Benchmark Almanac

LOKI

GPT-4o (snapshot unspecified) 68.8% tops its leaderboard LLaVA-OneVision 72B 60.8% leads the models we track
Not in model cards

Models distinguish synthetic content from real content across images, videos, audio, text, and 3D data. The benchmark checks real-or-synthetic labels and annotated abnormal details, with separate tasks for explaining anomalies.

The suite spans 26 categories. Explanation tasks have different evaluation requirements from classification tasks.

Made by
OpenDataLab
Released
October 2024
Kind
General visual questions
Level
Everyday
Scored
Checked automatically
Answers
Multiple choice
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Images, Video, Audio, 3D → Text
Interaction
Single turn
Refresh
One fixed set
Run by
Anyone
Last model added
August 2024
Scores refreshed
November 20th, 2026

In the authors’ words

With the rapid development of AI-generated content, the future internet may be inundated with synthetic data, making the discrimination of authentic and credible multimodal data increasingly challenging. Synthetic data detection has thus garnered widespread attention, and the performance of large multimodal models (LMMs) in this task has attracted significant interest. LMMs can provide natural language explanations for their authenticity judgments, enhancing the explainability of synthetic content detection. Simultaneously, the task of distinguishing between real and synthetic data effectively tests the perception, knowledge, and reasoning capabilities of LMMs. In response, we introduce LOKI, a novel benchmark designed to evaluate the ability of LMMs to detect synthetic data across multiple modalities. LOKI encompasses video, image, 3D, text, and audio modalities, comprising 18K carefully curated questions across 26 subcategories with clear difficulty levels. The benchmark includes coarse-grained judgment and multiple-choice questions, as well as fine-grained anomaly selection and explanation tasks, allowing for a comprehensive analysis of LMMs. We evaluated 22 open-source LMMs and 6 closed-source models on LOKI, highlighting their potential as synthetic data detectors and also revealing some limitations in the development of LMM capabilities. More information about LOKI can be found at https://opendatalab.github.io/LOKI/

Abstract of LOKI: A Comprehensive Synthetic Data Detection Benchmark using Large Multimodal Models (arXiv:2410.09732, CC0)
20
Entries in its paper, 15 models we track

GPT-4o (snapshot unspecified) tops the leaderboard with 68.8% LLaVA-OneVision 72B leads with 60.8% among models released by August 2024

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. This is the plain mean of the overall judgment and multiple choice accuracies in the multimodal evaluation. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is LLaVA-OneVision 72B, at 60.8%. This view shows only models we track, each at its best.

The newest model with a score here came out in August 2024, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 GPT-4o (snapshot unspecified) 68.8%
  2. 2 Claude-3.5-Sonnet (snapshot unspecified) 68.2%
  3. 3 LLaVA-OneVision 72B 60.8%
  4. 4 Qwen2-VL-72B-Instruct 60.5%
  5. 5 Gemini-1.5-Pro (snapshot unspecified) 58%
  6. 6 Qwen2-VL-7B-Instruct 57.4%
  7. 7 LLaVA-OneVision 7B 55.2%
  8. 8 Mantis-8B 54%
  9. 9 VILA1.5-40B 53.2%
  10. 10 InternVL2-40B 52.5%
  11. 11 LongVA-7B 52.1%
  12. 12 mPLUG-Owl3-7B 52.1%
  13. 13 InternVL2-8B 52%
  14. 14 InternVL2-26B 51.5%
  15. 15 Phi-3.5-vision-instruct 51.5%
  16. 16 MiniCPM-V-2_6 51.3%
  17. 17 InternLM-XComposer2.5-7B 49.6%
  18. 18 VILA1.5-13B 48.5%
  19. 19 Llama-3-LongVILA-8B 48%
  20. 20 Idefics2-8B 47.9%

Source: The LOKI paper

#EntryScoreSettingDateSource
1 GPT-4o (snapshot unspecified) 68.8% Official leaderboard
2 Claude-3.5-Sonnet (snapshot unspecified) 68.2% Official leaderboard
3 LLaVA-OneVision 72B 60.8% Aug 2024 Official leaderboard
4 Qwen2-VL-72B-Instruct 60.5% Aug 2024 Official leaderboard
5 Gemini-1.5-Pro (snapshot unspecified) 58% Official leaderboard
6 Qwen2-VL-7B-Instruct 57.4% Aug 2024 Official leaderboard
7 LLaVA-OneVision 7B 55.2% Aug 2024 Official leaderboard
8 Mantis-8B 54% Official leaderboard
9 VILA1.5-40B 53.2% Apr 2024 Official leaderboard
10 InternVL2-40B 52.5% Jul 2024 Official leaderboard
11 LongVA-7B 52.1% Jun 2024 Official leaderboard
12 mPLUG-Owl3-7B 52.1% Aug 2024 Official leaderboard
13 InternVL2-8B 52% Jul 2024 Official leaderboard
14 InternVL2-26B 51.5% Jul 2024 Official leaderboard
15 Phi-3.5-vision-instruct 51.5% Aug 2024 Official leaderboard
16 MiniCPM-V-2_6 51.3% Aug 2024 Official leaderboard
17 InternLM-XComposer2.5-7B 49.6% Jul 2024 Official leaderboard
18 VILA1.5-13B 48.5% Apr 2024 Official leaderboard
19 Llama-3-LongVILA-8B 48% Official leaderboard
20 Idefics2-8B 47.9% Apr 2024 Official leaderboard
#ModelScoreSettingDateSource
1 LLaVA-OneVision 72B 60.8% Aug 2024 Official leaderboard
2 Qwen2-VL-72B-Instruct 60.5% Aug 2024 Official leaderboard
3 Qwen2-VL-7B-Instruct 57.4% Aug 2024 Official leaderboard
4 LLaVA-OneVision 7B 55.2% Aug 2024 Official leaderboard
5 VILA1.5-40B 53.2% Apr 2024 Official leaderboard
6 InternVL2-40B 52.5% Jul 2024 Official leaderboard
7 LongVA-7B 52.1% Jun 2024 Official leaderboard
8 mPLUG-Owl3-7B 52.1% Aug 2024 Official leaderboard
9 InternVL2-8B 52% Jul 2024 Official leaderboard
10 InternVL2-26B 51.5% Jul 2024 Official leaderboard
11 Phi-3.5-vision-instruct 51.5% Aug 2024 Official leaderboard
12 MiniCPM-V-2_6 51.3% Aug 2024 Official leaderboard
13 InternLM-XComposer2.5-7B 49.6% Jul 2024 Official leaderboard
14 VILA1.5-13B 48.5% Apr 2024 Official leaderboard
15 Idefics2-8B 47.9% Apr 2024 Official leaderboard

Rankings by part

We read 10 parts from the benchmark’s paper in 2 groups: judgment and multiple choice. Each ranks models on one part of the test. Qwen2-VL-72B-Instruct leads 3 of them, and 4 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

Video

15 models we track have a score here.

  1. 1 InternVL2-40B ↑5 62%
  2. 2 InternVL2-8B ↑7 60.8%
  3. 3 LongVA-7B ↑4 60.4%
  4. 4 Qwen2-VL-72B-Instruct ↓2 59.6%
  5. 5 Qwen2-VL-7B-Instruct ↓2 59.5%
  6. 6 VILA1.5-40B ↓1 59.2%
  7. 7 InternLM-XComposer2.5-7B ↑6 58.4%
  8. 8 MiniCPM-V-2_6 ↑4 57.2%
  9. 9 LLaVA-OneVision 7B ↓5 56.8%
  10. 10 Phi-3.5-vision-instruct ↑1 56.8%

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Other labs
Benchmark paper Not tested since Oct 2024

Source: The LOKI paper

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2024 Near the ceiling (85%) LLaVA-OneVision 72B VILA1.5-40B VILA1.5-13B 0% 20% 40% 60% 80% 100% 2024 Near the ceiling (85%) LLaVA-OneVision 72B VILA1.5-40B

Source: Official LOKI leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Vision

All 636 in Vision
  • MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy. Qwen 3.6 Plus 86%
  • MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
  • CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
  • MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy. Seed2.1 Pro 90.7%
  • MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy. Qwen 3.8 Max 97.7%
  • MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy. Qwen 3.8 Max 85.9%

Data audit

9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 2 years ago, in August 2024, for Qwen2-VL-72B-Instruct. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
  • Pass Independent runs Scores from the benchmark’s paper (20 entries).
  • Pass Models scored 15 models we track, plus 5 other entries on its leaderboard.
  • Warning Newest score The newest came 2 years ago, in August 2024, for Qwen2-VL-72B-Instruct.
  • Note Frontier models None of its models are on the capabilities index.
  • Pass Headroom The best is 60.8%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 15 rows for models we track and 5 other entries at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check