LongNovel (32K): leaderboard
LongNovel (2026) asks whether a summary of a long novel contains a hallucination, over Chinese novels and English BookSum chapters with eight hallucination types and a human-revised test set. This board is balanced accuracy in percent at the 32K-token context scale, where 50 is chance; higher is better.
Metric: Balanced accuracy (%; M, about 32k tokens context scale; binary detection of hallucinated summaries of Chinese novels and English BookSum chapters (8 hallucination types, human-revised test set); direct prompt with the summary at the end; an output without a valid consistency score counts as an incorrect prediction; chance is 50). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Flash | 81 | #258 |
| 2 | GPT-5.2 Instant | 77 | #293 |
| 3 | Claude Sonnet 4.5 | 66.5 | #250 |
| 4 | Qwen 3 8B | 60.8 | #1830 |
| 5 | Qwen 3 14B | 59.8 | #1233 |
| 6 | Qwen 3 32B | 57.5 | #960 |
| 7 | GLM-4 9B Chat | 51 | #2759 |
| 8 | internlm2.5-20B-chat | 50.5 | #2546 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=longnovel-32k · How It Works · Data refreshed daily, snapshot 2026-11-13.