acceptodds
Withdrawn from ICLR 2027

Can Federated Agents Agree and Still Be Unsafe? FedSafeBelief for Repairing False Safety Consensus

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated agents may agree on a safety decision because they share the same blind spot. Pooling such decisions can preserve an unsafe rule and return it to clients as apparently reliable feedback. At the same time, differences between local policies can make a rule valid for one client and invalid for another. We study this combination as false safety consensus and propose FedSafeBelief, a correction procedure over client scoped safety beliefs. Clients expose abstract decision tuples while retaining their raw interaction traces. The coordinator aligns these tuples, checks policy scope, searches for counterevidence to both disputed and unanimously allowed decisions, and returns a field level correction. Its central restriction is fix not add: correction may revise an existing client belief but may not create a new client key. We characterize when union aggregation dilutes client scoped precision, why shared errors survive voting, and when verified correction contracts the remaining error. The evaluation design separates belief repair, unsafe feedback, response safety, and legitimate policy heterogeneity. Experiments across controlled belief correction, harmful request, and heterogeneous policy settings show that FedSafeBelief consistently reduces unsafe allows and improves client scoped belief precision while preserving legitimate policy differences.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
ML: Deep Generative Models & Autoencoders

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_USW4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies false safety consensus in federated LLM agents. Multiple clients may agree on an unsafe decision because they share similar blind spots. Meanwhile, different clients may follow different local policies, making direct aggregation unreliable. To address these issues, the authors propose FedSafeBelief. It represents client safety knowledge as structured beliefs and keeps each belief tied to its local policy scope. The method aligns beliefs across clients, checks their policy scopes, searches for counterevidence, and verifies potential errors before making corrections. It also follows a fix-not-add rule, which only allows existing client beliefs to be corrected rather than adding new ones. The paper provides theoretical results on rule union, shared errors under consensus, and error reduction through verified correction. Experiments on controlled settings, harmful-request benchmarks, and heterogeneous policy settings show that FedSafeBelief improves belief precision and reduces unsafe decisions compared with several aggregation baselines.

Critical strengths
  1. This paper highlights that agreement among federated agents does not necessarily imply a safe or correct decision.
  2. FedSafeBelief considers both shared unsafe errors and differences in local policies, and the fix-not-add design helps preserve client-specific knowledge.
  3. This paper evaluates belief quality response-level safety, policy heterogeneity, robustness, ablations, and system cost, and shows consistent improvements over the considered baselines.
Critical weaknesses
  1. The method mainly combines scope checking, counterevidence generation and verification, and field-level correction. These components are reasonable, but the overall pipeline appears relatively straightforward.
  2. The method relies heavily on counterevidence generation and verification, but there is no strong baseline that applies the same verification process locally without cross-client information. This makes it difficult to isolate the benefit of the federated design.
  3. The main theorem essentially assumes successful error correction and then shows that the remaining error decreases. It does not establish why the proposed method can achieve such correction in practice, making the theoretical guarantee somewhat circular.
  4. Some results follow directly from their assumptions. For example, the error contraction result assumes a lower-bounded probability of correctly repairing an error. The analysis does not explain how this probability depends on the actual components of FedSafeBelief.
  5. Most baselines are simple aggregation strategies such as rule union, majority voting, confidence weighting, and scope filtering. Comparisons with stronger verification- or critic-based methods would better demonstrate the advantage of FedSafeBelief.
  6. Since verification is a key component of the method, it is unclear how incorrect or unreliable verification affects belief correction and whether such errors may introduce new safety risks.
Questions for authors

see weaknesses.

Reviewer_8KYyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies false safety consensus in federated agents, where multiple agents may agree on an unsafe decision because they share the same blind spot. At the same time, different clients may legitimately follow different local policies, making naive global aggregation potentially harmful. The paper formulates this issue as client-scoped safety-belief correction and proposes FedSafeBelief, which combines belief alignment, policy-scope screening, verified counterevidence, and field-level patching. A key design principle is “fix not add”: the coordinator may correct an existing client-owned belief but may not introduce a new belief key.

Critical strengths
  1. The paper identifies an important and underexplored safety problem.
  2. The client-scoped belief formulation is conceptually appropriate.
Critical weaknesses
  1. FedSafeBelief mainly integrates scope filtering, counterevidence verification, and constrained belief patching. The contribution is stronger in problem formulation and system design than in introducing a fundamentally new learning or aggregation algorithm.
  2. The effectiveness of FedSafeBelief largely relies on obtaining valid counterevidence and a reliable verifier. It remains unclear how much of the improvement comes from the proposed framework itself rather than from the verifier.
  3. Shared errors are manually injected in the controlled experiments. More evaluation on naturally occurring correlated failures among LLM agents would strengthen the practical relevance.
  4. The current baselines are mostly simple aggregation or filtering strategies. Comparisons with stronger alternatives would make the empirical conclusions more convincing.
  5. The current setting assumes cooperative clients and does not consider malicious clients that may manipulate beliefs or evidence, which is an important concern in federated systems.
Questions for authors

Please see Weaknesses.

Reviewer_4siBOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a mechanism to verify and possibly correct “unsafe” decisions in a scenario where multiple LLM agents agree on a wrong decision. The proposed mechanism seems to introduce a centralized verification and correction mechanism for rules stored by those LLM agents. In practice, each agent maintains a set of so-called “safety beliefs” that outline allowed actions under what conditions. The proposed solution introduces two components, a coordinator and a verifier. The coordinator compares beliefs across agents, identifies compatible policies, and selects potentially unsafe decisions for verification. The verifier then checks whether counterexamples invalidate those decisions, allowing the coordinator to correct the corresponding beliefs when necessary.

Critical strengths
  1. The paper addresses an interesting problem related to LLM-agent safety that is the possibility that multiple agents agree on the same unsafe decision.

  2. The proposed framework explicitly considers heterogeneous policies across agents and attempts to preserve client-specific rules while correcting unsafe beliefs.
Critical weaknesses
  1. Unclear contribution to machine learning. I do not see an apparent contribution to machine learning advancements, although the paper can be relevant to the broad topic of LLM agents. The actual innovation seems to be applying counterexample-based verification to a collection of structured agent memories.
  2. Insufficient motivation for federation. I am not convinced about the premises of the paper: why should a federation of agents be necessary or beneficial in the first place? The example of multiple customer-support agents in the Introduction does not seem to justify the need for federation.
  3. Unclear benefit of coordination and privacy concerns. The verifier seems to be effectively the source of correctness, raising questions about what additional benefit cross-client coordination provides over independent local verification.
  4. Possible privacy issue. The coordinator receives substantial semantic information, including conditions, decisions, policy scopes, and evidence summaries, which may be sensitive even without raw interaction histories.
Questions for authors

Q1. I am not convinced about the premises of the paper: why should a federation of agents be necessary or beneficial in the first place? In the Introduction, the authors provide the example of multiple customer-support agents, but it remains unclear why these agents should federate their safety beliefs. Could the authors provide a concrete real-world scenario that justifies this setting?

Q2. A potential contradiction seems to emerge: as far as I can understand, the verifier is effectively the source of correctness. If the verifier can access authoritative policies and reliably interpret them, then the coordinator already has the information needed to correct an individual agent without consulting other agents. Could the authors clarify what additional benefit cross-client coordination provides, and why the same corrections could not be achieved through independent local verification?

Q3. Related to the previous point, could the authors include a baseline where each agent independently verifies and corrects its own beliefs using the same counterexample generator and verifier? This would help establish whether the improvements originate from the proposed federated coordination or simply from the verification mechanism.

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

Related papers

Discussion (0)

Sign in to comment.