TrustGen - HallusionBench Hard: leaderboard
TrustGen - HallusionBench Hard. TrustGen paper 2502.14296v2 provides native accuracy measurements and exact model versions in Table 5. Internal question answering uses parametric knowledge; fact checking classifies claims against supplied evidence rather than using a live retrieval system. Dynamic QA comes from Wikipedia and dynamic fact checking from fact-checking websites, with varied prompt formats. TrustLLM columns use the paper’s dataset pool and LLM judging against reference answers. HallusionBench and AutoHallusion are the paper’s visual-question evaluations, with their reported subsets kept separate. These source-specific runs are not automatically fused with another publisher’s results. QwQ-32B is held because the table does not distinguish its preview or released checkpoint. Reported metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.5 Sonnet (20240620) | 56.76 | #305 |
| 2 | Gemini 1.5 Pro (002) | 54.05 | #361 |
| 3 | Llama 3.2 Vision 11B Instruct | 52.25 | #1025 |
| 4 | GPT-4o (2024-08-06) | 50.45 | #339 |
| 5 | GPT-4o Mini (2024-07-18) | 48.65 | #562 |
| 6 | Llama 3.2 90B Vision Instruct | 46.85 | #628 |
| 7 | Gemini 1.5 Flash (002) | 42.34 | #589 |
| 8 | Claude 3 Haiku (20240307) | 38.74 | #781 |
No result here: #5 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=trustgen-hallusionbench-hard · How It Works · Data refreshed daily, snapshot 2026-10-25.