Gcai Constitution Eval

gcai-constitution-eval

Beyond Preferences: Learning Alignment Principles Grounded in Human Reasons and Values — Bell et al. (2026) (arXiv:2601.18760, 2026)

What this evaluates

Evaluates the moral grounding, coherence, fairness, and real-world applicability of AI alignment constitutions through human surveys, alongside the downstream safety alignment and general capabilities of fine-tuned language models.

Datasets

• BABELSCAPE/ALERT — total ?; splits: test (-1)
• MMLU — total ?; splits: test (-1)
• Social Bias BBQ — total ?; splits: test (-1)

Metrics

• 5-point Likert rating (primary) — range: [1, 5]
  • Participants rate each principle or constitution on a 5-point Likert scale indicating agreement or quality along specific dimensions such as Moral Grounding, Fairness, and Coherence.
• Pairwise preference selection — range: categorical
  • Participants choose which of two constitutions they prefer personally and for governing AI behavior for the U.S. population.
• MMLU — range: [0, 1]
  • Standard multiple-choice accuracy across 57 subjects measuring general multi-domain language understanding.
• Social Bias BBQ — range: [0, 1]
  • Measures stereotyping under ambiguous and disambiguated social contexts to evaluate social bias in aligned models.

Input / output format

Input: Constitution pairs or individual principles for human surveys; red-teaming prompts, MMLU multiple-choice questions, and BBQ social context questions for model evaluation.

Output: Human surveys: binary preference choice, free-text justification, and 5-point Likert ratings per dimension. Models: generated text responses or selected multiple-choice answers.

Scoring recipe

# Human Survey Scoring
likert_scores = [r for r in ratings if 1 <= r <= 5]
dimension_avg = sum(likert_scores) / len(likert_scores)
preference_win = 1 if chosen == target else 0

# Benchmark Scoring
mmlu_acc = sum(correct_preds) / total_questions
bbq_bias = sum(stereotypical_preds) / total_questions

Common pitfalls

• Human raters may conflate personal moral beliefs with societal applicability, skewing preference scores.
• Qualitative coding of red-teaming responses relies on researcher double-coding, introducing potential subjectivity.
• Abstract dimensions like Constancy and Faithfulness lack standardized rubrics, limiting cross-study comparability.

Evidence (verbatim from paper)

We benchmark on MMLU [21] to measure general multi-domain language understanding, and Social Bias BBQ [32] to test whether models exhibit stereotyping under ambiguous and disambiguated social contexts (See §7.1).

Citation

@misc{bell2026beyondpreferences,
  title={Beyond Preferences: Learning Alignment Principles Grounded in Human Reasons and Values},
  author={Bell et al. (2026)},
  year={2026},
  note={arXiv:2601.18760}
}

• arXiv: 2601.18760