Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2606.28958

When Latent Agents Lie: KV-Cache Integrity in Multi-Agent LLM Collaboration

Luís Brito·Carlos Baquero
View original source ↗Read source document ↗Export metadata ↓

Abstract

LLM agents can share more than text. In some systems, an agent can send a short visible message while also passing its full KV-cache state to another model. This hidden state can help the final model combine evidence from several agents, but it is also hard to inspect. A visible message may look harmless even if the hidden state has been changed. We study this problem in a multi-agent question-answering setup. Specialists each see part of the evidence, send a short commitment, and pass full KV-cache state to a coordinator. In clean runs, this latent collaboration improves over a matched text-only version. On transformed HiddenBench with Qwen3-4B, it reaches EM/F1 of 0.338/0.486, compared with 0.231/0.369 for text collaboration. Qwen3-8B and HotPotQA runs show the same direction of improvement. The problem appears when one specialist is malicious. Some false visible commitments can steer answers. More seriously, changing the hidden KV state can collapse performance even when the visible commitment still looks plausible. A verifier that checks only text misses this failure mode. Simple magnitude checks catch some obvious corruptions, but adaptive attacks can evade them while still damaging the final answer. The most reliable fix we find is not to guess whether hidden state looks normal, but to protect it in transport. We implement an HMAC-SHA256 manifest that binds the specialist, session, model, visible commitment, tensor metadata, and payload digest. It accepts all 774 honest replayed payloads and rejects all 295 recorded tampered payloads. The main lesson is that full-KV latent memory can be useful, but it should be treated as a security-sensitive object, not as ordinary internal model state.

Explore related subjects

cs.MA

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Luís Brito, Carlos Baquero. 2026-06-27. When Latent Agents Lie: KV-Cache Integrity in Multi-Agent LLM Collaboration. https://arxiv.org/abs/2606.28958

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

A Task-Oriented Multi-Agent Framework for Complex Wearable Health Analysis

Wearable health questions often combine data retrieval, longitudinal analysis, and health advice over structured records. Prompting a single large language model with a complete record and a composite query obscures whether every request is executed and which evidence supports the answer. We propose a task-oriented multi-agent framework that represents a composite query as distinct intents and typed tasks with explicit intra-intent dependencies. Specialized agents execute retrieval, analysis, and advice tasks; isolated intent states preserve request boundaries and evidence relationships before aggregation. We evaluate the framework on a synthetic dataset of $10{,}000$ virtual users with one month of longitudinal wearable records, covering structured data retrieval, multi-intent recognition, and overall response quality. Across $1{,}500$ retrieval questions, the Query Agent achieves $98.3\%$ accuracy, compared with $97.9\%$ for the Direct LLM baseline, while reducing average query-stage token consumption from $6{,}869$ to $3{,}136$. On $180$ multi-intent questions, the Manager Agent achieves $100.0\%$ Multi-Intent Coverage and $94.4\%$ Multiset Jaccard Similarity. Under the current synthetic evaluation setting, our method receives higher mean Trustworthiness and Transparency scores on both question categories, whereas Actionability does not improve consistently. These results provide preliminary evidence that explicit task organization can support task-relevant data access and data-grounded longitudinal analysis, while leaving health advice generation and validation on real wearable data as open challenges.

cs.MA↗
arXiv2026

Mixed-integer flow formulations for motion planning and decision-making of networked multi-agent systems

This work investigates the use of flow-based connectivity maintenance constraints in mixed-integer linear programming (MILP) trajectory planning and decision-making models for networked multi-agent systems (MAS). We integrate flow-based encodings for standard and k-hop connectivity into MILP multi-vehicle maneuvering models that are widely used alongside receding horizon planning strategies. Their necessity and sufficiency is demonstrated, guaranteeing full coverage of potential network topologies. The flow formulation for standard connectivity decreases the growth of the required inequality constraints from exponential to polynomial w.r.t. the size of the MAS when compared to the state-of-the-art subtour elimination (SEC) method. The flow-based k-hop connectivity constraints decrease the number of required binary variables and decouple its growth from the number of hops. However, the impact of these formulations in performance is not straightforward due to the introduction of a substantial number of continuous flow optimization variables and, in the case of k-hop connectivity, additional inequality constraints. We investigate this trade-off through a statistical evaluation of costs and optimization times using a conventional branch-and-bound commercial solver and trials performed with randomized environments for increasingly larger MAS. The results show that the flow formulation outperforms SEC in standard connectivity problems, enabling the solutions to be computed for larger MAS considering the imposed optimization time limit. The reduction in number of binary variables enabled by the k-hop flow formulations decreases the theoretical worst-case number of iterations required by the branch-and-bound algorithm to compute the global optimal solution. Our results show that this advantage did not translate into improvements in the average performance when compared to the baseline.

cs.MA↗
arXiv2026

Perception-Aware Communication Middleware for Distributed Visual Perception in UAV Swarms

Unmanned Aerial Vehicle (UAV) swarms increasingly support safety-critical applications that rely on distributed visual perception. Meeting the low-latency requirements of these applications can require perception models to execute within the swarm on inference-capable UAVs, creating a need for efficient UAV-to-UAV transport of high-bandwidth perception data. However, the Quality-of-Service (QoS) requirements of perception differ from conventional packet-level QoS; successful delivery of individual packets does not ensure that a complete, timely, and usable image is available for inference. We present a novel perception-aware communication middleware that treats complete perception-data samples as the communication objects for which QoS must be satisfied. The middleware extends a lightweight UDP broker-based publish-subscribe architecture with perception-specific services, including image fragmentation and reconstruction, concurrent packet transmission, priority-aware scheduling, and image quality assessment. The middleware is evaluated on a heterogeneous hardware testbed emulating a UAV swarm using YOLOv8n object detection. Experimental results demonstrate low end-to-end application latency, substantially higher throughput than a lightweight UDP broker, effective prioritization of perception traffic under increasing background load, and mitigation of object-detection degradation through middleware-level image quality assessment. This work provides an initial framework for integrating AI-specific data handling into communication middleware to support emerging distributed AI applications in multi-agent mobile cyber-physical systems.

cs.MA↗