EQ-Bench

Rates the intensity of four emotions a character feels at the end of a GPT-4-written dialogue, scored by distance from an author-set reference; the version harnesses run today is now legacy on the publisher's own site.

Also known as: Emotional Intelligence Benchmark for Large Language Models

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categoryreasoning
Subcategoryemotional and social intelligence (dialogue emotion-intensity prediction)
Page statussuperseded
MetricEQ-Bench score (distance from reference ratings)
Directionhigher_is_better
Unitpoints
Dataset size171
Dataset licenceMIT
PublisherIndependent researcher (no institutional affiliation given in the paper)

What it measures

EQ-Bench shows a model a short GPT-4-generated dialogue depicting a scene of conflict or tension between two characters, then asks it to rate the intensity (0-10) of four named emotions one character is likely feeling at the end of the scene -- typically one clearly present emotion, one clearly absent one, and two that require a careful, nuanced reading of the exchange. The format was designed as an improvement on an earlier psychometric test (SECEU) adapted for LLMs: it removes the requirement that ratings sum to a fixed total, uses reference answers set by the benchmark's author rather than averaged from a human crowd, and can be scored objectively without a human or model assessor interpreting free text. It is a single-turn, English-language, text-only task that the paper frames as testing "emotional understanding" specifically, one of four branches of the psychological construct of emotional intelligence.

Task format

Given a dialogue and four named emotions, output an intensity rating from 0 to 10 for each in a fixed format; scored by a distance-from-reference formula rather than exact match (see How it is scored). The original protocol requests both a first-pass and a self-revised answer and keeps whichever scores higher, though later pipeline versions disable the revision step by default since it was found to help only about 8% of the time.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub