Pushing Back, Not Deleting: Scale, Channels and Causal Control of Collusion Between LLM Pricing Agents

We study when collusion emerges between LLM pricing agents, what carries the coordination, and which interventions actually break it. In a repeated Bertrand oligopoly whose competitive and joint-profit prices are solved analytically, the collusion index has a principled zero and one, so no rubric or LLM judge is needed. Across Qwen2.5 at 14B, 32B and 72B, neutral markets converge to competition at the small scale and negotiate explicit cartels at the large one, with message content inverting from competitive narration to cartel language.

Two countermeasure results invert expectations. Restricting the channel to four symbols, the natural defence against covert coordination, made collusion almost four times worse, because free-text negotiation anchored on fairness and capped the cartel. And ablating the causal direction in the residual stream did nothing, while negative steering along the same direction at the same layer destroyed the cartel. For defenders, the operation matters, not just the feature.

We further find two conditions that reach equal collusion while leaving opposite evidentiary traces, so transcript auditing and price-parallelism auditing each miss one; that internals become readable only after the model reasons; and that one compromised agent raises prices across a market it does not control, with an interior optimum that makes over-steering self-defeating.