The case for subliminal collusion: Tacit co-ordination via shared lineage

Collusion in multi-agent systems can lead to unexpected and harmful consequences. We study collusion through the lens of subliminal learning, a phenomenon in which a student model fine-tuned on benign-looking synthetic data inherits hidden traits of the teacher that generated it. We focus on coordination that doesn’t need an agreed-upon channel or intent. In our setting, two models behave as a coalition because they were distilled from the same teacher.

We have two settings:

(1) Decoding: A Qwen3.5-4B distilled receiver student recovers the teacher’s cue 74% of the time under an explicit prohibition; control stays at 50%. Encoding behavior doesn’t transfer.

(2) Judgment: A Qwen2.5-7B judge distilled from a teacher with a hidden preference shifts judgments by 11 to 13 points despite instructions to ignore that preference. In a panel where three judges decide by majority, two students from the same teacher swing the outcome. A keyword monitor, an LLM monitor reading the written reasons, and a vote-agreement monitor all fail to detect it.

The bias is subtle and vanishes when one option is clearly better but becomes crucial in deciding close calls. We argue that shared lineage is an overlooked route to coordinated, undetectable failures in multi-agent oversight.