acceptodds
Under review as a conference paper at ICLR 2027

LatentShield: Composable Latent Multi-Agent Systems with Decision-Boundary Monitoring

13 views

Abstract

Latent communication lets language-model agents exchange continuous representations directly, but removes the textual handoffs inspected by conventional text-based monitors. We introduce LatentShield, which combines a two-scale communication architecture with safety monitoring at decision boundaries. Rather than propagate latent states through an entire agent graph, we partition the workflow into *Multi-Agent Decision Units* (MDUs): agents communicate in latent space within each MDU, while only token-space messages cross MDU boundaries. Each MDU ends at a designated decision node. Immediately before that node generates the unit's token output, LatentShield reads the contributor latent trajectory and predicts the risk of harmful compliance. A trainable readout applies learned-query attention to pool the variable-length trajectory and uses an MLP aligner to map the pooled representation into the hidden space of a frozen Llama Guard, without decoding the intermediate communication to text. If the risk score exceeds a validation-selected threshold, LatentShield adds a safety instruction before generation. Across the five backbones, LatentShield achieves the lowest mean failure rate among the evaluated latent-monitoring methods while retaining competitive benign utility overall. On Qwen3-4B, it reduces mean failure rate by 36.2% relative to unguarded LatentMAS. A monitor trained on one MDU structure also transfers without retraining to the evaluated unseen MDU sizes and communication graphs. We further evaluate adaptive prompt-injection attacks and agentic workflows. Together, these results support decision-boundary monitoring as a practical safety interface for latent multi-agent systems.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Latent multi-agent systems, latent communication, AI safety, safety monitoring, multi-agent collaboration, modular agent architectures, structural generalization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ztKtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces LatentShield, a safety-monitoring framework for latent multi-agent systems. It organizes latent collaboration into Multi-Agent Decision Units (MDUs), where latent communication is restricted within each unit and token-space communication is used across units. At each decision boundary, LatentShield aggregates contributor latent trajectories using learned-query attention, maps them into the hidden space of a frozen Llama Guard through an MLP aligner, and predicts harmful-compliance risk. If the risk exceeds a calibrated threshold, a safety instruction is added before output generation. The method is evaluated on five language-model backbones across safety, over-refusal, benign reasoning, unseen MDU topologies, adaptive attacks, agentic tasks, and composed workflows, generally achieving lower failure rates than the evaluated baselines while maintaining competitive benign utility.

Critical strengths
  1. The paper addresses an important emerging safety problem in latent multi-agent systems. Existing text-based monitoring techniques cannot directly inspect latent inter-agent communication, and the decision-boundary formulation provides a clear and practical point at which safety intervention can be introduced.
  2. The empirical evaluation is relatively broad. The paper evaluates multiple backbones, several safety and benign benchmarks, unseen MDU topologies, adaptive attacks, and agentic/composed workflows. In particular, the structural-transfer experiments provide useful evidence that the monitor can generalize across different communication graphs within the same backbone representation space.
Critical weaknesses
  1. The demonstrated composability is limited to homogeneous-backbone settings. All agents within an MDU share the same backbone, and a separate monitor is trained for each backbone representation space. Therefore, the reported generalization is mainly across graph structures and team sizes, rather than across heterogeneous model families. This limits the applicability of the framework to practical MAS settings involving specialized agents built on different backbones.
  2. The technical novelty of the monitoring mechanism appears relatively incremental, and the experiments do not fully isolate where the gains come from. The core monitor consists of learned-query attention pooling, an MLP representation mapper, and a frozen Llama Guard classifier. Since the Linear Probe baseline is already competitive on several backbones, it is unclear how much of the improvement specifically comes from alignment with Llama Guard rather than simply using a more expressive nonlinear classifier and trajectory aggregation. A stronger ablation or nonlinear-probe baseline would make the methodological contribution more convincing.
Questions for authors
  1. The current experiments assume that all agents within an MDU use the same backbone, and a separate monitor is trained for each backbone representation space. Could the authors clarify how the claimed composability should be interpreted under this restriction, and whether the framework can be extended to heterogeneous-agent settings with different model families?
  2. Could the authors compare LatentShield against a stronger nonlinear latent classifier with comparable capacity, but without the frozen Llama Guard component? This would help clarify whether the observed gains come from leveraging Llama Guard’s safety representations or mainly from the increased expressiveness of the pooling and nonlinear mapping.
Reviewer_h2anOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces LatentShield, a safety-monitoring framework for latent multi-agent systems. It organizes collaboration into Multi-Agent Decision Units (MDUs), with latent communication within each unit and token-based communication between units. Before a decision node generates its output, a trainable attention pooler and MLP aligner map the contributors’ latent trajectory into the representation space of a frozen Llama Guard. The resulting risk score determines whether to insert a refusal-oriented safety instruction. The evaluation covers five agent backbones, unseen collaboration structures, adaptive attacks, and agentic workflows.

Critical strengths
  1. Broad empirical evaluation with demonstrated structural transfer. The experiments consider harmful compliance, over-refusal, and benign capability, providing a more informative assessment than attack success alone. LatentShield achieves the lowest reported MFR among the evaluated latent-monitoring methods across five backbones. The structural-transfer experiment is particularly useful: a monitor trained on a four-agent Qwen3-4B chain is applied without retraining or threshold recalibration to six unseen structures, reducing average MFR from 0.180 to 0.098.
Critical weaknesses
  1. Backbone-specific training limits portability, and cross-model generalization remains untested. Each agent backbone uses a separately trained monitor, so the results across five models demonstrate the applicability of the method after independent training rather than the transferability of a learned guard. The demonstrated structural generalization keeps the agent backbone fixed. There is no evaluation of transferring a Qwen-trained monitor to another model family, transferring between model variants, or monitoring heterogeneous teams containing different backbones. Since the attention query and MLP aligner depend on the backbone’s representation space, the current approach introduces a recurring adaptation burden when the underlying agents change. This limits its suitability as a reusable safety component for evolving or heterogeneous multi-agent systems. Cross-backbone zero-shot and few-shot transfer experiments, together with measurements of data collection, training, and recalibration costs, would substantially strengthen the practical deployment claims.
Questions for authors

How is the cross-model generalization performance of LatentShield?

Reviewer_z3FXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LatentShield, a safety monitoring method for multi-agent systems that communicate through latent representations. The idea is to monitor the agents' hidden states at decision boundaries using attention pooling, an MLP, and a frozen Llama Guard. If harmful behavior is detected, a safety instruction is added before the model generates its response.

The authors test the method across five backbones, different communication structures, and several safety benchmarks. The results show improvements over the baselines, including a 36.2% reduction in mean failure rate on Qwen3-4B.

Critical strengths
  1. The problem is relevant. Safety monitoring in latent multi-agent systems is challenging since intermediate communication is not readable as text. Monitoring at decision boundaries is a reasonable approach.

  2. I like the MDU design. It provides a way to add safety monitoring without forcing agents to communicate entirely through text.

Critical weaknesses
  1. It is unclear how much of the improvement comes from the proposed alignment method and how much comes from Llama Guard itself. A stronger classifier baseline would help establish this.

  2. I am not convinced by the linear probe comparison. It uses a different pooling strategy, so the comparison does not really isolate the benefit of the frozen safety model.

  3. The computational overhead is missing from the evaluation. Using an 8B safety model at every decision boundary could be costly, particularly when multiple MDUs are involved.

  4. The safety intervention relies on adding an instruction rather than blocking unsafe behavior. A model could still ignore that instruction, so I am not sure how reliable this would be under stronger attacks.

Questions for authors
  1. Why not compare against a stronger classifier trained on the same pooled latent representations? Would Llama Guard still offer a significant advantage?

  2. What is the actual inference overhead per decision boundary? How does it scale with the number of MDUs?

  3. Do the authors have results showing how often the monitor correctly detects harmful behavior but the intervention fails to prevent it?

Comments on the paper

1 comment

Thanks for the authors' response. My concerns were basically addressed. I will keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.