.645.127.6
Tokens ↓148k149k78k
OpenResearcher-30BSucc. (%) ↑50.751.152.1
Time (s) ↓162.0189.3144.5
Tokens ↓2157k1771k1588k
OpenPangu-2.0-Flash-92BSucc. (%) ↑54.353.855.0
Time (s) ↓344.7393.8286.4
Tokens ↓970k982k739k

Even under simple per-search filtering without the context manager, QRAE (filtering only) consistently outperforms Rerank and QRHead across all models in both accuracy and efficiency, further confirming the robustness of the QRAE signal. Moreover, comparing QRAE (filtering only) with the full QRAE in the first table shows that our context-management framework provides complementary gains: by ranking candidates globally across the cache rather than truncating within each individual search, it prevents relevant chunks from being permanently discarded when a single search returns more useful chunks than the per-step quota allows. We have incorporated the controlled comparison of different scoring signals under the same context-management framework into the ablation study in Section 4.4 of the revised manuscript.

W2 & Q5: Sources of token savings. Compression effects are mixed with cross-step deduplication, fewer searches, and shorter trajectories, leaving the source of token savings unclear and weakening the main efficiency claim. How much of the token reduction comes from context filtering versus fewer search calls and shorter trajectories?

A2: We thank the reviewer for this question. For cross-step deduplication, all evaluated methods share the exact same retrieval configuration and cross-step URL deduplication policy (see Appendix C.3 for details), so deduplication does not contribute to QRAE's relative token savings. Instead, the token reduction comes from two primary sources: (1) reduced per-step input volume, as QRAE selectively admits only high-scoring search chunks into the reasoning context while keeping uncertain candidates outside, and (2) shorter reasoning trajectories, as pruning irrelevant context noise enables the model to better leverage valid evidence and formulate more accurate search queries, thereby reaching the final answer in fewer steps.

To quantify the contribution of each source, we analyzed all trajectories on BrowseComp-Plus with DeepDiver-7B, recording the average cumulative tokens, number of reasoning steps, and per-step context growth:

MethodAvg. time (s)Avg. trajectory (tokens)Avg. stepsAvg. trajectory growth per step (tokens/step)
DeepSearch (no compression)29.5177k18.08763
QRAE (compression)20.554k12.07537

Based on these statistics, aligning the QRAE's step count to uncompressed baseline's 18.08 steps yields an estimated baseline cost of approximately 78k tokens. Accordingly, out of the total 123k token reduction (177k−54k177\text{k} - 54\text{k}), roughly 99k tokens (80.5%) stem directly from context filtering over search results, while roughly 24k tokens (19.5%) come from fewer reasoning and search steps. We have incorporated this breakdown into the revised manuscript.

Q1: What exactly is QQ in the evaluator at each step?

A3: At each step, QQ is always the original user question. By conditioning the evaluator on the original user question, we can filter for chunks that directly support or provide relevant evidence for solving the overall task. We apologize that our notation and wording in Section 3.2 were not sufficiently explicit on this point, and we have revised the manuscript to state this clearly.

Q2: Does the current trajectory state affect rescoring?

A4: No, the current trajectory state does not affect rescoring. The evaluator never uses the agent's previously retrieved context or intermediate reasoning progress as input; instead, it always scores candidate batches conditioned solely on the original user question QQ. Which chunks are selected for rescoring is determined strictly by their historical evaluation count, prioritizing candidates with the fewest prior evaluations.

Q4: Are cross-step deduplication and search behavior matched across methods?

A5: Yes, both cross-step deduplication and the underlying search framework are completely matched across all methods to ensure a fair comparison. Specifically, cross-step URL deduplication is consistently applied across every baseline and QRAE. Moreover, all methods follow the same Tongyi-DeepResearch agent framework and preserve the model's original think-act loop, without altering how the agent generates search actions.

Q6: Please provide the code and data for review.

A6: We have uploaded the core implementation reproducing our main results (Table 1) to an anonymous repository for review: https://anonymous.4open.science/r/QRAE-A337/. The remaining code, including offline head localization, experiments on the 30B or 92B model, and ablation studies, will be publicly released upon acceptance of the paper.

Reviewer_SJWmOpenReview ↗

Response from Reviewer SJWm

Thank you for the detailed response and the new experiments. I still have several concerns.

  1. Changed ablation numbers. In Table 5, the BrowseComp-Plus "Current-search filtering" row changed from 14.9 / 30.2 s / 129k to 15.4 / 27.6 s / 78k. It is not highlighted. The new numbers match the rebuttal's "QRAE (filtering only)". Please explain why this result changed.

  2. Token differences under the same manager. In Table 6, all scorers share the same manager and admit the same number of chunks per trigger. The per-step admitted volume should therefore be similar, yet QRAE uses 54k tokens and Qwen3-Reranker uses 139k. Please report the average number of steps, search calls, and admitted-chunk length for each scorer. Normalized entropy is length-sensitive (App. E.3), so QRAE may simply prefer shorter chunks.

  3. Token-saving decomposition. The 78k reference in D.9 seems to come from linear scaling. Because context accumulates, later steps cost more than average ones. With your reported growth of 537 tokens/step, 18.08 steps would give roughly 100–110k. That would put the split closer to 55/45 than 80/20. Please give the exact procedure, or compare cumulative tokens at matched step indices directly.

  4. Evaluator token count. Table 21 reports 4.9k compression tokens per BrowseComp-Plus trajectory. Table 15 reports 4.36 delayed admissions per trajectory with 3 admissions per trigger, which implies at least about 2.4 triggers on average. Each trigger runs two prefills over 5 chunks of up to 512 tokens. This suggests at least about 10k evaluator tokens. In the released code, the probe calls are also not added to token_usage. How were evaluator tokens measured, and are they included in Table 1?

  5. Rescoring. Since Q is fixed and the trajectory state is not used, a chunk's score changes only through batch composition. The paper still says chunks are reevaluated "when later evidence changes their utility" (Sec. 1, App. A.1). The score-once gains in Table 23 (0.2–0.7 pp, single run) are within the CI widths in Table 19. Please revise these claims.

  6. Cross-model claims. Table 19 shows no significant gain over DeepSearch on the 30B and 92B models. The abstract and conclusion still state "2.35 pp" and "benefits generalize to ~100B". The new claim about preliminary 200-step experiments (D.3) has no supporting data and should be removed or substantiated.

  7. Code. The BrowseComp-Plus README uses 256-token snippets and 20 turns, while the paper uses 512 tokens and 50 rounds. The baselines and the Table 6 variants are not included.

Thus, I keep my score.

All scorers use the same context manager and admission quota, and the retained chunks are comparable in length. We attribute the efficiency gap to how the selected evidence shapes the agent’s subsequent steps. QRAE selects chunks that more directly support the original query, helping the model draw more accurate inferences and choose more effective search actions. The agent therefore spends fewer think–act steps pursuing false leads. Each avoided step saves another pass over the growing context, substantially reducing token use and latency. This is consistent with context rot, where distracting passages can derail reasoning even when the necessary evidence is present.

Our central finding is that reasoning models encode a relevance signal in their own attention. We found this signal in all five models tested, and it became more discriminative with scale. The 7B or 30B or 92B model’s own signal outperformed a 0.6B reranker trained specifically for relevance scoring. DeepSearch context management is one application of this broader idea. By understanding how a model recognizes useful evidence, we can use that ability to improve its reasoning.

You can run the public anonymous code to evaluate QRAE directly and investigate the empirical questions you raised about evidence selection. The release predates the final implementation, but it uses exactly the same attention heads identified in DeepDiver-7B, L18H22 and L17H0. A single search is enough to inspect the signal. Run a search, examine how these heads respond to the returned chunks, and check whether QRAE ranks supporting evidence above distractors. The signal is visible regardless of the agent’s turn limit or how many chunks it subsequently admits. We released the code so readers can test whether QRAE reliably identifies useful evidence and whether using that evidence helps the model solve real search tasks.

Our goal is to improve model performance by understanding how a model recognizes useful evidence and putting that ability to work. DeepSearch context management is one concrete application. The agent framework and engineering optimizations fall outside the main focus of our work.

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

The paper proposes Query-Relevant Attention Entropy (QRAE), an online context compression framework for DeepSearch agents. QRAE first identifies a small set of evidence-assessment attention heads using answer-bearing examples from HotpotQA. During inference, it uses paired query-conditioned and query-free prefills to compute attention differences, followed by normalized entropy to estimate the relevance of retrieved text chunks. High-scoring chunks are admitted into a persistent evidence whitelist, while other chunks remain in a candidate cache for potential reevaluation. The method preserves the original retrieved text without requiring additional generative summarization.

The authors evaluate QRAE on four DeepSearch benchmarks using DeepDiver-7B, with additional experiments on OpenResearcher-30B and OpenPangu-92B. Across six model–benchmark configurations, QRAE reportedly improves task success by an average of 2.35 percentage points while reducing end-to-end latency by approximately 30% and processed tokens by 45%. The efficiency improvements are substantial, although the task-success gains are relatively modest and not consistent across all models.

Critical strengths

The paper addresses an important practical problem: the repeated processing of retrieved information across multiple reasoning steps. QRAE combines attention-based evidence scoring, persistent retention, and recoverable deferral without requiring additional generative summarization. This is a useful design for reducing inference overhead while preserving potentially relevant evidence.

The evaluation covers multiple benchmarks, model sizes, and architectures, including models with approximately 30B and 92B parameters. The reported reductions in token consumption and end-to-end latency are substantial. The component ablations and cross-model experiments further strengthen the empirical evaluation, even though the improvements in task success are less consistent.

Critical weaknesses
  1. The evidence for the proposed attention-based mechanism is not sufficiently conclusive. The paper's central mechanistic claim is that query-relevant attention entropy provides a reliable signal of semantic evidence relevance. However, the paired prefills read attention from different positions: the final query token for the query-conditioned input and the final candidate token for the query-free input (Section 3.2). Consequently, the resulting attention difference may reflect changes in the readout token, positional context, or attention normalization rather than query relevance alone. Although the theoretical analysis discusses control mismatch, it does not establish that this discrepancy is sufficiently small in the actual experiments.

  2. The experiments do not fully isolate the benefits of QRAE scoring from the context-management policy, and the accuracy improvements are not uniformly convincing. Although the ablations support the usefulness of individual scoring components, the comparisons with alternative retrieval methods do not consistently hold the context-management policy fixed. For example, QRAE employs persistent admission and recoverable caching, while the passage-selection baselines generally select chunks within individual searches. Therefore, some improvements may arise from the retention strategy rather than the proposed entropy-based relevance signal. A stronger evaluation would compare alternative relevance scores and inexpensive retrieval-order heuristics under the same caching, reevaluation, and admission policy.

Questions for authors

Since the query-conditioned and query-free prefills use different final-token readout positions, how do the authors distinguish query-induced relevance from positional and readout-token effects? Could the authors provide an empirical comparison using more closely matched readout conditions? Also, does the query used during online scoring refer to the original user question or the agent's current search query?

Could the authors compare QRAE with conventional relevance scoring, retrieval-order selection, and QRHead under exactly the same persistent whitelist, candidate cache, reevaluation schedule, and retention budget? This would help isolate the contribution of the proposed entropy-based scoring mechanism.

Since admission to the evidence whitelist is permanent, how does QRAE prevent irrelevant or outdated chunks from accumulating during long search trajectories? Is there an explicit upper bound on whitelist size, and how does the compression efficiency change as the number of search steps increases?

Could the authors provide paired confidence intervals or statistical significance tests for the reported task-success improvements, particularly for OpenResearcher-30B and OpenPangu-92B, where the differences are relatively small?

8 comments

We sincerely thank the reviewer for the detailed feedback, which has helped us further strengthen the demonstration of our method's practicality and effectiveness. Below, we provide our detailed point-by-point responses to each comment. For ease of reference, all corresponding revisions in the updated manuscript are highlighted in blue. Please feel free to let us know if any further clarification is needed.

W1 & Q1.1 & Q1.2: Readout mismatch in the paired prefills. The query-conditioned input reads attention from the final query token, whereas the query-free input reads from the final candidate token. The resulting attention difference may reflect changes in the readout token, positional context, or attention normalization rather than query relevance alone. Although the theoretical analysis discusses control mismatch, it does not establish that the discrepancy is sufficiently small in the actual experiments. How do the authors distinguish query-induced relevance from positional and readout-token effects? Could the authors provide an empirical comparison using more closely matched readout conditions?

A1: We thank the reviewer for this question regarding readout positions and control matching. We clarify that QRAE operates solely on a tiny subset of heads. Unlike positional or previous-token heads that are sensitive to local token offsets, these specialized heads strongly concentrate attention onto query-relevant spans upon processing the query. As illustrated in Figure 2(b), their positive attention responses emerge sharply on the evidence-bearing chunk. The query-free prefill serves as a paired control. While differing readout positions introduce minor variations in attention normalization, the magnitude gap between query-relevant and irrelevant chunks on these specialized heads is substantial. Because QRAE is used for relative chunk ranking, the query-induced signal easily dominates minor positional perturbations.

In fact, contrasting attention readouts with and without a query is a well-established practice in prior literature. For instance, one of the baselines evaluated in our paper, QRHead (Zhang et al., 2025), similarly employs specialized attention heads for passage reranking by subtracting the attention distribution of a null-query ("N/A") control, even though its readout length does not strictly match that of the actual query. The signals generated by query-focused attention heads are inherently robust, they do not fail due to differences in the readout position of the query-free control, much less do they reflect positional attention preferences.

To empirically verify that positional and readout-token differences do not drive our results, we conducted controlled readout-matching experiments on BrowseComp-Plus following the evidence-ranking setup in Tables 3 and 4 on our primary model DeepDiver-7B. Specifically, we compare:

  • QRAE (original readout): Our default setup with XtQ=[I;Bt;Q]\mathcal{X}_t^Q = [I; \mathcal{B}_t; Q] and Xt0=[I;Bt]\mathcal{X}_t^0 = [I; \mathcal{B}_t].
  • QRAE (shared-marker readout): Appending an identical readout marker token ⟨end⟩\langle\text{end}\rangle to both sequences, i.e., XtQ=[I;Bt;Q;⟨end⟩]\mathcal{X}_t^Q = [I; \mathcal{B}_t; Q; \langle\text{end}\rangle] and Xt0=[I;Bt;⟨end⟩]\mathcal{X}_t^0 = [I; \mathcal{B}_t; \langle\text{end}\rangle].
  • QRAE (shared marker + length-matched null): Replacing QQ in the query-free input with a placeholder sequence Q~\tilde{Q} of the exact same token length followed by the identical marker token, i.e., XtQ=[I;Bt;Q;⟨end⟩]\mathcal{X}_t^Q = [I; \mathcal{B}_t; Q; \langle\text{end}\rangle] and Xt0=[I;Bt;Q~;⟨end⟩]\mathcal{X}_t^0 = [I; \mathcal{B}_t; \tilde{Q}; \langle\text{end}\rangle], thereby matching both the readout token identity and its exact sequence position.

In addition, we include the baselines and ablations from the paper, removing query-free subtraction, applying QRAE scoring to Heads selected by QRHead and Random heads (see Tables 4 and 20 for details).

Variant5 chunks: AUC5 chunks: Recall@15 chunks: Recall@320 chunks: AUC20 chunks: Recall@120 chunks: Recall@10
QRAE (original readout)83.254.6%83.2%84.444.6%88.7%
QRAE (shared-marker readout)83.354.9%83.4%84.245.0%88.6%
QRAE (shared marker + length-matched null)83.454.8%83.6%85.044.8%89.2%
w/o query-free subtraction79.648.6%80.9%78.834.2%84.5%
Heads selected by QRHead68.324.5%70.0%70.413.9%63.3%
Random heads45.79.9%38.8%53.05.3%40.2%

As shown in the table, all three QRAE readout variants achieve virtually identical AUC and Recall across both candidate-set sizes, demonstrating that the scoring signal is insensitive to the exact readout position and comes from query-induced relevance. Furthermore, any variant using our localized heads substantially outperforms Heads selected by QRHead and Random heads, confirming the effectiveness and robustness of the signal. We have incorporated this controlled readout-matching analysis into the revised manuscript.

Q1.3: Does the query used during online scoring refer to the original user question or the agent's current search query?

A2: The query QQ used during online scoring always refers to the original user question. While the agent's search queries vary across steps to explore local sub-questions or specific keywords, scoring against the original user question keeps a consistent global objective, identifying chunks that directly support or provide relevant evidence for answering the overall question. We apologize for the ambiguity in Section 3.2 and have updated the manuscript to explicitly clarify this definition.

W2 & Q2: Isolating the scoring mechanism from context management. Although the ablations support individual scoring components, comparisons with alternative retrieval methods do not consistently hold the context-management policy fixed. QRAE employs persistent admission and recoverable caching, while passage-selection baselines generally select chunks within individual searches, so some gains may arise from the retention strategy rather than the entropy-based relevance signal. Could the authors compare QRAE with conventional relevance scoring, inexpensive retrieval-order selection, and QRHead under exactly the same persistent whitelist, candidate cache, reevaluation schedule, and retention budget?

A3: We sincerely thank the reviewer for pointing out this limitation. We agree that holding the context-management policy fixed is essential to isolate the contribution of the entropy-based scoring signal. Following your suggestion, we evaluated alternative scorers on BrowseComp-Plus across DeepDiver-7B, OpenResearcher-30B, and OpenPangu-2.0-Flash-92B (spanning 7B to nearly 100B parameters and diverse architectures) under the exact same QRAE context-management framework, identical persistent whitelist, candidate cache, reevaluation schedule, and retention budget. Specifically, we compare QRAE with Qwen3-Reranker and QRHead (the baselines from our paper), Retrieval order (selecting top-ranked chunks by their returned search order), and Attention sum (using the summed positive attention):

ModelMetricQwen3-RerankerQRHead (full scoring)Retrieval orderAttention sumQRAE
DeepDiver-7BSucc. (%) ↑14.912.713.314.617.2
Time (s) ↓33.342.731.627.320.5
Tokens ↓139k141k127k92k54k
OpenResearcher-30BSucc. (%) ↑51.251.450.452.453.0
Time (s) ↓174.3194.3227.7149.5134.7
Tokens ↓2056k1840k1947k1696k1453k
OpenPangu-2.0-Flash-92BSucc. (%) ↑54.354.753.554.955.3
Time (s) ↓320.5315.4395.1295.8255.4
Tokens ↓847k815k944k749k660k

Under the identical context manager, QRAE consistently achieves higher task success and substantially lower latency and token usage than all alternative signals across all three models, demonstrating the clear superiority of our entropy-based scoring signal.

Conversely, we also conducted the reverse comparison by stripping away QRAE's cache-and-whitelist manager and applying the QRAE signal solely to filter chunks within each individual search (filtering only), matching the per-search filtering setup of Rerank and QRHead in the paper:

ModelMetricRerankQRHeadQRAE (filtering only)
DeepDiver-7BSucc. (%) ↑14.711.615.4
Time (s) ↓35.645.127.6
Tokens ↓148k149k78k
OpenResearcher-30BSucc. (%) ↑50.751.152.1
Time (s) ↓162.0189.3144.5
Tokens ↓2157k1771k1588k
OpenPangu-2.0-Flash-92BSucc. (%) ↑54.353.855.0
Time (s) ↓344.7393.8286.4
Tokens ↓970k982k739k

Even under simple per-search filtering without the context manager, QRAE (filtering only) consistently outperforms Rerank and QRHead across all models in both accuracy and efficiency, further confirming the robustness of the QRAE signal. Moreover, comparing QRAE (filtering only) with the full QRAE in the first table shows that our context-management framework indeed provides complementary gains: by ranking candidates globally across the cache rather than truncating within each individual search, it prevents relevant chunks from being permanently discarded when a single search returns more useful chunks than the per-step quota allows. We have incorporated the controlled comparison of different scoring signals under the same context-management framework into the ablation study in Section 4.4 of the revised manuscript.

Q3: Since admission to the evidence whitelist is permanent, how does QRAE prevent irrelevant or outdated chunks from accumulating during long search trajectories? Is there an explicit upper bound on whitelist size, and how does the compression efficiency change as the number of search steps increases?

A4: In mainstream DeepSearch frameworks (such as the Tongyi-DeepResearch paradigm used in our work), retrieved documents are appended into the reasoning context (either in the main agent or a subagent) across steps. QRAE acts as a selective gate before chunks enter the reasoning context, admitting at most ⌈ρK⌉\lceil \rho K \rceil chunks per evaluation trigger with ρ=0.5\rho = 0.5. Although we do not impose an explicit hard cap on the whitelist size, its growth rate is bounded by at most half of the retrieval stream. Consequently, chunk accumulation under QRAE is strictly and substantially lower than in standard uncompressed agents.

To illustrate how compression efficiency evolves over long trajectories, we tracked the average cumulative context length and the resulting retention ratio at the end of each 2 steps from Step 1 to Step 20 on BrowseComp-Plus with DeepDiver-7B:

ConfigurationStep 2Step 4Step 6Step 8Step 10Step 12Step 14Step 16Step 18Step 20
Uncompressed2,5575,3957,6149,55210,99312,07412,94913,59714,25614,944
Compressed1,9573,3174,5955,7086,5546,9317,3107,6567,9958,316
Retention ratio76.5%61.5%60.3%59.8%59.6%57.4%56.5%56.3%56.1%55.6%

As shown in the table, the retention ratio is initially higher because both settings share the same fixed system prompt and user instruction. As the number of steps increases, the retention ratio steadily decreases and stabilizes around 55.6% by Step 20. Because QRAE applies a 50% retention rate specifically to retrieved search chunks—which dominate the context—while leaving system instructions, generated thoughts, and non-search tool outputs (e.g., lookup actions) intact, the asymptotic retention ratio naturally converges slightly above 50%.

Q4: Reliability of task-success improvements. Could the authors provide paired confidence intervals or statistical significance tests for the reported task-success improvements, particularly for OpenResearcher-30B and OpenPangu-92B, where the differences are relatively small?

A5: We thank the reviewer for this suggestion. Following your request, we report the task-success differences (Δ=QRAE−Baseline\Delta = \text{QRAE} - \text{Baseline}) along with their 95% paired confidence intervals on BrowseComp-Plus for OpenResearcher-30B and OpenPangu-2.0-Flash-92B in the table below (statistics for DeepDiver-7B are provided in Appendix D.3):

ModelBaselineSuccess difference Δ (pp)95% paired CI (pp)
OpenResearcher-30BDeepSearch (no compression)-0.2[-1.58, +1.18]
LLMLingua-2+0.9[-0.28, +2.08]
ACON+3.6[+2.03, +5.17]
Rerank+2.3[+0.69, +3.91]
QRHead+1.9[+0.25, +3.55]
SideQuest+5.4[+3.69, +7.11]
OpenPangu-2.0-Flash-92BDeepSearch (no compression)+0.4[-1.29, +2.09]
LLMLingua-2+1.1[-0.46, +2.66]
ACON+5.2[+3.64, +6.76]
Rerank+1.0[-0.42, +2.42]
QRHead+1.5[+0.21, +2.79]
SideQuest+4.0[+2.39, +5.61]

As shown in the table, QRAE achieves statistically significant improvements over most compression baselines (e.g., ACON, QRHead, SideQuest, and Rerank on 30B) while fully matching the uncompressed DeepSearch agent at a fraction of the token and latency cost. We have included these paired statistical significance results in Appendix D.3 of the revised manuscript.

Reviewer_Jy8mOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes an online context compression method that uses sparse evidence-assessment heads to decide which intact retrieved chunks should persist across reasoning steps. High-scoring chunks enter a persistent evidence whitelist and are included in the reasoning context. It also shows improvements on several benchmarks.

Critical strengths
  1. The paper study an important and practical problem
  2. The evaluation is relatively broad and also have ablation study.
  3. Keeping the original test check is good for preserving the original evidence and source information.
  4. The efficiency gains are substantial and encouraging.
Critical weaknesses

If each search returns K new candidates that have not appeared before, then the scoring step is only applied to those K new candidates. A candidate that was scored earlier but not admitted at that time is not scored again. In that case, an old chunk can still enter the whitelist later. This is because admission is based on the ranking over the full cache of scored candidates. Once some higher-ranked candidates are admitted and removed from consideration, an older candidate can move up in the ranking. So I think there are two different effects that need to be separately discuss:

  1. evidence being admitted later;
  2. evidence being rescored and then admitted because its score changes. You can test an ablation: Score each chunk only once, while keeping the cache, whitelist, ranking over the full cache, and admission quota exactly the same.
Questions for authors

What exactly is Q during scoring: the original user question, the current search query, or a dynamically updated reasoning state?

4 comments

We sincerely thank the reviewer for the positive assessment and detailed feedback. We greatly appreciate your constructive insights, which have helped us improve the presentation of our method. We have carefully addressed every comment and revised the manuscript accordingly, with all changes highlighted in blue. We hope our clarifications resolve any remaining concerns, and we would be glad to provide further information if needed.

W1: Delayed admission versus rescoring. If each search returns KK new candidates that have not appeared before, then the scoring step is only applied to those KK new candidates. A candidate that was scored earlier but not admitted at that time is not scored again. An old chunk can still enter the whitelist later because admission ranks the full cache of scored candidates: once higher-ranked candidates are admitted and removed from consideration, an older candidate can move up in the ranking. Please distinguish evidence being admitted later from evidence being rescored and then admitted because its score changes. Test an ablation that scores each chunk only once while keeping the cache, whitelist, ranking over the full cache, and admission quota exactly the same.

A1: We sincerely thank the reviewer for this insightful observation and careful reading of our mechanism. As the reviewer accurately noted, both effects coexist: (1) delayed admission through full-cache ranking promotion once higher-ranked chunks enter the whitelist, and (2) admission triggered by updated scores after rescoring. In fact, ranking across the global cache is an intentional design choice because the number of useful chunks returned by a single search is unknown a priori. Consider an extreme case where all KK results from one search are essential for reasoning, whereas a subsequent search with a modified query returns entirely irrelevant results. A per-search selection policy (e.g., retaining only the top 50% of each search) would permanently discard half of the useful evidence from the first search. Global ranking across the full cache ensures that candidates from different steps are compared horizontally, allowing temporarily deferred relevant chunks from information-dense searches to be admitted in later steps.

To isolate the two effects as suggested, we evaluated an ablation that scores each chunk only once (score-once) while keeping the candidate cache, persistent whitelist, global cache ranking, and admission quota identical. We conducted this experiment on the most challenging benchmark, BrowseComp-Plus, across DeepDiver-7B, OpenResearcher-30B, and OpenPangu-2.0-Flash-92B (spanning 7B to nearly 100B parameters and distinct model architectures):

ModelMetricQRAE (score-once)QRAE (full)
DeepDiver-7BSucc. (%) ↑16.517.2
Time (s) ↓21.520.5
Tokens ↓59k54k
OpenResearcher-30BSucc. (%) ↑52.753.0
Time (s) ↓141.5134.7
Tokens ↓1549k1453k
OpenPangu-2.0-Flash-92BSucc. (%) ↑55.155.3
Time (s) ↓270.6255.4
Tokens ↓687k660k

As shown in the table, QRAE (score-once) still substantially outperforms the baselines (cf. Tables 1 and 2), confirming that global cache ranking with delayed admission already captures most of the gains, while QRAE (full) consistently achieves higher task success with lower latency and token usage across all three models. By inspecting the agent trajectories, we find that the performance gap mainly stems from losing the rescoring mechanism. Because attention responses are normalized across the evaluation batch Bt\mathcal{B}_t before computing positive differences, when a single search retrieves multiple relevant chunks together, they compete for attention mass within the same batch; consequently, some valid evidence chunks have their positive responses compressed and receive lower initial QRAE scores. When such a chunk is later rescored alongside mostly irrelevant candidates in a new batch, its positive attention breadth and QRAE score increase markedly, enabling admission. Because retrieving multiple relevant chunks within a single search occurs in a relatively small fraction of steps, removing rescoring causes a modest performance drop, while confirming that rescoring provides consistent additional gains in both accuracy and efficiency. We have incorporated this distinction and ablation into the revised manuscript.

Q1: What exactly is QQ during scoring: the original user question, the current search query, or a dynamically updated reasoning state?

A2: During online scoring, QQ is always the original user question. Because the agent's intermediate search queries often focus on local sub-questions or exploratory keywords, conditioning the evaluator on the original user question ensures a consistent global objective, identifying chunks that directly support or provide relevant evidence for solving the overall task across steps. We apologize that our notation and wording in Section 3.2 were not sufficiently explicit on this point, and we have revised the manuscript to state this clearly.

Reviewer_Jy8mOpenReview ↗

Response to authors

Thank you for your response. I will maintain my score of 3.

Thank you for your positive assessment of our work. We would be happy to address any further questions or provide additional clarification.

Comments on the paper

1 comment

Dear Area Chair,

We thank the reviewers for their thoughtful feedback. All three recognized that the paper addresses an important practical problem. We responded to their comments point by point and added the following comparisons and analyses.

  • With the context manager held fixed, QRAE achieved higher task success, lower token use, and lower latency than all four alternative scorers on the 7B, 30B, and 92B models.
  • When all methods filtered each search independently, QRAE retained its advantage over Qwen3-Reranker and QRHead on all three models.
  • We added paired 95% confidence intervals for task-success differences at 30B and 92B.
  • Additional ablations and trajectory analyses separated rescoring from delayed admission, tested sensitivity to the attention readout token and position, and examined the sources of token savings.

The reviewers’ main remaining concern is that the source of the observed efficiency gap between QRAE and the other scoring methods is unclear. We attribute the savings primarily to QRAE selecting chunks that more directly support the original query. With better evidence, the model can draw more accurate inferences and choose more effective search actions. It therefore needs fewer think–act steps, reducing token use and latency. This interpretation is consistent with context rot, where distracting passages can derail reasoning even when useful evidence is present.

Our goal is to improve model performance by understanding how a model recognizes useful evidence and putting that ability to work. DeepSearch is one application of this insight. The agent framework and its engineering optimizations are secondary to that scientific goal. We hope these clarifications help inform your decision.

Thank you very much for your time and consideration.

Best Regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.