Reordering Competitors or Adding Sentiment Text, With Identical Numbers, Shifts LLM Pricing Agents' Behavior
arXiv 2609.18357 (16 Sep 2026) introduces market signal injection, which manipulates numerical formatting, competitor ordering or qualitative commentary without issuing any instruction, and evaluates nine open-weight models in simulated Bertrand duopoly and triopoly markets plus three proprietary models in duopoly. Sentiment-based attacks caused the largest behavioral shifts, propagated to other firms and altered profits and consumer surplus, with susceptibility varying by model family and larger models not consistently more robust. Episode-held-out probes separate baseline from attacked activations in all eleven re-evaluated model-condition pairs at linear AUC 1.00, though the authors note this separability does not by itself identify harmful decisions; input canonicalization removes the tested attacks.
↳ Follow the thread