LogicScore - HotpotQA Essentiality: leaderboard
LogicScore (Yan et al., 2026) Essentiality on 1,939 chain-structured HotpotQA questions (distractor setting): the model writes a cited step-by-step answer and a short answer from the retrieved documents, and GPT-4o-mini splits the long answer into propositions and entity triples. Essentiality is the share of those propositions that lie on a recovered path from the gold answer back to a question entity, and 0 when no path is found. Percent, higher is better; single run.
Metric: Essentiality (%; share of the answer's extracted propositions that lie on the recovered question-to-gold-answer path, 0 when no path is recovered; 1,939 HotpotQA distractor-setting items). Source: arxiv.org. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 14B | 52.01 | #1235 |
| 2 | Qwen 3 30B A3B | 50.97 | #1124 |
| 3 | Qwen 3 4B | 50.39 | #2297 |
| 4 | Qwen 3 1.7B | 50.07 | #5551 |
| 5 | DeepSeek R1 | 49.6 | #467 |
| 6 | GPT-5.1 | 48.37 | #225 |
| 7 | Qwen 3 8B | 47.47 | #1838 |
| 8 | O3 | 47.44 | #224 |
| 9 | Qwen 3 235B A22B | 45.62 | #622 |
| 10 | GPT-4o | 44.9 | #741 |
| 11 | DeepSeek V3.2 | 44.81 | #403 |
| 12 | Gemini 3 Pro | 39.69 | #135 |
| 13 | Qwen 3 0.6B | 33.11 | #8392 |
Interactive version: theaggregate.ai/benchmark?slug=logicscore-hotpotqa-essentiality · How It Works · Data refreshed daily, snapshot 2026-11-12.