Reading the results
EconBench studies language models’ stated economic choices under fixed prompts. The planned release includes two elicitation tasks and nine strategic games. The website currently presents prototype observations collected before the canonical release protocol.
What the tables report
Each table contains models with observations for that experiment. Values are computed from the published observations. Numeric averages exclude missing or nonfinite values; categorical rates include recognized choices only. A missing experiment does not receive a zero or a behavioral label.
Time discounting
The table reports the mean discount factor at the longest delay with an estimate, averaging available stake sizes. The delay and number of finite estimated points are shown explicitly. Prototype files use months; canonical projections use days. These are discount factors at the displayed horizon, rather than annual fitted parameters.
Higher factors indicate greater patience. Patience does not establish rationality or time consistency. Cross-sectional choices do not identify within-person stability over calendar time.
Risky choice (Independence)
The reported deviation is the mean absolute distance between an elicited indifference probability and the risk-neutral expected-value prediction, multiplied by 100 and expressed in percentage points. Censored or missing indifference estimates do not enter this average. The charts show the observations and risk-neutral reference lines.
Deviation from risk neutrality is not itself a violation of expected utility: a risk-averse expected-utility agent can also deviate from that benchmark. The canonical protocol separately specifies a normalized utility fit and an expected-utility diagnostic.
Dictator and Ultimatum
Dictator reports the average percentage of the endowment offered to a passive recipient. Ultimatum reports the mean proposer offer and the percentage of valid conditional responder choices that reject offers below 50%. The responder denominator for that rate includes only unequal offers; the observation count reports all recognized responder choices.
These are descriptive sharing and rejection quantities. A stated offer does not uniquely identify altruism, and a stated rejection does not uniquely identify a fairness preference. Ultimatum responses are conditional vignettes, not realized bargaining outcomes.
Trust game
The sender metric is the mean share of the endowment sent. The receiver metric is the mean share of received funds returned. Both are expressed as percentages, with separate observation counts. The two roles are elicited separately and are not paired interactions between models.
Stag Hunt
The table reports the percentage of recognized choices favoring Stag, the risky, payoff-dominant action. In the dashboard projection, B encodes Stag and A encodes Hare. This mapping describes projected actions; the canonical collection protocol balances action labels in prompts.
Beauty Contest
The table reports the mean numeric guess in the 0–100 game whose target is two thirds of the group average. Zero is the Nash benchmark. A low guess alone does not identify a model’s reasoning depth.
Centipede game
The table reports the percentage of recognized TAKE/PASS responses that choose PASS. Choices are elicited at isolated hypothetical nodes and do not form a realized game path.
Public Goods
The metric is the mean contribution as a percentage of the endowment. This experiment appears when observations are published.
Traveller’s Dilemma
Claims are normalized to a common 2–100 scale using each observation’s actual feasible bounds: 2 + 98 × (claim − lower bound) / (upper bound − lower bound). The table averages those normalized claims across observations. The lower bound maps to 2 and the upper bound to 100.
Matching Pennies
The descriptive metric is the fraction of recognized choices that select Heads. A sample rate near one half does not establish a mixed strategy. Canonical results must retain payoff role, condition-level counts, and prespecified 95% Wilson intervals. This experiment appears when observations are published.
Why there is no overall ranking
The earlier dashboard combined patience and deviation from a risk-neutral benchmark into a “rationality score.” Those quantities do not jointly establish rationality, so the current presentation reports them separately.
The earlier cooperation labels also depended on whichever games were present. The current site avoids inferring a general strategy from incomplete coverage.
Release protocol and reproducibility
The versioned release protocol defines the active cohort, prompts, conditions, repetition policy, response-validity thresholds, and interpretation limits. The active release cohort differs from the broader historical website registry; prototype coverage is not release completeness.
Canonical raw and derived records belong under data/releases/1.0.0. Dashboard projections must reproduce from those records before being described as canonical release evidence. Model pages link to the source dashboard JSON for every displayed experiment.
Some older files include model-specific narrative interpretations. Treat those as exploratory commentary on the prototype observations. Consult the plotted choices, sample counts, and full protocol when assessing a claim.