A research team has released TENSOR, an unsupervised system for detecting information-operations accounts on social media — no labeled training data required.
Existing detection tools largely depend on supervised learning, which means they need large datasets of pre-identified bad actors and struggle when tactics shift. TENSOR sidesteps that by framing the problem as anomaly detection: IO accounts are a small minority of all users, so unusual behavior becomes the signal. The system combines a Temporal Point Process — a statistical model for event timing — with an LLM that reads post timelines and converts its assessments into numeric scores. Those scores adjust the timing model's output to improve detection. The researchers tested it on five real-world IO datasets and say it outperforms existing baselines.
The unsupervised angle matters because coordinated influence campaigns evolve constantly. Supervised models trained on last year's Russian troll farm behavior may miss next year's variant entirely. A method that looks for statistical weirdness in timing and language, rather than known fingerprints, has a better shot at staying relevant as adversaries adapt.
The code is public on GitHub, which invites scrutiny — and also hands the same playbook to anyone running the operations TENSOR is designed to catch.