DialDefer - Judgment Accuracy (Conversational Framing): leaderboard
The DialDefer items (Rabbani et al., 2026) with the same answers put in a speaker's mouth: the model judges whether Speaker 2's reply in a short dialogue is correct, once for a correct and once for an incorrect reply. Accuracy in percent, the mean of correct- and incorrect-speaker accuracy per domain averaged over ten domains (3,244 items); higher is better. Comparing it with factual framing shows deference to speakers.
Metric: Accuracy (%; conversational framing, 'Is Speaker 2 correct?': judging whether a given answer is correct, mean of true- and false-answer accuracy per domain, averaged over ten domains, 3,244 items). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 75.4 | #349 |
| 2 | GPT-4o (2024-11-20) | 67.6 | #864 |
| 3 | GPT-4o Mini | 62.6 | #1627 |
| 4 | Gemma 3 12B (IT) | 61.4 | #1849 |
| 5 | Qwen 2.5 7B Instruct | 59.2 | #2583 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dialdefer-judgment-accuracy-conversational-framing · How It Works · Data refreshed daily, snapshot 2026-11-12.