acceptodds
Under review as a conference paper at ICLR 2027

Defending against GEO Poisoning with Hierarchical and Explainable Byzantine-Robust Aggregation

Abstract

Large language models (LLMs) increasingly rely on retrieval systems and external tools to obtain up-to-date information beyond their parametric knowledge. This reliance introduces a new attack surface. Knowledge corruption attacks can inject misleading evidence into retrieval pipelines, while Generative Engine Optimization (GEO) poisoning can make adversarial content more likely to be retrieved and used during reasoning. We propose a hierarchical and explainable Byzantine-robust aggregation framework with dynamic multi-source evidence retrieval. Our method uses reinforcement learning-guided retrieval, evaluates the stance and reliability of individual evidence items, and organizes claims in a dynamically expanded argument tree. It then aggregates node scores bottom-up to limit the influence of unreliable or conflicting evidence. Experiments on four datasets and two backbone LLMs, together with component and retrieval-tool ablations, demonstrate the effectiveness of ToE on standard claim-verification benchmarks and under controlled GEO poisoning.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
GEO Poisoning, Byzantine-Robust Aggregation, Knowledge Corruption

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

This paper proposes Tree of Evidence, or ToE, for automated fact-checking. Instead of asking an LLM to directly judge a claim, ToE searches for evidence from several sources and organises the evidence in a tree. It explicitly searches for both supporting and contradicting evidence. When the evidence is not considered reliable enough, the system breaks the claim into smaller subclaims and searches again.

The system contains three main parts: an RL-based multi-source retrieval agent, a learned evidence evaluator, and a tree aggregation model. The authors also provide a theoretical analysis of the retrieval process.

The method is evaluated on LIAR, PolitiFact, Check-COVID, and an adversarial dataset called AdvFact. It is compared with direct prompting and several existing fact-checking approaches using two LLM backbones. The reported results show that ToE is often better than the compared methods, with the largest improvement appearing on AdvFact with DeepSeek-V3.2. However, it does not outperform the strongest baseline in every dataset and backbone combination.

Critical strengths
  1. The paper addresses an important and timely problem. Fact-checking systems that retrieve information from the web can be misled when fabricated content is deliberately made easy to retrieve. The paper directly studies this problem and proposes a structured defence rather than relying only on the LLM’s internal knowledge.

  2. The explicit search for counter-evidence and the evidence-tree structure are useful ideas. ToE does not search only for evidence supporting the claim. It also deliberately searches for contradicting evidence, which helps reduce one-sided retrieval. The evidence tree also preserves subclaims, evidence, and intermediate scores, making the result easier to inspect than a direct LLM verdict. The paper’s retrieval ablation suggests that removing counter-evidence search substantially reduces performance, although this result is based on a small evaluation set.

Critical weaknesses
  1. The experiments do not justify the full complexity of the proposed pipeline. ToE adds many components at the same time: RL-based source selection, multiple retrieval tools, LLM query generation, document cleaning, source and stance analysis, learned reliability estimation, claim decomposition, neural tree aggregation, pruning, and stopping rules.

The reported ablations mainly remove groups of retrieval sources. They do not separately show the value of the main architectural components, such as the learned retrieval policy, claim decomposition, learned reliability score, neural tree aggregation, or pruning. A simpler baseline using the same LLM, retrieval sources, retrieved-document budget, and evidence-processing steps would help determine whether the tree and RL components provide gains beyond additional retrieval and inference.

  1. The adversarial evaluation is difficult to interpret because all AdvFact examples are false. The paper states that AdvFact contains 200 examples and that all are labelled false. If ordinary accuracy is calculated only on this dataset, a trivial system that always predicts false would score 100%. This makes the reported AdvFact accuracy insufficient by itself to demonstrate robustness to poisoning.

A more convincing evaluation would include both true and false claims, matched clean and poisoned examples, an always-false baseline, attack success rate, and the change in performance from clean to poisoned evidence. This concern is important because AdvFact provides a central part of the evidence supporting the paper’s poisoning-robustness claim.

Questions for authors
  1. Clarification and additional evidence: Can the authors provide separate ablations for the learned retrieval policy, claim decomposition, learned reliability estimation, neural tree aggregation, and pruning? These results would materially affect whether I view the full architecture as justified.

  2. Clarification and additional comparison: How would ToE compare with a simpler pipeline using the same LLM backbone, retrieval sources, number of retrieved documents, and evidence extraction, but using a fixed retrieval strategy and transparent weighted aggregation instead of RL and a tree?

  3. Clarification: How should the AdvFact accuracy be interpreted when all examples are false? What accuracy would an always-false baseline achieve under the reported evaluation? Can the authors report results using both true and false claims, or matched clean and poisoned versions of the same claims?

  4. Clarification: How much inference does ToE use relative to the baselines? At minimum, could the authors report the average number of searches, documents, LLM calls, tokens, and tree nodes per claim? It is needed to separate the effect of the proposed method from the effect of greater inference-time computation.

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

This paper studies fact verification under GEO-style evidence poisoning, where retrieved documents may contain misleading, fabricated, or adversarially optimized content. The authors propose ToE, a hierarchical verification framework that combines reinforcement-learning-based multi-source retrieval, counter-evidence collection, reliability-aware claim decomposition, and bottom-up tree aggregation. The method reports strong results under the evaluated GPE poisoning conditions and mixed but generally competitive performance on LIAR, PolitiFact, and Check-COVID. While the problem is important and the overall system is intuitively appealing, the current submission does not yet provide sufficiently controlled experiments or complete technical details to substantiate several of its main claims, particularly those concerning Byzantine robustness, the contribution of the learned retrieval policy, and the comparability of the GPE results.

Critical strengths

1、The paper addresses an important and timely problem. Retrieval-augmented systems increasingly depend on open and potentially untrusted information sources, and maliciously optimized documents can plausibly manipulate both retrieval and downstream verification. The paper therefore targets a realistic weakness of current LLM-based verification systems. 2、The proposed pipeline is coherent and reasonably well motivated. In particular, actively retrieving counter-evidence, decomposing uncertain claims into smaller subclaims, and aggregating evidence through an explicit argument tree are sensible design choices for handling conflicting evidence. The resulting tree also provides a more inspectable reasoning structure than a single end-to-end LLM judgment. 3、The reported results under the evaluated GPE conditions are strong, with ToE outperforming the listed baselines across the reported poisoning ratios and retrieval settings. The paper also evaluates conventional fact verification with two LLM backbones and reports token-based efficiency measurements rather than discussing accuracy alone.

Critical weaknesses

1、The novelty and robustness claims need sharper justification. ToE mainly integrates existing ideas—including claim decomposition, iterative retrieval, reliability estimation, and tree-based aggregation—into a unified framework. Moreover, the paper calls the method “Byzantine-robust” without formally specifying attacker capabilities, the tolerated fraction of malicious sources, or a corresponding robustness guarantee. The authors should clarify both the technical novelty and the precise meaning of Byzantine robustness. 2、The comparability of the GPE results is unclear, and the baseline coverage is incomplete. The claim-local baseline values in Table 2 appear identical to the DeepSeek-V4-Flash results in the original GPE paper, although this paper states that DeepSeek-V3.2 and GPT-OSS-20B are used. Table 2 should clearly report whether each result was reproduced or imported, as well as the backbone and inference configuration. The evaluation also omits some GPE conditions and recent relevant defenses such as SecureCollaRAG and PRA-RAG. Comparisons under a unified protocol are needed to establish ToE’s position relative to recent methods. 3、The ablations do not sufficiently isolate the main contributions. The learned retrieval policy is not compared with heuristic, random, round-robin, or fixed-source policies, while several reward terms and reliability-related components are not evaluated separately. Moreover, the reported results lack uncertainty estimates, and the tool ablation uses only 30 claims. Consequently, it remains difficult to determine which components are responsible for the observed improvements and whether smaller differences are statistically meaningful. 4、Several implementation and mathematical details are insufficiently specified. In particular, the six-class GPE classifier lacks information about its architecture, training data, loss, and decoding procedure; the reliability target y_r is undefined; and the mutual-information interpretation and bound in Eqs. (2) and (4) require clearer assumptions and derivations. These omissions limit reproducibility and make some theoretical claims difficult to assess.

Questions for authors

Could you formally define the adversary model assumed in your "Byzantine-robust" claim? Specifically, what is the maximum fraction of poisoned/retrieved evidence the system can tolerate while maintaining correctness? Does this guarantee hold under the GEO poisoning settings tested (e.g., 2/3 contamination ratio in Table 2), where accuracy drops significantly (to ~46-50%)?

If no formal robustness bound is derived, would it be more accurate to describe ToE as "resilient to noisy/adversarial retrieval" rather than "Byzantine-robust"? Please clarify if the RL agent’s ability to select diverse sources constitutes a formal Byzantine defense mechanism or a heuristic mitigation.

Were the baseline results in Table 2 reproduced by you using the specified backbones (DeepSeek-V3.2/GPT-OSS-20B) and your implementation, or were they directly imported from the GPE paper? If imported, how do you justify the comparability given potential differences in backbone capabilities and prompt engineering?

Can you provide an ablation comparing the RL-guided retrieval policy against simpler baselines such as: (a) a fixed round-robin source selection, (b) a random source selection, or (c) a heuristic-based policy (e.g., always prioritize fact-checking sites for political claims)? This is crucial to determine if the performance gain comes from the RL optimization itself or merely from the multi-source diversity.

Reviewer_4jVxOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Authors propose a pipeline for fact checking under generative engine optimization attacks. It has an RL agent that collects evidence across multiple sources to verify the claim, an agent that scores the claim based on the collected data, and an agent that extracts a tree from the main claim by breaking it into smaller claims and scoring the sub-claims when the collected data is not conclusive.

The pipeline is implemented on two different LLMs, evaluated in four datasets, and compared to multiple baselines. The improvements are significant.


Update after reading the author response: I improved my score from weak reject to weak accept

Critical strengths
  • The paper is clear and the method is reasonable
  • The results are good
  • The analysis is detailed
Critical weaknesses
  • The pipeline uses a specific algorithm to decompose uncertain claims. What if the attacker uses this exact type of verification to create the fake documents? It appears that experimental setup, which uses generic attackers, is unrealistic
  • Byzantine-robust methods are not explained for the average reader in the paper. Authors have not proposed any argument to justify why their method is byzantine-robust either
  • No statistical confidence tests are reported for the results in Section 4.2
Questions for authors

None

1 comment

Thank you for the response. I increased my review score.

Comments on the paper

1 comment
Area_Chair_UdzAOpenReview ↗

Reminder: Discussion Period Closes in 24 Hours

Dear Reviewers,

This is a gentle reminder that the discussion period will close in 24 hours.

If you have not yet done so, please review the authors’ responses to your comments and participate in the discussion. Please let the authors know whether their responses have addressed your concerns or whether you have any remaining questions.

Thank you again for your time and contribution to the review process.

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.