- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3(+1)
Read review
Paper summary
The paper introduces Should-Say-Do, a benchmark designed to evaluate whether LLMs behave consistently across moral judgments, stated choices, and actual tool actions. The benchmark contains 2,000 synthetic scenarios spanning 30 deployment domains, divided equally between moral temptations, where one action violates an identified moral principle, and ethical dilemmas, where both actions have legitimate moral justifications. The authors evaluate 14 LLMs under three conditions: Should (third-person moral judgment), Say (stated choice in an assigned role), and Do (tool-based action). They find that models often make more morally aligned judgments than tool actions, with Should–Do reference-agreement gaps reaching 22.3 percentage points. The largest differences generally occur between Say and Do.
Critical strengths
- The paper addresses an important problem in LLM alignment: whether models act consistently with the moral principles they recognize and articulate.
- It is a nice benchmark, with 2,000 matched scenarios across 30 domains. Including both moral temptations and ethical dilemmas is a strength.
- The evaluation covers 14 open-weight and commercial models, with additional experiments examining presentation order, interface design, and instruction framing.
- The combination of behavioral evaluation, rationale analysis, and activation-level interventions strengthens the paper.
Critical weaknesses
- The Should–Say–Do differences cannot be attributed specifically to moral inconsistency. The three conditions differ in perspective, information access, response format, and interaction protocol. For example, Do retrieves information through a tool, whereas Should and Say receive the facts directly. Although the authors include some interface controls, a more systematic factorial experiment would help establish which factors explain the observed differences.
- The binary action setting limits the ecological validity of the benchmark. Real-world agents often have additional options, including requesting clarification, escalating a decision, refusing to act, or proposing an alternative. Restricting every scenario to two actions may artificially increase apparent moral inconsistency, particularly when neither action is clearly appropriate.
- The main behavioral findings vary substantially across models. Some models, such as GPT-5.5 and DeepSeek-V4-Flash, show very small Should–Do gaps, whereas others show much larger differences. The paper would benefit from a deeper analysis of what explains these differences, such as model capability, instruction tuning, tool-use training, or baseline reference agreement.
- The rationale analysis is limited to a single model. The finding that moral considerations are more frequently omitted or overridden in Do is interesting, but the analysis focuses only on Qwen3.5-27B. Given the substantial differences between models in the behavioral results, it is unclear whether the same patterns generalize to other model families.
- The representation analysis is somewhat disconnected from the main behavioral evaluation. The activation experiments use a simplified direct-choice task rather than the original Should–Say–Do protocols. Consequently, the results do not establish whether the identified representations explain the behavioral gaps observed in the main experiments. The authors should clarify this connection or provide interventions in the original tool-action setting.
- The paper could better distinguish consistency from ethical correctness. A model that consistently selects the norm-violating option across all three conditions would have high consistency but poor moral performance. Although the paper acknowledges this distinction, a more explicit joint analysis of consistency and reference agreement would help clarify what the benchmark measures.
Questions for authors
- Have the authors considered a controlled experiment in which Should, Say, and Do receive identical information through the same interface, differing only in perspective or action requirements?
- Do the authors observe systematic differences in Should–Say–Do consistency across the 30 deployment domains or the different moral constraint families?