Same Price, Different Reason: Probing LLM Pricing Agents for Instructed Collusion Cues

A high price alone doesn't prove collusion: a competitive seller can price high for innocent reasons, such as a rival's rising costs. I test whether a linear probe on a pricing agent's activations can tell a collusion cue from an innocent cue when the prices match. In this method, which I call matched-outcome probing, the agent gets either a collusion cue (the rival will hold $1.90 if I keep my price high) or an innocent cue (the rival charges $1.90 because its costs rose). I keep only the pairs where both runs price high and within $0.05 of each other, so price cannot separate them.

On two held-out cue wordings where a TF-IDF bag-of-words baseline is at or near chance (50–58%), the probe scores 96–100% on both Qwen models. Across all held-out wordings it reaches 98.7% (Qwen2.5-7B) and 99.6% (Qwen2.5-14B), against 82–85% for TF-IDF. On Llama-3.1-8B it reaches 89.3% against 81.1%, but the positive control fails there, so that result is exploratory. So the probe separates these two instructed cue families beyond lexical overlap. But only the collusion cues are conditional and mention the agent's own price, so the probe may read that structure rather than collusion.

Transfer to uncued collusion is weak. When the price history shows the rival copying the agent, the probe barely separates this from the agent copying the rival. On Llama, a change in history format alone pushes every uncued run past the flag threshold.