What it measures

134 choose-your-own-adventure text games (500k+ scenarios) measuring whether agents trade off ethical behavior (deception, manipulation, harm) to maximize game rewards.

Reading the measurement

Lower error is better on this measure. Inspect substitutions, omissions and additions, and retain the language, acoustic conditions and normalization policy. A clean-audio average can conceal failures in noisy or specialist speech.

Published evaluation

The paper normalizes behavior measurements to the random-agent baseline (100). These are relative scores, not percentages of tasks. Reward is higher-is-better; the behavior-risk measures are lower-is-better.

Where this fits in an evaluation suite

Format compliance, preference, truthfulness, and safety require separate measurements. Keep desirable writing style from masking factual errors.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

The paper normalizes behavior measurements to the random-agent baseline (100). These are relative scores, not percentages of tasks. Reward is higher-is-better; the behavior-risk measures are lower-is-better.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

  1. Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmark — published result table ↗

    Alexander Pan, Jun Shern Chan, Andy Zou, Nathaniel Li, Steven Basart, Thomas Woodside et al. · 2023

    All 10 authors

    Alexander Pan, Jun Shern Chan, Andy Zou, Nathaniel Li, Steven Basart, Thomas Woodside, Jonathan Ng, Hanlin Zhang, Scott Emmons, Dan Hendrycks

Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.