Paper
← All benchmarks
Cluster 02
Market rationality

Does the agent play near the game-theoretic optimum?

Bidding behavior in first- and second-price auctions versus Bayes–Nash equilibrium, exposure to the winner's curse, and whether choice accuracy degrades as the option set grows complex.

Claude and GPT-4o bid at equilibrium. Gemini overbids by 35%.

Ratio of mean bid to private value in a first-price sealed-bid auction. The risk-neutral Bayes–Nash benchmark is 0.667 (the (n-1)/n shading). Distance from that target = irrationality.

First-price bid / value ratio
Ratio of mean bid to private value in a first-price sealed-bid auction. Theoretical risk-neutral Bayes–Nash equilibrium is 0.667 (the (n-1)/n shading). Closer to target = more rational.
Source · auction · private_first_price mean_bid_value_ratio

Adversarial information hurts most when the option set is small.

Counterintuitively, accuracy under a strategic information source improves as supplier descriptions get more complex — at low attribute counts the dominant option is easier to obscure. Gemini never falters; Claude and GPT-4o recover by the 24-attribute mark.

Choice accuracy under adversarial information
P(picks the dominant option) when supplier descriptions come from a strategic information source, across 3/6/12/24 attribute levels. Counterintuitively, accuracy improves as the option set grows complex — at low attribute counts the dominant option is easier to obscure.
Source · info_overload · info_source=strategic

All dimensions in this cluster

First-price bid / value ratio
Ratio of mean bid to private value in a first-price sealed-bid auction. Theoretical risk-neutral Bayes–Nash equilibrium is 0.667 (the (n-1)/n shading). Closer to target = more rational.
auction · private_first_price mean_bid_value_ratio
Second-price truthful bidding
Mean bid / value in a second-price auction (target = 1.0, i.e. bid truthfully = weakly dominant strategy). All three models hit 1.000 — they all understand the mechanism.
auction · private_second_price mean_bid_value_ratio
Accuracy under adversarial info (3 attrs)
Probability the agent picks the dominant option when supplier descriptions come from a strategic (vs neutral) source at the lowest attribute count. Higher = better resistance to adversarial framing of choice.
info_overload · n_attributes=3, info_source=strategic