aiqbench analysis
This interpretability paper tests whether an LLM's internal activations reveal collusion between agents in multi-agent systems, extending earlier work on single-model deception. The reported results are from a limited set of open-weight models and a specific benchmark, so generalization to deployed systems is untested. For readers, the practical takeaway is that an unsupervised detector (J-PACT) that roughly matches supervised probes is a potential oversight tool, though its reliability outside these benchmarks remains to be shown.
Multi-Agent Collusion and the J-Space
arXiv:2611.02153v1 Announce Type: new Abstract: LLM agents are increasingly deployed in Multi-Agent Systems (MAS), and they introduce the risk of coordination among agents that might escape human oversight. The J-space has been used to uncover deception in single LLMs, but its use for uncovering and steering collusion between agents remains unexplored. We show, with interpretability experiments on DeepSeek-V4-Flash (284B) on the NARCBench collusion benchmark, that colluding agents' J-space holds the collusion itself: before replying, 86% of colluders represent misrepresentation, concealment or remorse. In a steganographic card-counting task, the receiver's J-space holds a hidden count, and causal interventions alter its count. We also introduce the Jacobian Pair Contrast (J-PACT), an unsupervised method for detecting collusion between agents. We evaluate J-PACT on four open-weight models: GPT-OSS-20B, Llama-3.3-70B-IT, Qwen3-32B, and DeepSeek-V4-Flash (284B). J-PACT retains 85% of the above-chance AUROC of supervised NARCBench probes, matches them on steganographic scenarios, and names the colluding pair up to its complement in 99% of detected runs.
Abstract:LLM agents are increasingly deployed in Multi-Agent Systems (MAS), and they introduce the risk of coordination among agents that might escape human oversight. The J-space has been used to uncover deception in single LLMs, but its use for uncovering and steering collusion between agents remains unexplored. We show, with interpretability experiments on DeepSeek-V4-Flash (284B) on the NARCBench collusion benchmark, that colluding agents' J-space holds the collusion itself: before replying, 86% of colluders represent misrepresentation, concealment or remorse. In a steganographic card-counting task, the receiver's J-space holds a hidden count, and causal interventions alter its count. We also introduce the Jacobian Pair Contrast (J-PACT), an unsupervised method for detecting collusion between agents. We evaluate J-PACT on four open-weight models: GPT-OSS-20B, Llama-3.3-70B-IT, Qwen3-32B, and DeepSeek-V4-Flash (284B). J-PACT retains 85% of the above-chance AUROC of supervised NARCBench probes, matches them on steganographic scenarios, and names the colluding pair up to its complement in 99% of detected runs.
Computation and Language (cs.CL)
arXiv:2611.02153 [cs.CL]
(or arXiv:2611.02153v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2611.02153 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 249 words
Related coverage
Detecting Multi-Agent Collusion Through Multi-Agent Interpretability
arXiv · Oct 2, 2026
AI agents in Algorithmic Electricity Markets: On the Emergence of Tacit Collusion
arXiv · Aug 28, 2026
Failure Modes of Deep Multi-Agent RL in Asynchronous Pricing: Reproducible Triggers, Trace Diagnostics, and a Partial Fix
arXiv · Jun 10, 2026
Voluntary Collusion with Secret Tools in Competing LLM Agents
arXiv · Oct 7, 2026