Price Afterimage: Markets Remember What Agents Forget

The paper turns history sensitivity into a useful measurement problem: how much of a planted price gap survives after the planted rows leave every prompt. Fresh-seed replication, scripted relay replacements and the hint-versus-sham comparison make the central result persuasive. The distinction between imitation and demonstrated punishment is especially welcome. The next step should be a direct mid-game log-reset arm: a cold start is not automatically the counterfactual outcome of resetting an established market. Please also make natural, economically consistent histories a primary replication, since the effect falls from 0.84 to about 0.45 with real openings. Add a sham containing an irrelevant number to isolate the informational value of the best-response hint, and explain sensitivity to scoring gates that exclude near-cost models. Finally, a near-zero afterimage within this design does not establish that a model has a fixed intrinsic pricing trait. Clarify the advance over Keppo et al.: On the Fragility of AI Agent Collusion already documents leader-follower price anchoring, and Trusting AI in Competitive Markets describes adaptive advice-feedback paths. Your more specific contribution is persistence after direct context expiry, with randomized plants and relay interventions, rather than the general observation that pricing depends on interaction history.

This paper shows how AI pricing agents get stuck following old price patterns left behind by previous sellers, but it demonstrates that adding a simple "best-response hint", i.e., a prompt intervention that explicitly provides the agent with its optimal pricing move, snaps the model out of historical inertia and pushes high prices back down to competitive levels. This fix is a practical and lightweight tool to break tacit algorithmic collusion, though the study is limited by the brief sprint's simple computer simulations. Since transparent marketplaces cannot simply erase public price history, future research could explore whether presenting past prices through summary statistics like rolling averages can break this "afterimage" anchoring without hiding data from agents.

This paper convincingly shows that an opening price history can shape later prices even after its original records have left every agent’s prompt. The fresh-seed replication and experiments with different rival policies strengthen this finding. A price-following rival keeps the effect strong, while a best-responding rival reduces it to almost zero, providing useful evidence that later interaction carries the earlier influence. Reporting the cold-start price level and sensitivity to inherited history separately is also a valuable contribution to model evaluation.

One qualification concerns the size of the effect. Qwen3-14B retains about 0.84 of the initial price gap with level-shifted histories, compared with about 0.45 using unchanged real openings. Both results support persistence, but its magnitude depends on the history provided. This limits the reach of the headline estimate without weakening the central finding.

Overall, the paper offers a clear insight supported by well-chosen experiments. It makes a useful case for including initial histories when evaluating pricing agents.