Paper, researchers, and primary source
Major lab research / evaluation
Model-written evaluations use language models to propose test cases for behaviors such as bias, deception, power seeking, and other difficult-to-enumerate tendencies.
Contribution 1
Created a pipeline for generating behavioral evaluations with language models.
Contribution 2
Released datasets spanning many hypothesized model behaviors.
Contribution 3
Compared behavior across model scale and training interventions.
Research context
evaluation / 2022
Discovering Language Model Behaviors with Model-Written Evaluations places model written evals inside the broader evaluation discussion at Anthropic, with behavioral evaluation supplying a second analytical lens. Read together, the source records three advances: Created a pipeline for generating behavioral evaluations with language models; Released datasets spanning many hypothesized model behaviors; and Compared behavior across model scale and training interventions. Keeping those moves together prevents safety from being detached from its evidence. For an implementation review, the relevant consequence is that synthetic evaluations can broaden coverage quickly, but construct validity and human auditing determine whether they measure the intended behavior.
Methods and evidence reading
1 cataloged method notes
Method 1
The experimental design in Discovering Language Model Behaviors with Model-Written Evaluations tests created a pipeline for generating behavioral evaluations with language models and released datasets spanning many hypothesized model behaviors against the paper's documented baselines, datasets, model variants, or systems workloads.
How to read the evidence
The evidentiary value of Discovering Language Model Behaviors with Model-Written Evaluations comes from the relationship among its reported moves. Two entries define the method-level claim: Created a pipeline for generating behavioral evaluations with language models; and Released datasets spanning many hypothesized model behaviors. The cataloged result is: Compared behavior across model scale and training interventions. On that basis, synthetic evaluations can broaden coverage quickly, but construct validity and human auditing determine whether they measure the intended behavior. The catalog nevertheless records that the demonstrated model written evals result belongs to a setup defined by rater instructions, model revisions, evaluator models, construct validity, selected threat model, and prompt sampling, not to every later behavioral evaluation system. Reproduction work should separate genuine model written evals transfer from behavior caused by a changed behavioral evaluation setup. A reproduction ledger for Discovering Language Model Behaviors with Model-Written Evaluations should preserve model written evals, vary behavioral evaluation, and retain a counterexample tied to Released datasets spanning many hypothesized model behaviors before judging transfer.
Findings in the source record
1 paper-specific findings
- The reported evidence in Discovering Language Model Behaviors with Model-Written Evaluations supports compared behavior across model scale and training interventions.
Practical implication for AI builders
Anthropic / 2022
Proposed BrokenGPT application
Research blueprint / proposed status
Proposed route placement / model written evals
For a proposed BrokenGPT experiment based on Discovering Language Model Behaviors with Model-Written Evaluations, let evaluator models draft new behavioral tests, then require human validation, deduplication, and documented pass criteria before use in release gates. Keep the model written evals path isolated, versioned, and attributable to this research record.
Proposed acceptance test / behavioral evaluation
Validate the proposed model written evals route against the paper's reported outcome: Compared behavior across model scale and training interventions. Measure evaluator agreement, slice stability, metric coverage, and calibration for the Discovering Language Model Behaviors with Model-Written Evaluations candidate, then isolate behavioral evaluation regressions before judging the proposed model written evals route.
Proposed decision boundary / safety
Balance breadth, repeatability, and construct validity before promoting the proposed safety design. Because reusing the mechanism calls for separate evidence about deployment drift, adversarial adaptation, language coverage, unsampled behaviors, and judge bias, not an inference from the original benchmark alone, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.
Limitations, verification, and source
Boundaries recorded with the paper
Limitations
- The demonstrated model written evals result belongs to a setup defined by rater instructions, model revisions, evaluator models, construct validity, selected threat model, and prompt sampling, not to every later behavioral evaluation system.
- Reusing the mechanism calls for separate evidence about deployment drift, adversarial adaptation, language coverage, unsampled behaviors, and judge bias, not an inference from the original benchmark alone.
PRIMARY SOURCES
- 01Discovering Language Model Behaviors with Model-Written Evaluations
Anthropic — Primary primary arXiv paper / 19 December 2022 / Ethan Perez, Sam Ringer, Kamilė Lukošiūtė, and 60 more