acceptodds
Under review as a conference paper at ICLR 2027

Beyond Centrality: Position-Decoupled Verification for Finite-Horizon Attacks in LLM Multi-Agent Systems

Abstract

Large language model-based multi-agent systems (LLM-MAS) are vulnerable to misinformation that propagates through inter-agent dependencies. Topology-aware defenses often prioritize structurally central sources, implicitly assuming that lower-centrality sources pose lower security risk. We challenge this assumption with matched Hub/Nonhub attacks that differ only in source position, and find that nonhub sources can pose risks comparable to hubs. We further show formally that dominant asymptotic Perron source rankings need not order finite-horizon task-relevant adversarial influence; when such uncalibrated rankings drive selective verification, residual risk can shift toward structurally deprioritized sources. To address this risk–protection mismatch, we propose CORA-GG, a position-decoupled assertion-governance framework that separates verification allocation from source centrality, prioritizes risky assertions within each source, tracks semantic lineages across downstream reuse, and gates factual influence on verifier-supported evidence. This prevents unverified or repeatedly inherited assertions from gaining factual authority through propagation. Across Quant, Rigid, and MMLU on MetaGPT, LangGraph, and AutoGen, CORA-GG records no observed targeted attack successes and achieves lower worst-source residual risk than evaluated baselines under matched token, false-positive-rate, and clean-utility constraints.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Multi-Agent Systems, Multi-Agent Security, Finite-Horizon Adversarial Influence, Position-Decoupled Verification, Assertion-Level Governance

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_xxoJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies whether static source centrality reliably predicts misinformation risk in finite-horizon LLM multi-agent systems. It proposes CORA-GG, which combines position-independent verification quotas, assertion-level risk prioritization, semantic lineage tracking, and verification-gated downstream influence.

Critical strengths
  1. The paper addresses a meaningful problem: structurally peripheral agents can introduce consequential misinformation and may receive insufficient verification under centrality-based allocation.
  2. The controlled interventions on verification allocation are useful for examining how allocation bias affects residual risk. The additional stress tests also explore relevant limitations, including evidence poisoning and constrained verification budgets.
Critical weaknesses
  1. The reported ASR and accuracy values ​​appear inconsistent with the stated evaluation definition. Appendix C.1.2 defines a targeted attack success as a terminal task failure attributable to the injected misinformation. Under the same evaluation population and denominator, targeted ASR therefore cannot exceed the task error rate. However, Table 20 reports 93.3% ASR alongside 86.7% attacked-task accuracy for undefended MMLU under Hub attacks, and 96.7% ASR alongside 93.3% accuracy under Nonhub attacks. Table 1 similarly reports 82.2% mixed ASR and 32.8% attacked-task accuracy for the undefended system. These discrepancies require clarification of the denominators, evaluation populations, or metric implementation. As presented, they undermine confidence in the quantitative conclusions.

  2. The model coverage is limited. The main comparisons use Qwen2.5-7B-Instruct for both agents and verification. Although Table 24 includes Llama-3.1-8B and a Mistral-7B verifier, these are limited transfer checks for CORA-GG without corresponding baseline comparisons. The evaluated models also remain within the 7–8B scale. Consequently, the experiments do not establish whether the relative advantages persist across substantially different agent and verifier capabilities.

  3. The evaluation lacks direct comparison with closely related misinformation-specific defenses. In particular, ARGUS (Goal-aware identification and rectification of misinformation in multi-agent systems, ICLR 2026) is neither cited nor evaluated. ARGUS directly addresses misinformation localization and correction, combining topology with semantic relevance and dynamic feedback. Comparing against a static Hub-preferential policy does not adequately represent this alternative. A direct comparison under matched attack and resource settings would help establish represent CORA-GG’s advantage over existing misinformation defenses.

  4. The experimental presentation relies too heavily on tables and provides limited visibility into the underlying mechanisms. Although Figure 1 illustrates the architecture, the experimental findings are presented almost entirely as tabulated aggregates.


If authors could address these concerns, I will consider raising my grade.

Questions for authors

See weaknesses

5 comments
AuthorsOpenReview ↗

Response to Reviewer xxoJ: Corrected Metrics, Cross-Backbone Comparisons, Resource-Matched ARGUS Evaluation, and Assertion-Genealogy Visualization

Dear Reviewer xxol,

Thank you for your review. We corrected the reporting error and added cross-backbone, ARGUS, and genealogy evidence.

1. ASR–Accuracy Consistency

We sincerely apologize for the accuracy-reporting error and concerns about evaluation reliability. Its cause was an incorrect field reference: accuracy used a pre-attack prediction, not the attacked terminal answer. ASR already used attacked terminal outcomes and semantic attribution, so ASR remains unchanged. We recalculated accuracy from terminal answers using task-specific criteria.

Appendix C.1.2 (app:targeted_attribution) defines correctness CeC_e, attribution AeA_e, and Se=(1−Ce)AeS_e=(1-C_e)A_e. Both metrics use identical attacked outputs and denominators: ASR=∑Se/∣E∣\mathrm{ASR}=\sum S_e/|\mathcal E|, AttackAcc=∑Ce/∣E∣\mathrm{AttackAcc}=\sum C_e/|\mathcal E|, so ASR + AttackAcc ≤ 1. All runs count. Appendix C.1.3–C.1.4 (app:run_accounting, app:statistics): 3,000 runs per task-family/position (200 tasks × 3 frameworks × 5 repeats), 9,000/position, 18,000/method; mixed ASR is a 1:1 Hub/Nonhub aggregate, not another sample; 600 task IDs are inference clusters.

Appendix F.1.2 (tab:task_subgroups) corrects undefended MMLU Hub ASR/accuracy: 93.3%/6.0%; Nonhub: 96.7%/2.9% (sums 99.3%/99.6%). Section 4.2 (tab:main_comparison) corrects pooled undefended accuracy to 4.25% (82.2% mixed ASR), Reflection to 20.6%; Appendix D (tab:minimal_system) corrects Minimal Hub-preferential/Uniform to 13.2%/28.5%. Pooled accuracy: (10.0+3.3+3.3+0.0+6.0+2.9)/6=4.25%. Zero observed ASR is not zero risk: the one-sided exact 95% bound is 0.50% on task-level any-failure probability over the 600-task grid, not per-run risk (Section 4.2; Appendix F.2.2, tab:zero_bounds).

2. Cross-Backbone Baseline Comparison

Standalone transfer tests were insufficient. Sections 4.1/4.5 and Appendix G.3 (tab:paired_backbone) compare None, Balanced, G-Safeguard, ARGUS, CORA-GG on Qwen2.5-3B-Instruct, Qwen2.5-7B-Instruct, DeepSeek-R1-Distill-Qwen-1.5B, and DeepSeek-R1-Distill-Llama-8B agent backbones. Each of the three newly run backbones uses matched tasks, attacks and seeds across methods (200 tasks × 2 positions × 5 repeats = 2,000 runs/method; 30,000 new runs total). Qwen2.5-7B reuses Exp-M aggregates (600 IDs; 18,000 runs/method), not an independent sample.

ARGUS→CORA-GG mixed ASR: Qwen-3B 7.9%→0.8%; Qwen-7B 5.4%→0.0%; DeepSeek-Distill-Qwen-1.5B 9.3%→1.5%; DeepSeek-Distill-Llama-8B 6.1%→0.6%. CORA-GG worst-source ASRs: 1.1%, 0.0%, 2.1%, 0.9%, respectively—lowest among the five methods throughout.

3. Direct and Resource-Matched ARGUS

ARGUS adaptively localizes propagation channels and rectifies goals; it is not static Hub prioritization (Sections 2, 4.2; Appendix E.1). Exp-M ARGUS Hub/Nonhub/mixed/worst ASR: 2.2%/8.6%/5.4%/8.6%, versus CORA-GG 0.0% observed. CORA-GG vs. ARGUS: attacked accuracy 41.3%/40.1%, FPR 2.3%/2.6%, tokens 29.8k/46.5k, latency 171.6s/270.4s.

Appendix E.1 (tab:constrained_summary, tab:constrained_asr, tab:constrained_aux) validation-tunes/fixes ARGUS operating points on a held-out slice (200 clusters; 2,000 runs/method/constraint). ARGUS mixed/worst ASR: iso-token 10.0%/14.5% (29.9k vs. 29.8k tokens); iso-FPR 5.9%/9.2% (2.3%); iso-clean-utility 4.9%/7.8% (72.8%). CORA-GG remains 0.0% observed. ARGUS leads baseline mixed ASR under all three; G-Safeguard beats ARGUS on iso-token worst (13.5% vs. 14.5%).

4. Log-Verified Assertion Genealogy

Section 4.4/Appendix D.3 (app:assertion_lineage_case, fig:assertion_lineage_trace) add a log-verified observed Nonhub trace. Retriever injects false 48×25=1400, inherited by Solver as L-A0. Balanced adopts it without a timely verdict; CORA-GG returns CONTRADICTED, makes L-A0 RED with zero factual weight, and aggregates independently supported 48×25=1200 through GREEN L-E1. Assertion identities, lineage edges, verdicts, and terminal outcomes were checked against paired execution logs.

Controlled dose–response (Section 4.4, tab:gamma_dose): γ=−2: 4.5%/0.1%; γ=+2: 0.0%/2.2%; γ=0: 0.0%/0.0%, Hub/Nonhub ASR, with equal verification shares at zero.

Conclusion

These changes improve metric consistency, cross-backbone and resource-matched comparisons, and log-grounded interpretation. We appreciate your reconsideration.

Thanks for the response and the new experiments. The revision addresses several of my concerns. I have a few follow-up questions about the reported results and the main insight.

  1. The original metric mismatch is fixed, but a new one appears in Table 17. The corrected ASR and accuracy values resolve the contradiction I raised. However, page 27 contains another mismatch. The text says ARGUS has 66.4% clean accuracy and leaves its SCR blank because the supplied 68.6% is inconsistent. Table 17 instead reports 68.4% clean accuracy and 64.2% SCR. The paragraph below again uses 66.4%. Which values are correct? Is the text outdated, or does the table need correction? Please clarify where the final values come from.

  2. The ARGUS comparison helps, but what does it tell us about position-decoupled allocation? The new ARGUS results largely address my missing-baseline concern. I also appreciate the clearer discussion of its adaptive localization and goal-aware rectification. My remaining question is about the main insight. ARGUS already uses semantic information and dynamic feedback, so it does not simply assume that central sources deserve more protection. The full-system comparison shows that CORA-GG performs better in the reported setting. But CORA-GG also changes verification, lineage handling, and downstream influence. What evidence connects this advantage over ARGUS specifically to position-decoupled allocation? The routing controls support an advantage over the tested allocation rules. How far do the authors think that conclusion extends to adaptive, content-aware allocation?

  3. The link between the allocation theory and the implemented gate is still unclear to me. The revision clarifies that Eq. (3.5) is a simplified allocation-level model. It also states that unverified assertions have zero direct factual weight. Under this rule, checking fewer assertions does not by itself explain why more false information reaches the final answer. Appendix G.6 lists possible explanations, but does not identify which occurred in the reported failures. What do the existing failed runs actually show? Did the verifier approve a false assertion, did the lineage tracking fail, or did the attack act through another stage? This matters for the paper’s main explanation. Does the allocation model capture the failure mechanism of CORA-GG, or does it mainly motivate the allocation rule?

The new comparisons and visualization are helpful. My remaining concerns are the unresolved reporting mismatch and the connection between the allocation insight and the observed gains. Clarifying these points would help me update my assessment.

Dear Reviewer,

Thank you for the follow-up. We answer each point directly.

1. Which ARGUS clean-task values are correct?

Table 17 is correct: ARGUS Clean Accuracy is 68.4% and SCR is 64.2%. The original default-configuration clean-execution logs confirm these values; 66.4% Clean Accuracy, 68.6% SCR, and the blank-SCR explanation came from outdated prose. We apologize and have corrected Appendix E (tab:clean_utility). Both metrics use the same clean-task executions and denominator: Clean Acc = correct / N_clean; SCR = (safe AND correct) / N_clean. The corrected CORA-GG vs. ARGUS comparison is 72.8% vs. 68.4% Clean Acc and 70.9% vs. 64.2% SCR.

2. Does the ARGUS advantage specifically come from position decoupling?

The ARGUS comparison alone cannot establish that. ARGUS uses adaptive localization and goal-aware rectification, while CORA-GG also changes verification, lineage handling, and gating. Hence ARGUS's 5.4% mixed/8.6% worst-source ASR versus CORA-GG's 0.0% observed ASR (tab:main_comparison) is a whole-system comparison, not an allocation-only causal estimate.

The allocation-specific evidence is the controlled routing intervention, with other CORA-GG components fixed (Section 4.4; Appendix D): Hub-preferential, Uniform, global Risk-only, and Position-Decoupled routing yield 6.7%, 6.0%, 4.8%, and 0.0% mixed ASR, respectively (tab:routing_control). Risk-only is content-aware but lacks per-source base quotas. Changing only the centrality preference shifts Hub/Nonhub ASR from 4.5%/0.1% (γ=−2) to 0.0%/2.2% (γ=+2), versus 0.0%/0.0% (γ=0) (tab:gamma_dose). Risk shifts to the deprioritized source. Without genealogy or gating, the minimal router-plus-verifier control reduces mixed ASR 81.9%→17.4% (tab:minimal_system); routing × hard-gating controls show 6.7%→0.0% with gating on and 13.4%→0.6% with gating off (tab:routing_gate_factorial).

Scope: These interventions isolate a routing benefit among tested policies, not the full ARGUS gap or universal superiority. A calibrated adaptive allocator might perform as well or better. Our claim: without calibrated source hazards, static position should not set base verification quotas; content risk still ranks assertions within quotas. Theorem 2 applies only to this uncalibrated routing-level surrogate.

3. What actually caused the failures, and what does Eq. (3.5) explain?

Fewer checks do not directly let YELLOW/RED assertions through the GREEN-only gate. Appendix G.6 now reports the audit of the 60 successful executions under the 180-call, 8× assertion-volume condition (tab:capacity_failure_attribution): 25 Hub + 35 Nonhub successes out of 1,000 executions each, or 3.0% mixed ASR.

56/60 (93.3%) were primarily capacity-related cross-stage failures: verifier-budget saturation delayed checks of attack-relevant assertions, and attack-affected intermediate workflow states influenced later processing and terminal decisions. This is indirect cross-stage influence, not direct YELLOW/RED admission at the fixed-stage gate. Not every subsequent step was localized. 4/60 (6.7%) were false verifier approvals: false assertions received SUPPORTED verdicts and became GREEN, making them eligible for factual aggregation. These are verifier errors, not promotion-rule bypasses. No separate genealogy error was identified as a primary failure category; secondary contributors are not excluded.

The capacity intervention supports this diagnosis: at 180 calls, increasing volume 1×→8× decreased verified share 30.1%→4.2% and increased mixed ASR 0.0%→3.0%; at 360 calls/8×, mixed ASR fell to 0.5% (tab:global_budget_full). These trends are not per-run counterfactual proofs. Proposition A5 guarantees only stage-local zero direct factual weight for YELLOW/RED, not cross-stage non-influence or verifier correctness.

Eq. (3.5), q_i^D=q_i^0(1−ν_iη), mainly motivates the allocation rule. It abstracts coverage-dependent residual hazard; Theory motivates allocation; routing controls isolate its tested effect; Appendix G.6 identifies observed failure pathways.

Thank you.

Reviewer_tnA8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper questions the common practice in LLM-MAS defense of allocating verification resources according to static source centrality. The authors first use matched Hub/Nonhub attacks (differing only in source position) to show that a non-hub source can pose risk comparable to a hub source (undefended ASR 83.3% vs 81.1%), and formally prove Theorem 1: the asymptotic Perron source ordering need not coincide with the finite-horizon task-relevant adversarial influence ordering. They then introduce coverage skew and Proposition 1 (asymmetric coverage can reverse the post-defense source-risk ordering). The method, CORA-GG, decouples each source's baseline verification fraction from centrality, ranks within each source by risk, tracks assertion semantic lineages via a genealogy graph, and allows only evidence-verified lineages to carry factual weight in downstream aggregation. On MetaGPT/LangGraph/AutoGen × Quant/Rigid/MMLU, CORA-GG records 0% targeted ASR (one-sided 95% upper bound 0.50%), and outperforms the validation-tuned strongest baselines under token/FPR/utility-matched comparisons; the centrality dose–response and routing-control experiments show that changing only the allocation preference suffices to reverse the residual-risk direction.

Critical strengths
  1. A carefully controlled mechanism demonstration rather than a correlation. The matched Hub/Nonhub design plus the centrality dose–response experiment is a genuinely good controlled manipulation: changing only the allocation preference changes which source is under-covered and where residual risk remains. As a piece of causal evidence about allocation bias, it is more persuasive than the cross-setting comparisons typical of this literature.
  2. A coherent, deployable governance design. Decoupling baseline verification entitlement from centrality, ranking within-source by risk, and gating downstream influence on verified lineages is a clean and internally consistent architecture, and the authors evaluate it under adaptive-attack stress tests rather than only on the primary grid.
Critical weaknesses
  1. The headline "0% targeted ASR" is statistically and methodologically fragile. Three separate issues compound here. (i) ρ=0.30 is selected on validation as the smallest ratio with zero validation ASR—an explicit search for a zero, which biases the subsequent fresh-set 0% result optimistic even though the fresh set is disjoint. (ii) The reported 0% is a finite-sample outcome whose honest characterization is a one-sided 95% upper bound of 0.50%, yet the abstract and tables present "0.0" prominently, inviting readers to read it as a demonstrated impossibility rather than a non-detection. (iii) The strongest security claim is thus carried by a self-selected operating point and a non-detection bound, while the operating points where the method actually fails (verifier saturation, retrieval poisoning, 8× assertion volume) are relegated to a robustness appendix.
  2. The defense's security rests on trusted components it does not protect, and its failure modes are the realistic attack surface. CORA-GG's guarantee presumes an uncompromised verifier and trusted evidence; the paper's own stress tests show the method degrading precisely where a real adaptive adversary would push (retrieval poisoning 3.5%, verifier-capacity saturation 3.0%, larger collusion up to 1.5% source-specific ASR). Because the contribution is fundamentally a verification-allocation scheme, it inherits and concentrates the single-verifier trust assumption rather than reducing it—so the system moves the vulnerability rather than eliminating it, and the paper does not analyze the compound risk of an adversary that jointly targets allocation and the verifier.
Questions for authors
  1. If ρ is chosen by nested cross-validation on validation and ASR then reported on a fully independent hold-out, does the 0% result still hold? Please report the distribution of selected ρ and the corresponding fresh-set ASR, not just the single selected point.
  2. Given that the 0% figure is a non-detection with a 0.50% one-sided upper bound, will you revise the abstract/tables to report the bound rather than "0.0"? If not, what justifies presenting a finite-sample zero as a headline result?
  3. The centrality critique rests on two matched positions and a 45-position association with R²≤0.032 for centrality scores. Can you provide a positive characterization—conditions (task family, horizon, topology) under which static centrality does order finite-horizon risk—so the applicability boundary of Theorem 1 is explicit rather than implied to be general?
  4. If the verifier itself is compromised or its trusted evidence is poisoned (retrieval poisoning already yields 3.5%), what is the end-to-end guarantee? Does CORA-GG retain any advantage over baselines when the adversary jointly attacks the allocation rule and the verifier?
2 comments
AuthorsOpenReview ↗

Response to Statistical Reliability, Verification-Ratio Selection, and Centrality Applicability

Dear Reviewer tnA8,

Thank you for your review. We clarify validation selection, uncertainty, centrality, and verifier trust below.

1. Verification-Ratio Selection (Q1)

We agree that choosing the smallest ratio with zero validation ASR induces selection bias; validation zero is not an unbiased deployment-risk estimate. Appendix C.2 (app:rho_selection) reports a prespecified 15-ratio sweep over 45 disjoint validation task IDs, paired Hub/Nonhub attacks, and five repeats: 45×2×15×5=6,750 validation executions. The first zero-observed-ASR ratio, ρ=0.30, was selected and frozen before testing.

A five-fold stability check on 60 validation task IDs selected (0.30, 0.30, 0.30, 0.30, 0.30). This is selection stability, not nested cross-validation with independent outer-test folds; we claim no unmeasured fold-specific test ASRs.

With frozen ρ, two disjoint tests yielded zero targeted successes: primary held-out, 600 task IDs/18,000 CORA-GG executions (Section 4.2, Table 1); independent replication, 600 additional IDs/6,000 CORA-GG executions (Appendix F.2.1, tab:ind_replication). Neither test set selected ρ. These results support the frozen setting, not zero risk.

2. Meaning of Zero Observed ASR (Q2)

Finite-sample zero does not mean attack impossibility. The revised Abstract specifies no successes observed in the finite held-out sample and tested settings. Section 4.2/Table 1 and Appendix C.1.4 (app:statistics) distinguish the observed rate from uncertainty. With targeted-success indicator Y for task u, framework f, position a, and repeat r, define Z_{u,a}=max_{f,r}Y_{u,f,a,r} and Z_u=max_a Z_{u,a}; Z_u=1 denotes any targeted failure anywhere in the prespecified grid.

The primary 18,000 runs reuse 600 base-task clusters, not 18,000 independent trials. For zero observed cluster failures, the one-sided exact 95% upper bound is 1−0.05^(1/600)=0.498%≈0.50%. Thus 0.0% is observed execution-level ASR, while 0.50% bounds base-task any-failure probability under the specified protocol, not arbitrary deployment or per-execution risk. This distinction appears in Section 4.2, the Table 1 caption, and Appendix F.2.2 (tab:zero_bounds).

Section 4.5 reports observed limits: 3.0% mixed ASR under verifier-capacity saturation, 3.5% under retrieval poisoning, and other adaptive-attack degradation; we make no universal zero-risk claim.

3. When Is Structural Centrality Informative? (Q3)

Theorem 1 establishes non-sufficiency, not universal failure, of asymptotic centrality. Section 3.2 and Appendix A.2 (app:thm1) give a sufficient condition. If v_M^T M=λ₁v_M^T, task readout r^T=c v_M^T (c>0), and all sources have the same nonnegative injection schedule u(t), then Φ_i(T)=c v_{M,i} Σ_{t=0}^{T−1}λ₁^(T−1−t)u(t). When this common factor is positive, finite-horizon influence follows M's Perron source ordering. For dependency centrality from A, M must additionally preserve A's source ordering; M=τA+dI, τ>0, d≥0, suffices. These conditions are not universal LLM-MAS assumptions.

Section 4.3/Appendix D.1.1 (tab:topology_control) show topology-dependent Hub/Nonhub ASR: chain 78.3%/65.0%; star 85.0%/80.0%; mesh 76.7%/88.3%. Hub dominates in the tested chain, Nonhub in the mesh; task-family orderings also differ (Appendix F.1.2, tab:task_subgroups). Hence centrality is not useless, but uncalibrated centrality cannot universally justify unequal finite-horizon verification allocation.

4. Verifier Compromise, Poisoning, and Joint Attacks (Q4)

Position-decoupled allocation cannot itself guarantee verifier correctness or evidence integrity. Section 3.3 and Appendix B.2 (Proposition A5) establish only a stage-local GREEN-only factual-influence invariant: verifier-supported assertions alone have direct factual weight at a fixed aggregation stage; a corrupted verifier or evidence source may still produce false GREEN assertions.

Appendix G.7 (tab:adaptive) tests retrieval poisoning: CORA-GG 3.5% ASR, Balanced 16.7%, G-Safeguard 14.4%. CORA-GG performs better in this tested setting, but its nonzero ASR exposes reliance on trusted evidence. Appendix G.4–G.5 further report 3.0% mixed ASR under a 180-call budget at 8× assertion volume, and up to 1.5% source-specific ASR under larger tested collusion.

We have not established robustness to jointly optimized verification-allocation pressure and verifier compromise. Separate poisoning, capacity, and collusion tests cannot imply a joint-attack guarantee; this limitation is explicit in Appendix H.2 (Practical Limitations).

Summary

These revisions clarify selection versus testing, observed zero versus cluster-level uncertainty, conditions for informative centrality, and verifier/evidence trust limits. We appreciate the reviewer's concerns.

I thank the authors for the clarifications on ρ, the 0% figure, Theorem 1, and verifier trust.

Q2. Distinguishing the observed 0.0% execution-level ASR from the 0.50% cluster-level one-sided bound, and stating in the abstract that this is a finite-sample non-detection, is the minimum correction. I agree with that wording. It does not make “0% targeted ASR” a demonstrated security guarantee.

Q1. I asked for nested cross-validation: the distribution of selected ρ and the corresponding fresh-set ASR on independent outer folds. The reply reports a 15-ratio validation sweep and a five-fold stability check in which ρ=0.30 is always selected, then two frozen-ρ tests with zero observed successes. The authors explicitly “claim no unmeasured fold-specific test ASRs.” That is not the nested-CV number I asked for. Selection stability of ρ is not the same as an unbiased test ASR after selection.

Q3. The sufficient condition in Section 3.2 / Appendix A.2, plus the chain/star/mesh Hub–Nonhub splits (78.3/65.0, 85.0/80.0, 76.7/88.3), is a useful positive characterization. I accept that Theorem 1 is non-sufficiency, not universal failure, of centrality.

Q4 / W2. The authors agree that allocation cannot protect a compromised verifier, that retrieval poisoning still yields 3.5% ASR, and that they “have not established robustness to jointly optimized verification-allocation pressure and verifier compromise.” That was the core of W2: the scheme moves the trust assumption onto the verifier and trusted evidence rather than reducing it. Naming the limitation in Appendix H.2 does not close it. A better number than Balanced/G-Safeguard under a single poisoning test is not an end-to-end guarantee under a joint attack.

W1 therefore still stands. The headline security claim is a self-selected operating point plus a non-detection bound; the realistic failure modes remain in the stress tests; the joint attack is untested. I keep my rating at 2 (weak rejection).

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

The paper studies whether structurally central agents should receive more verification resources in LLM-based multi-agent systems. Using attacks that differ only in source position and a finite-horizon propagation analysis, it argues that lower-centrality sources can be similarly consequential. The proposed CORA-GG assigns a common base verification fraction across sources, prioritizes assertions within each source, tracks their semantic lineages, and permits only verifier-supported assertions to contribute factual information. Experiments on three task families and three agent frameworks report reduced worst-source attack success, together with allocation controls and robustness analyses.

Critical strengths
  • The paper addresses a useful question in resource-constrained defense: whether structural importance is a reliable basis for allocating verification. The controlled source-position experiments and centrality-weight interventions are appropriate ways to distinguish structural rank from the consequences of an attack within a limited number of rounds.

  • The evaluation design includes separate comparisons that control for token usage, false-positive rate, or clean-task utility, a disjoint replication set, and stress tests for collusion, assertion flooding, and evidence poisoning. These controls are relevant to assessing the practical value of a defense beyond its performance at one operating point.

Critical weaknesses
  • The reported metrics are inconsistent with the definition of targeted attack success. Appendix C.1.2 requires both an incorrect terminal result and attribution to the injected assertion. However, Table 20 reports 93.3% Hub ASR and 86.7% Hub accuracy for undefended MMLU. These events cannot have these rates on the same executions if attack success is a subset of task failure. The discrepancy directly affects confidence in the evaluation and requires corrected results or an explanation of different denominators.

  • The connection between the allocation theory and the implemented defense is unclear. Eq. (3.5) leaves unverified assertions with their baseline hazard, whereas Stage III excludes unverified factual content. Table 26 attributes increased ASR to reduced verification coverage, but does not explain whether failures arise from false verifier approvals, changes to later reasoning, or another pathway. Without this distinction, it is difficult to assess how the minimax allocation result supports the actual pipeline.

Questions for authors
  • Table 20 reports 86.7% accuracy for undefended MMLU under Hub attacks, leaving only 13.3% incorrect answers. Since Appendix C.1.2 requires an incorrect final answer for attack success, why is the reported ASR 93.3%? Are ASR and accuracy computed on the same evaluated runs and with the same denominator?
  • In the routing and verifier-capacity experiments, how do successful attacks reach the final decision despite verified-only factual influence? Which observed failure mechanism is modeled by Eq. (3.5)?
3 comments
AuthorsOpenReview ↗

Response to Evaluation Consistency and the Scope of Verification-Gated Security

Dear Reviewer skhE,

Thank you for your review. We corrected the accuracy error and clarified the theory and experiments.

1. ASR–Accuracy: Cause, Corrections, Denominators

We sincerely apologize for this reporting error and its impact on evaluation credibility. Appendix C.1.2 (app:targeted_attribution) defines terminal correctness CeC_e, attribution AeA_e, and targeted success Se=(1−Ce)AeS_e=(1-C_e)A_e. On the same executions E\mathcal E: ASR=∑Se/∣E∣ASR=\sum S_e/|\mathcal E|, AttackAcc=∑Ce/∣E∣AttackAcc=\sum C_e/|\mathcal E|; thus ASR + AttackAcc ≤ 1.

Cause: accuracy aggregation referenced a pre-attack prediction field instead of the attacked terminal answer. ASR used attacked terminal outputs and semantic attribution, so ASR is unchanged. Accuracy was recomputed from final answers using task-specific criteria—not different denominators or success definitions.

Appendix F.1.2 (tab:task_subgroups; former Table 20) corrects undefended MMLU Hub ASR/accuracy: 93.3%/6.0%; Nonhub: 96.7%/2.9% (sums 99.3%/99.6%). Section 4.2 (tab:main_comparison) corrects undefended accuracy to 4.25% (82.2% mixed ASR) and Reflection to 20.6%. Appendix D.1.4 (tab:minimal_system) corrects Minimal Hub-preferential/Uniform to 13.2%/28.5%; Appendix E (tab:clean_utility) is aligned. Pooled accuracy equals (10.0+3.3+3.3+0.0+6.0+2.9)/6 = 4.25%.

Appendix C.1.3–C.1.4 (app:run_accounting, app:statistics): 3,000 runs per task-family/position (200 tasks × 3 frameworks × 5 repeats), 9,000/position, 18,000/method. Both metrics use all attacked executions, not only errors, contamination, or successes. Mixed ASR is a 1:1 Hub/Nonhub aggregate, not an independent sample; 600 unique base-task IDs are inference clusters.

2. Why Does GREEN-Only Aggregation Permit Residual ASR?

Section 3.3 requires verifier SUPPORTED for GREEN; YELLOW/RED have zero direct factual weight, and paraphrases add no independent evidence. Appendix B.2 (Proposition A5) proves stage-local non-influence with fixed GREEN assertions, genealogy, ownership, and weights—not end-to-end multi-round security.

Section 3.3 and Appendix G.6 (app:failure_path) list possible, unmeasured pathways: (i) erroneous SUPPORTED promotions; (ii) faulty genealogy/evidence attribution granting false trust; (iii) indirect influence on subsequent retrieval, candidate generation, genealogy, or verification. Nonzero ASR does not establish bypass of the GREEN-only gate.

Minimal-System Isolation (Appendix D.1.4, tab:minimal_system) removes genealogy, hard gating, and verification-gated aggregation; no gate exists to bypass. Routing-Only (Appendix D, tab:routing_control) retains governance, changing allocation only. Hub-preferential routing yields 91.4%/28.7% Hub/Nonhub effective coverage and 3.3%/10.0% ASR (tab:structural_bias). Centrality reversal shifts under-covered source risk (Section 4.4, tab:gamma_dose); routing × gating: tab:routing_gate_factorial. These show allocation effects, not per-run causes.

3. What Does Eq. (3.5) Model?

Section 3.2 and Appendix A.1 (Assumption A3) define qiD=qi0(1−νiη)q_i^D=q_i^0(1-\nu_i\eta): baseline hazard qi0q_i^0, effective pre-terminal coverage νi\nu_i, conditional interception efficacy η\eta. This phenomenological hazard model neither admits unverified Stage III content nor predicts exact end-to-end ASR. It omits separate terms for false approvals, genealogy errors, and cross-stage propagation. With reliable verification, enforced gating, and no indirect paths, low coverage need not cause success.

Routed share fif_i differs from coverage νi\nu_i, requiring timely verdicts. Section 3.2 and Appendix A.4 (app:minimax_proof) limit Theorem 2 to minimax equal base routing shares for the specified worst-source surrogate under uncalibrated hazards—not end-to-end ASR optimality or equal coverage under latency/call caps.

4. What Does the Capacity Experiment Show?

Appendix G.5 (tab:global_budget_full; former Table 26): with 180 verifier calls, assertion volume 1×→8× reduces verified share 30.1%→4.2%, while mixed targeted ASR rises 0.0%→3.0%. At 8×, 360 calls lower ASR to 0.5%. At 180 calls/8×: 25 Hub successes/1,000 (2.5%), 35 Nonhub/1,000 (3.5%): 60/2,000 = 3.0% mixed. The table's 30 normalized mixed events/1,000 are not another sample.

Lower coverage increases observed risk; more calls improve containment. This does not show YELLOW/RED bypass. Appendix G.6 (app:failure_path) cannot assign successes to false approvals, genealogy errors, or indirect effects. Eq. (3.5) models coverage-dependent hazard, not an identified implementation-level failure cause.

Conclusion

We corrected the field error and denominators (Section 4.2; Appendices C–F), distinguished routing theory from stage-local GREEN-only governance (Sections 3.2–3.3; Appendices A–B), and bounded routing/capacity evidence (Section 4.4; Appendices D, G). We appreciate the reviewer's feedback.

Thank you for the detailed response. The ASR-accuracy consistency is mostly resolved. Please ensure that the corrected aggregation was re-run for every method, not only for the undefended rows. I have several follow-up questions.

  • The rebuttal states that Eq. (3.5) is a phenomenological hazard model and that Theorem 2 applies only to routing fractions. Although this removes the ambiguity, it also weakens the link between the theory and the proposed pipeline. The core allocation principle (equal per-source verification fractions) is conceptually simple, and Theorem 2 reduces to max–min allocation under a linear budget. The case for CORA-GG therefore rests mainly on the empirical routing interventions.

  • Table 6 suggests that equal allocation alone (Uniform: 6.0% mixed ASR) does not account for the gains, and risk prioritization and gating appear to drive most of the improvement. The contribution attributed to position decoupling specifically is therefore unclear.

  • The rebuttal says Appendix G.6 cannot assign successes to false approvals or indirect effects. However, Table 29 does attribute them: 56 of 60 to "budget saturation, delayed verification, and cross-stage influence," and 4 of 60 to false SUPPORTED verdicts. It does not explain how unresolved content reaches the terminal answer if governed messages contain only GREEN factual content. Specifically: (1) Do downstream agents read the raw messages or only the governed messages? (2) Is the terminal answer itself treated as an assertion and subject to the GREEN-only gate?

AuthorsOpenReview ↗

Clarifying the Independent Contribution of Position Decoupling and the Scope of Verification Gating

Dear Reviewer,

Thank you. Your questions correctly distinguish the allocation principle from the complete defense and, more importantly, the Stage III gate from the full workflow. We have clarified these boundaries in Appendices B, D, and G.6.

1. Aggregation consistency across methods. We agree that the corrected per-execution ASR/accuracy calculation must be applied to every defense, not just undefended rows. Appendix C defines the same execution-level variables and denominator for all methods: AttackAcc = mean(C_e) and ASR = mean((1-C_e)A_e), hence ASR <= 1-AttackAcc. The displayed main-comparison rows satisfy this necessary consistency check.

2. Independent Contribution of Position Decoupling.

We agree that equal allocation alone is insufficient: Uniform achieves 6.0% mixed ASR, versus 0.0% for CORA-GG's position-decoupled routing. The latter combines equal source quotas with within-source risk prioritization; global Risk-only routing achieves 4.8% ASR.

Crucially, our controlled centrality intervention isolates position decoupling itself (tab:gamma_dose). Holding risk scoring, verification, lineage handling, and gating fixed, Hub/Nonhub ASR shifts from 4.5%/0.1% (γ=−2) to 0.0%/2.2% (γ=+2), while γ=0 yields 0.0%/0.0%. Thus, residual risk shifts toward the structurally deprioritized source even when all other components remain unchanged.

This demonstrates an independent benefit of avoiding centrality-induced coverage skew, while risk prioritization and gating provide complementary gains. We do not claim that equal quotas alone explain the full improvement or outperform all calibrated adaptive allocators.

3. Precisely what enters downstream prompts, and is the terminal answer gated? These are two distinct interfaces:

(a) Downstream inputs: The factual message channel uses the governed messages M_tilde_i^(t), containing only GREEN-supported factual assertions. But a downstream agent's full prompt also includes persistent planning/task state, some of which was updated from raw messages before verification. Accordingly, the next agent does not receive only governed content across all context fields. An unverified assertion can have zero factual weight in Stage III yet still affect subsequent processing through such pre-gate state.

(b) Terminal output: The evaluated implementation does not perform a separate assertion-by-assertion GREEN-only verification pass on the final generated answer. GREEN-only is enforced on the governed factual aggregation input, not on every inference or conclusion produced later. We have stated both implementation boundaries explicitly after Algorithm A1 (Appendix B).

(c) What the 60 failed executions show: Under 180 calls and 8x assertion volume, there were 25 Hub and 35 Nonhub targeted successes among 2,000 executions (3.0% mixed ASR). The execution-trace attribution in Appendix G.6 (tab:capacity_failure_attribution) identifies 56/60 (93.3%) instances of the concrete path budget exhausted -> attack-relevant assertion remains YELLOW and is excluded from the governed factual channel -> pre-verification attack-affected state persists in later agents' context -> terminal answer inherits the erroneous state without independent final verification. This is not direct YELLOW/RED passage through the Stage III gate. In the remaining 4/60 (6.7%), false assertions received incorrect SUPPORTED verdicts, became GREEN, and were eligible for governed aggregation. Those are verifier-correctness failures. The categories denote the primary observed path, not the absence of secondary factors.

Proposition A5 guarantees only fixed-stage non-influence of YELLOW/RED through the factual aggregation channel. Eq. (3.5) captures a coverage-dependent allocation-level hazard; it motivates the routing policy but does not itself model pre-gate state reuse, terminal generation, or erroneous verifier support. The appendix now makes this implementation-level causal boundary explicit rather than treating the 56/4 classification as a mechanism explanation by itself.

Thank you for highlighting these distinctions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.