acceptodds
Under review as a conference paper at ICLR 2027

SAGE-RAG: Defending Retrieval-Augmented Generation with Trust Calibration and Graph Selection

10 views

Abstract

Retrieval-augmented generation (RAG) augments language model inputs with retrieved evidence to improve performance on knowledge-intensive tasks. When attackers can inject content into the retrieval corpus, however, this reliance on external evidence creates opportunities to manipulate generated answers. We study coordinated evidence poisoning, in which individually plausible and relevant passages reinforce misleading claims that conflict with benign evidence. Such attacks motivate assessing relations among claims, since local plausibility and apparent agreement alone do not establish trustworthiness. We propose SAGE-RAG, which represents retrieved claims and their relations in an evidence graph and combines trust calibration with joint evidence selection. Attack-Aware Trust Propagation (AATP) uses structural coordination patterns as signals to calibrate claim trust. Maximum Consistent Reasoning Subgraph selection (MCRS) then jointly selects claims under a token budget by balancing utility, calibrated trust, and relational consistency. It downweights claims with low trust and penalizes the joint selection of contradictory claims. On an 8,000-sample benchmark spanning four QA datasets and four attack types, SAGE-RAG reduces attack absorption from with BM25 to and achieves the highest exact-match (EM) score among ten evaluated systems, reaching compared with for RobustRAG. Additional evaluations show gains across four generators and on the RGB benchmark. Ablations show that trust calibration helps retain benign evidence, while joint selection further reduces the inclusion of adversarial claims.

Primary area:
generative models
Keywords:
Retrieval-Augmented Generation (RAG), Coordinated Evidence Poisoning, Adversarial Robustness, Evidence Graph, Trust Propagation, Knowledge Conflict Resolution

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_2kPiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies coordinated evidence poisoning in retrieval-augmented generation, where individually plausible passages collectively reinforce misleading claims. It proposes SAGE-RAG, which represents retrieved evidence as a graph of atomic claims connected by support, contradiction, and relatedness relations. The framework combines Attack-Aware Trust Propagation (AATP) to assess claim trust with Maximum Consistent Reasoning Subgraph (MCRS) selection to balance utility, trust, and consistency under an evidence budget. The resulting evidence is organized into structured context for a frozen generator. Evaluation includes adversarial QA benchmarks, multiple generators, component ablations, adaptive attacks, and qualitative failure analyses.

Critical strengths
  1. The paper addresses a meaningful failure mode of evidence aggregation. Individually relevant passages can collectively create misleading corroboration, making passage-level relevance insufficient for assessing evidence quality. Representing atomic claims and their relations provides a useful framework for studying this problem. This framing is consequential for RAG systems that combine evidence from multiple potentially untrusted sources.

  2. The selection method balances usefulness, trust, and consistency. MCRS considers both individual claim quality and how claims support or contradict one another. This is useful for selecting evidence that works together, particularly for multi-hop questions.

  3. The evaluation covers a broad range of settings. The experiments include multiple QA datasets, attack types, and generators, with comparisons against existing retrieval and defense methods. Additional robustness tests, component ablations, and case studies provide useful evidence about the method’s effectiveness, the roles of its components, and its limitations.

Critical weaknesses
  1. Parameter selection and sensitivity remain unclear. The method involves multiple objective weights and detection thresholds. Although the same settings are used across datasets, their selection rationale and the sensitivity of the results remain unclear. The authors should explain how these parameters were determined and examine whether answer quality and defense effectiveness remain stable as key weights and detection thresholds vary within reasonable ranges.

  2. The construction of coordinated attacks in the main benchmark is insufficiently described. Although the appendices provide attack examples and some adaptive attack strategies, it remains unclear how the malicious documents in the main experiments were generated and coordinated to reinforce the same incorrect conclusion. The authors should clarify their sources and construction procedure, and whether the samples were checked against the stated conditions of coordinated evidence poisoning. This would help establish how well the evaluation covers the intended threat model.

  3. Clarification of EM and CM Evaluation. Could the authors clarify why EM exceeds CM for TrustRAG on TriviaQA in Table 6 (0.320 vs. 0.295)? Under the stated definitions, an exact match should also satisfy contains match when both metrics use the same predictions, references, and normalization. Are different evaluation subsets or aggregation procedures used?

Questions for authors

I would appreciate clarification on the issues raised in the Critical Weaknesses section.

3 comments

Response to Reviewer

W1. Parameter selection and sensitivity

We thank the reviewer for the constructive request to justify parameter selection and characterize sensitivity. To address this point, we added a one-at-a-time analysis of the seven objective parameters α\alpha, β\beta, γ\gamma, δ\delta, η\eta, λ\lambda, and κ\kappa on 2,000 questions (500 per dataset). We vary each parameter at 0.50, 0.75, 1.00, 1.25, and 1.50 times its Table 4 default while fixing the other settings. The defaults were chosen to balance adversarial exclusion and benign retention and are used unchanged across all datasets and attack types. The analysis measures how both selection metrics and answer EM change around the deployed configuration.

One-at-a-time parameter sensitivity. The analysis uses the seven objective parameters and 2,000 questions.

ParameterDefaultMax ∣Δ Selection F1∣\lvert\Delta\,\mathrm{Selection\ F1}\rvertRange across the five values
α\alpha (utility)1.00.0540.740–0.793
β\beta (trust)1.00.0780.716–0.793
γ\gamma (SUPPORT bonus)0.50.0410.753–0.793
δ\delta (CONTRADICTION penalty)1.50.0760.717–0.793
η\eta (per-token cost)0.010.0390.754–0.793
λ\lambda (contradiction density)0.80.0930.700–0.793
κ\kappa (trust-gating exponent)2.50.0630.731–0.793

At the default setting, exclusion/retention/Selection F1 are 0.791/0.804/0.793. In the one-at-a-time sweep, the default is the best of the five tested values for every parameter. The parameters differ in sensitivity. λ\lambda, β\beta, and δ\delta have the largest effects, with maximum ∣Δ Selection F1∣\lvert\Delta\,\mathrm{Selection\ F1}\rvert values of 0.093, 0.078, and 0.076, respectively. η\eta, γ\gamma, and α\alpha are milder, with values of 0.039, 0.041, and 0.054, while κ\kappa is intermediate (0.063). Stronger penalties improve exclusion but cost benign retention. For example, halving β\beta raises exclusion from 0.791 to 0.854 but lowers retention from 0.804 to 0.622. Increasing β\beta by 50% does the reverse, with exclusion at 0.642 and retention at 0.846. The sweep also records EM. The default has the highest EM among the tested values for every parameter, and the largest EM drop is 0.054 (λ\lambda).

The defaults in Table 4 are fixed across all datasets and attack types. We will add the full five-point sweep to the supplementary appendix, with a short pointer in Section 3.3. The sweep covers the seven objective weights, with the detection thresholds fixed at their Table 4 values.

Response to Reviewer — continued

W2. Attack-data sources, construction, and coordination

We thank the reviewer for the helpful request for greater clarity on attack-data sources, construction, and coordination. To clarify, Section 2 defines the threat model. An attacker inserts passages into the retrieval corpus to steer the answer toward a target answer, and the defender sees only the query and the retrieved documents. This follows PoisonedRAG (Zou et al., 2025), which demonstrates that five malicious texts per target question can achieve about 90% attack success in a knowledge database with millions of texts. The benchmark instantiates this threat model with a controlled construction. It is not produced by PoisonedRAG's official attack optimizer. A subset of the passages follows PoisonedRAG's P=Q+IP=Q+I construction. For each question, we set a common incorrect answer and steer the malicious passages toward it. The attack styles are surface realizations of that target. This choice keeps the generation procedure fixed and the evidence pool identical across methods.

Construction. Questions and gold answers come from NQ, TriviaQA, HotPotQA, and MuSiQue. Benign documents come from the contexts released with each dataset. For HotPotQA, these are the supporting and distractor paragraphs. For MuSiQue, they are the supporting and non-supporting paragraphs. For TriviaQA, we use Wikipedia contexts. For NQ, we use the released answer candidates because its open release provides no contexts. For each question, we set a common incorrect answer derived from the gold answer by a controlled perturbation. We then construct at least three malicious passages that steer toward it and are designed to support one another. The passages use template-based instruction injection and controlled false factual statements. A subset follows PoisonedRAG's P=Q+IP=Q+I construction. Construction is human-guided. We fix the target answer and the coordination goal, and the passages are generated by a controlled pipeline rather than by PoisonedRAG's official attack optimizer. The passages are shuffled into the same candidate pool as the benign documents. Every method sees the same pool and the same top-K=4K=4 retrieved set.

Coordination. Coordination is generated at the answer level. The passages are steered toward a common incorrect answer and placed in the same candidate pool, so they jointly compete with the benign evidence. The attack styles are surface realizations and are not the coordination variable. We verified the resulting document-level structure. Specifically, 42.93% of the benchmark documents are malicious documents participating in cross-document SUPPORT components. These are SUPPORT-connected components with malicious claims from at least two different malicious documents. The remaining 57.07% are benign documents. The malicious documents therefore form coordinated support structures, while the benign documents provide the competing evidence.

Validation. We checked construction automatically for schema validity, multiple malicious passages per sample generated by different mechanisms, absence of duplicate documents and meta-commentary, and the presence of instruction cues in template-generated passages. Samples were also spot-checked manually. Table 2 reports graph-level attack recall and benign retention, which quantify how much of the benchmark the structural signals capture.

W3. The TrustRAG entry in Table 6

We thank the reviewer for catching this issue. We agree that, under the stated definitions, an exact match implies a contains match when predictions, references, and normalization are the same. The TrustRAG/TriviaQA entry is therefore internally inconsistent. We checked the per-dataset results and found that the EM and F1 values had been swapped. The correct values are EM = 0.148, F1 = 0.320, and CM = 0.295. The remaining rows of Table 6 satisfy EM≤CM\mathrm{EM}\leq\mathrm{CM}. We will correct the two cells, the sentence that cites TrustRAG's EM as 0.320, and the ranking marks. After the correction, RobustRAG has the highest EM on this split (0.202), while TrustRAG keeps the lowest ASR (0.060). This is a cell-level error in one row of Table 6, not a difference in evaluation subsets or metric definitions. We are sorry for this error and appreciate the reviewer's attention to this detail.

We hope that our responses have addressed the reviewers' concerns. The new tables and future-work notes will be included in the revision. We remain happy to clarify any remaining questions during the discussion period.

Thank you for providing the additional detail on the construction of the coordinated attack benchmark and for conducting further parameter sensitivity analyzing.

These clarification and additional experiment have adequately addressed our main concern.

Reviewer_PG2TOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The motivation in this paper is robustness to injection attacks (coordinated evidence poisoning) in RAG. The authors utilize an evidence graph to represent relations between claims in the retrieved passage, and they use this graph to calibrate trust. In order to make use of the trust measure, the authors also introduce a joint claim selection mechanism to balance trust and relevance of selected claims.

Critical strengths
  1. Problem motivation is framed well and explained thoroughly
  2. The multi-faceted selection mechanism is sensible.
  3. Evaluation is broad, authors used multiple benchmarks, QA datasets, and attack types
  4. The authors are transparent about limitations in the discussion, and report zero-shot performance in comparisons.
Critical weaknesses
  1. Potential inconsistency in some metrics between appendix (table17) and main paper (table1). Needs clarification on difference in experimental settings.
  2. Some terms need additional elaboration, especially 3H judged-subset protocol. If this is a specific protocol from a previous study, it needs to be stated with an in-text citation, I was not able to find this.
  3. No confidence interval anywhere. Temperature of 0 with API models is not guaranteed to be fully deterministic. With long prompts or complex reasoning, the mismatch can be significant across sequential runs at 0 temprature.
  4. Zero-shot performance is better in several metrics, indicating the problem subset should be more selective. The authors actually do perform the necessary problem selection for case studies in the appendix, but they do not do this selection for the main experiment.

See details in next section

Questions for authors

Specific questions:

  1. For SAGE-RAG + GPT-5.4mini performance, table 1 says(EM: 0.451, F1:0.564) table 17 says (EM:0.32, F1:0.433). There is a similar issue for zero-shot gpt-5.4 mini. What is the setting difference between table 17 and table 1? Is it specified in the paper?

  2. Table 1 says it follows the 3H judged-subset protocol, I understand 3H here is likely “helpful,honest,harmless”, I cannot infer the detailed design from “judged-subset protocol”. If this is a specific protocol established in another paper, the authors should provide an in-text citation. If this is a protocol the authors built, the method section should include a detailed description of the protocol.

  3. The paper uses a temperature of 0 with the gpt-5.4 mini api, but api-based generation is not deterministic even with temperature of 0. Since the author use this model for both claim extraction and answer generation, the randomness compounds across stages. I believe this warrants some multi-run testing and reporting of variance across-runs.

  4. The authors are transparent with zero-shot performance being higher than any retrieval methods in table 1 for F1, ASR, and Margin. In table 7, for HotPotQA, even the EM is higher for zero shot than the retreival methods. This suggests to me that some of the questions in their benchmark is sufficiently solved with parametric knowledge, and the authors enforce indiscriminate retrieval. Since the authors use RAG-injection attacks as their motivation, it would make sense to evaluate on a question subset where parametric knowledge alone is insufficient, and retrieval is more warranted. I see the authors do describe the need for retrieval in case studies. The authors state for the case studies, “The three cases are re-selected under a criterion the earlier version lacked. Each question must be unanswerable without retrieval”. Why do the authors not bring this criterion to their main experiment? It seems like this criterion is highly relevant to the problem statement.

1 comment

Response to Reviewer

W1 (Q1). Table 1 versus Table 17

We thank the reviewer for this careful comparison of Tables 1 and 17. To clarify, Table 1 reports the main 8,000-question evaluation, with 2,000 questions per dataset. Table 17 reports a separate 2,000-question multi-model evaluation under the common-evidence setting of Appendix L. The two tables therefore describe different evaluation populations (2k versus 8k) and settings. They are not repeated measurements of the same questions, and scores should be compared within each table.

W2 (Q2). The 3H criteria and the Help/Hon evaluation

We thank the reviewer for the opportunity to clarify the 3H criteria and the Help/Hon evaluation. The 3H criteria are helpfulness, honesty, and harmlessness, following the HHH framework (Askell et al., 2021). We evaluate helpfulness and honesty, while harmlessness is outside the scope of this study. GPT-5.4 Mini scores the full set of outputs for each method on a 1–5 scale. The two dimensions are aggregated separately by arithmetic mean. The Table 1 caption and Appendix D state the model, coverage, scale, and aggregation.

W3 (Q3). Nondeterminism at temperature 0

We thank the reviewer for raising this important point about nondeterminism at temperature 0. We agree that temperature 0 does not make API inference deterministic. To address this concern, we ran three independent end-to-end runs, each on the full 8,000-question benchmark. Each run re-executed both claim extraction and answer generation rather than reusing earlier outputs. We held the data, prompts, and parameter configuration fixed.

Three independent end-to-end runs. The evaluation uses the full 8,000-question benchmark.

RunEMAnswer-F1CM
10.45030.55980.5203
20.45490.56450.5254
30.45200.56260.5221
Mean ±\pm std0.4524±0.00230.4524\pm0.00230.5623±0.00240.5623\pm0.00240.5226±0.00260.5226\pm0.0026
Min–max0.4503–0.45490.5598–0.56450.5203–0.5254

The table above reports the mean and sample standard deviation (ddof = 1) across the three runs. The submitted Table 1 values (0.451/0.564/0.521) fall within the observed run ranges: EM 0.4503–0.4549, answer-F1 0.5598–0.5645, and CM 0.5203–0.5254. The main results are therefore consistent with run-to-run variation. We will add the per-run table to the supplementary appendix and summarize it in Section 5.

W4 (Q4). Why the main benchmark is not filtered by retrieval dependence

We thank the reviewer for this thoughtful question about retrieval dependence. To clarify, zero-shot is included as a no-document reference. Its answer quality is informative. However, a low attack-absorption rate without document exposure is not evidence of a grounded defense. We keep the standard public QA distribution in the main benchmark rather than filtering questions by whether a particular generator can answer them without retrieval. The goal is to measure robustness of a retrieval-enabled pipeline under poisoned evidence on a common distribution.

Retrieval dependence is model- and prompt-dependent. Filtering by a particular generator's no-document failures would therefore change the evaluation population and complicate cross-model comparisons. The case studies use a no-document failure criterion only to select illustrative examples.

We consider a model-aware and time-aware retrieval-dependence benchmark a natural extension and add it as future work.

Reviewer_V4DwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates a RAG poisoning scenario in which attackers inject mutually reinforcing malicious evidence into an external knowledge corpus, contradicting legitimate evidence. To defend against such attacks, the authors propose SAGE-RAG, which consists of two key modules: (1) AATP, which evaluates evidence trustworthiness through an enhanced random-walk-based trust propagation mechanism, and (2) MCRS, which formulates joint evidence selection as a constrained integer programming problem. Experiments across multiple baselines show promising results, while ablation studies examine the contributions of the key modules and their components.

Critical strengths

S1. The paper moves beyond evaluating individual pieces of evidence independently by explicitly modeling their relationships. This allows evidence trustworthiness to be assessed based on both individual characteristics and interactions with other evidence, which is particularly relevant for detecting mutually reinforcing malicious evidence.

S2. Complementary module design. AATP and MCRS complement each other, forming a coherent and comprehensive methodological framework.

S3. Competitive performance and comprehensive evaluation. The paper includes a sufficient set of competitive baselines and achieves promising results. The experiments cover a broad range of settings, with extensive ablation studies and robustness analyses that further validate the effectiveness of the proposed approach.

Critical weaknesses

W1. The method assumes that malicious evidence mutually reinforces itself while contradicting benign evidence. However, benign evidence may also conflict due to temporal changes, and malicious evidence may not explicitly contradict legitimate evidence. Without external ground truth, how can the method distinguish benign and malicious subgraphs with similar structural characteristics?

W2. AATP and MCRS largely rely on existing techniques and heuristic designs, making their methodological novelty limited. Moreover, the ablations compare each module with only one alternative, without adequately demonstrating the necessity or superiority of the proposed designs over other competitive alternatives.

Questions for authors
  1. Could the authors provide empirical evidence demonstrating how commonly the assumed structural patterns occur in real-world scenarios and whether the underlying assumptions generally hold?

  2. How could the proposed method be improved to defend against adaptive attacks, particularly those that exploit the structural assumptions underlying the defense mechanism?

  3. What alternatives to PageRank could be considered for trust propagation? Why is integer linear programming (ILP) preferred over simpler optimization strategies, such as greedy selection?

2 comments

Response to Reviewer

W1. Temporal change and structurally indistinguishable evidence

We thank the reviewer for raising this important point about temporal change and structurally indistinguishable evidence. We agree that benign evidence can conflict for reasons unrelated to an attack, including temporal change. We also agree that a malicious subgraph need not explicitly contradict benign evidence. To clarify the scope of the method, Sections 3.2 and 3.2.3 use structural signals to calibrate trust, rather than to certify factual correctness. At inference, the defender has no reference answer or adversarial labels. SAGE-RAG lowers the trust of structurally suspicious evidence and does not assert that the remaining evidence is true.

On the RGB counterfactual task, conflicts come from tampered facts rather than coordinated poisoning. SAGE-RAG detects 94.0% of counterfactuals and corrects 91.5% of detections. This result shows that the pipeline is not limited to the coordinated pattern, although coordinated poisoning remains its target. Temporal validity is a separate source of conflict outside the current evaluation, and a deployment-level study of temporal conflicts is future work.

W2 and Q3. Component necessity, alternative designs, and the choice of ILP

We thank the reviewer for the constructive request to examine component necessity and alternative designs. We agree that k-core decomposition, restart random walks, and ILP are not new algorithms. Our contribution lies in the problem-specific coupling: calibrated trust gates claim utility, and the pairwise support/contradiction terms make the value of a claim depend on the other claims selected under the budget. As a result, selection can balance adversarial exclusion against retention of benign evidence.

To address the concern about component necessity, we added a 2,000-question comparison (500 per dataset) covering each module and the main alternatives. The table below reports the results. Full SAGE-RAG reaches EM/answer-F1 0.457/0.569. AATP-only reaches 0.428/0.540, and PageRank+ILP reaches 0.407/0.523. Greedy selection gives 0.407/0.520. The selection-level columns help explain these gaps. Full excludes 0.791 of adversarial claims while retaining 0.804 of benign claims, for a Selection F1 of 0.793. AATP-only, which removes MCRS, excludes only 0.526 and retains 0.697, for a Selection F1 of 0.600. PageRank+ILP, which replaces AATP with classic signed PageRank, excludes more claims (0.978) but retains only 0.651 benign claims, for a Selection F1 of 0.782. Its answer quality is also lower. Greedy does the opposite: it retains 0.878 benign claims but excludes only 0.548 adversarial claims, for a Selection F1 of 0.670. The joint objective is designed to control this trade-off. Maximizing exclusion alone is not the goal.

Component and alternative comparisons. The comparison uses 2,000 questions (500 per dataset). AATP-only and PageRank+ILP selection metrics follow Table 10.

VariantEMAnswer-F1Adv. exclusionBenign retainedSelection F1
Full0.4570.5690.7910.8040.793
AATP-only0.4280.5400.5260.6970.600
PageRank+ILP0.4070.5230.9780.6510.782
Greedy0.4070.5200.5480.8780.670
HITS0.4050.5250.7350.7520.738
Degree0.3830.5030.7300.7630.742
Uniform utility0.3740.4900.7430.6320.679

The trust-scoring alternatives are also below Full. HITS reaches EM 0.405 with a Selection F1 of 0.738. Degree-based scoring reaches 0.383 with 0.742, and uniform utility reaches 0.374 with 0.679. By comparison, Full reaches 0.457 with a Selection F1 of 0.793. Overall, these results support complementarity under the evaluated setting.

The comparison uses a 2,000-question subset and is separate from the 8,000-question main table. We add a short pointer in Section 5.3 and include the component/alternative table in the supplementary appendix.

Why ILP rather than greedy. In short, Equation 7 contains pairwise SUPPORT rewards and CONTRADICTION penalties, so the value of a claim depends on the selected set. The linearization in Appendix A lets ILP represent these interactions directly. The evaluated greedy variant is an approximation/fallback, and the solver runs under a time limit. The mean solve time in the original setting is 0.003 s (Appendix J), so the time limit is not the binding constraint in the reported experiments.

Response to Reviewer — continued

Q1. Prevalence of the assumed structural pattern

We thank the reviewer for this helpful question about how often the assumed structural pattern occurs in practice. To clarify, the three conditions in Section 2 define the target attack family: individually plausible passages, internal support among misleading claims, and external conflict with benign evidence. Our benchmark is a controlled evaluation of that family, not a prevalence study. What we can measure is how much of the benchmark the structural signals capture. Table 2 reports attack recall of 78.31% and benign retention of 79.84%. The ablation without coordinated-core detection retains 76.4% of benign evidence while dropping attack recall to 60.8% (Table 3). Accordingly, the defense does not rely on the pattern being universal, even though its strongest setting uses it.

The threat is motivated by PoisonedRAG (Zou et al., 2025). That work shows that injecting five malicious texts per target question into a knowledge database with millions of texts can reach about 90% attack success. A deployment-level prevalence estimate is beyond this paper and is added as future work.

Q2. Adaptive attacks and future improvements

We thank the reviewer for the helpful suggestion to analyze adaptive attacks and identify future improvements. To address this point, Appendix M evaluates attackers that exploit the structural assumptions. The defense advantage falls from +34.0+34.0 percentage points in the static setting to +2.0+2.0–20.020.0 points under the evaluated adaptive variants. The strongest variants (A2-2 and A4) reach 72.0% defended ASR under that appendix's protocol. A0 uses A4's instruction without oracle feedback and already cuts the advantage to 16.0pp. This result indicates that the bypass comes mainly from knowing the mechanism rather than from iterative optimization. These numbers use the appendix's ASR definition (1−EM1-\mathrm{EM} on samples where a malicious document entered the context).

Provenance signals that do not depend on the graph signature, multiple graph views rather than one fixed pattern, and a stronger separation between factual evidence and embedded instructions at generation time are natural directions for improvement. The existing adaptive results remain in Appendix M.

Comments on the paper

1 comment

We sincerely thank all reviewers for their careful reading and thoughtful, constructive suggestions. In our rebuttal, we have provided additional component and alternative comparisons, three independent end-to-end runs, and parameter sensitivity analyses. We have also clarified the benchmark construction and coordination verification, distinguished the evaluation settings and populations, explained key methodological choices. In addition, we have acknowledged the limitations concerning temporal conflicts, adaptive attacks, and retrieval dependence, and outlined corresponding future directions. We hope these additional results and clarifications address your concerns, and would greatly appreciate your feedback on any points that remain unresolved. If anything is still unclear or requires further evidence or discussion, we would be very happy to provide additional clarification and address any remaining concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.