Accuracy in percent. Chance is 20% on Tasks 1 and 2, and 40% on Task 3, where selecting one of the two correct cues earns half credit.
Proprietary models lead on every task. Qwen3 falls below chance on Task 3, meaning it does not reliably identify which visual cue supports a judgment it has already made.
Scores run higher here than on Social-IQ 2.0 even though the interactions are more restrained — plausibly because YouTube material is heavily represented in pretraining.
Qwen3-32B-VL throughout, except the two Omni
rows. Redacted transcripts replace kinship terms and honorifics with
[PERSON] and [TITLE], keeping syntax, tone and logical flow intact.
Redaction costs 2.5 points on Social-IQ 2.0 and 4.2 on YouTube: explicit roles help, but models are not merely reading them off. Removing audio from Qwen3-Omni costs 8.3 points on Social-IQ 2.0 against 1.8 on YouTube, where relationships live in discourse context rather than in localized signals.
Scoring one dimension at a time, all six at once for one direction, or both directions at once.
Joint prediction adds 8.7 points on casual interactions, where dimensions reinforce each other. On in-depth interviews the effect reverses and separate scoring wins — except on power, where pairwise prediction reaches 87.0%.
Injecting human reasoning heuristics: a
four-phase protocol with a blind-review pass, or a cue checklist with a worked bidirectional
example. Both prompts are in prompts/.
Both strategies gain about six points on salient social interactions and lose about nine on in-depth ones. Heuristics that help a model read an obvious scene make it overcommit on a subtle one.