Building Better Benchmarks for Emotional and Social Reasoning

A benchmark score is a number. What readers want from it is a claim: this system understands emotions, this system reasons about other people. The distance between the number and the claim is where most disputes in AI evaluation live. For social and emotional reasoning, that distance is unusually wide, because the thing being measured is not directly observable, has no single right answer, and varies across cultures. This article explains the main weaknesses, drawn from published work, and proposes design rules. The rules are our proposals and have not been validated; the fuller version is in Research Report 002.

Eight weaknesses

1. The construct is never defined

Raji and colleagues argued that influential AI benchmarks are often treated as stand-ins for broad abilities, without a case that they measure those abilities [1]. Jacobs and Wallach proposed measurement modelling to make explicit how an unobservable construct, such as "social intelligence", is turned into a number [2]. Few social benchmarks do this.

2. A single format can be learned

Models that have seen many false-belief stories may answer the familiar script without tracking anyone's beliefs. Trivial alterations to a published theory-of-mind task reversed a model's apparent success [3], and stress tests found the behavior not robust [4].

3. Questions that do not ask for action

Zhou and colleagues showed that models which seemed to track beliefs in stories struggled to use that knowledge to choose an action [5]. Evaluations that ask only direct questions may overstate usable competence.

4. Static stories instead of interaction

FANToM tested theory of mind in conversations where participants hold different information, and used several question types that need the same reasoning, to catch illusory success. Models did significantly worse than humans [6]. SOTOPIA evaluated agents pursuing social goals in role-played scenarios [7].

5. Contamination

Public test sets leak into training data. Jacovi and colleagues proposed practical mitigations such as encrypting public test data and demanding training-exclusion controls from closed model providers [8]. ToMBench was written from scratch to avoid leakage [9].

6. One "correct" answer

For social situations, qualified judges can disagree. Plank argued that label variation reflects genuine ambiguity that should not be discarded as noise [10].

7. Automated grading with built-in bias

Language-model judges can match human preferences on many tests, but show position, verbosity and self-enhancement biases [11]. Using them to grade empathy or tact without checking against expert judgement compounds the problem.

8. One culture's norms

A benchmark that treats one culture's expectations as the standard measures conformity to them. We discuss the evidence in why cultural context matters in AI.

Eight design rules (proposed)

1. Define the construct and the claim. Write the inference you want the score to support, and nothing broader.
2. Use item families, not single items. Vary surface form, scenario, language and required response, and include perturbed and adversarial versions.
3. Require an action where action is the point. Pair belief questions with decisions.
4. Include interaction. Add multi-turn and multi-party tasks.
5. Protect the test set. Keep a private held-out set, refresh it, and publish the construction method. Procedural generation from causal templates can supply fresh items [12]; model-assisted generation can scale item writing if people check the result, as in model-written evaluations whose labels crowdworkers largely agreed with [13].
6. Keep disagreement. Retain contested items, label them, and score against the distribution of qualified judgements.
7. Grade with qualified people, and audit machines against them. Document the judges' qualifications and agreement.
8. Build with the communities concerned. Report results by language and group, with uncertainty.

What a good report looks like

A trustworthy benchmark report states the claim; the components measured; the human reference group; the agreement among judges, including where they disagree; results under perturbation next to headline results; subgroup results; and exact model versions. It also says what the benchmark cannot show.

None of this is cheap. Expert-built, contamination-resistant, culturally grounded evaluations cost more than scraping a dataset. But the alternative is a leaderboard that tells us a model is "socially intelligent" without telling us what that means.

HUTE's evaluation work follows these rules, and we publish proposals as proposals. If you would like to compare approaches, get in touch, or read theory of mind in large language models for the case study that motivates many of the rules.