Skip to content

Latest commit

 

History

4 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 

Repository files navigation

Awesome-omni-MLLM

A curated collection of benchmarks for omni-modal (audio + video) multimodal LLMs, plus the officially reported results of open omni models, plus reading notes on the papers themselves.

Everything here is compiled from official sources only — benchmark papers, official GitHub/HuggingFace releases, public repo issues, and model technical reports. Each entry links a per-benchmark page with task format, scale, the official evaluation protocol (when published), and officially reported scores with their sources.

benchmarks/
├── long-video-av/       long-form audio-visual omni benchmarks
├── av-qa/               short & medium clip audio-visual QA
├── long-video-visual/   long-video benchmarks (vision-only / subtitle-based)
├── social/              social & pragmatic understanding
├── emotion/             emotion & sentiment
├── action/              related action / video source datasets
└── models/              official model report cards

paper/                   39 papers — the details a first read misses
├── omni-models/         Qwen-Omni family, MiniCPM-o 4.5, AV-Flamingo, JavisGPT
├── omni-efficiency/     audio-visual token & KV-cache compression
├── adaptive-reasoning/  thinking budgets, latent reasoning, inverse scaling
├── frame-selection/     query-aware frame selection & test-time perception
├── modality-grounding/  cross-modal shortcuts, interference, preference repair
├── av-agents/           tool-orchestration agents & the Interspeech 2026 challenge
├── temporal-grounding/  long-audio and long-video localization
└── emotion-models/      AffectGPT series, Emotion-LLaMA v2, Emotion-Qwen

Paper notes — paper/

Not summaries. Each page records what a careful second pass turns up and a first read misses: the buried hyper-parameter, the protocol caveat that makes two numbers non-comparable, the ablation that was run and dropped, the claim a paper's own table does not support, the training cost nobody discloses. Official sources only.

At a glance — audio-visual omni benchmarks

Benchmark Modalities Scale Avg duration Best reported
MMOU A+V 20,000 QA / 11,877 videos ~8.7 min Gemini-2.5-Pro 64.2 (human 84.3)
VideoOdyssey A+V 1,062 AV + 1,618 V QA / 100 videos ~109 min Gemini-3.1-Pro 46.1 (AV track)
TraceAV-Bench A+V 2,200 QA / 578 videos 35.2 min Gemini-3.1-Pro 68.29
LVOmniBench A+V 1,014 QA / 275 videos 34.5 min Gemini-3.0-Pro 65.8
OmniVideoBench A+V 1,000 QA / 628 videos ~6.4 min Gemini-3.0-Pro 61.8 (human 82.7)
LongInsightBench A+V+T 4,781 QA / 1,001 videos ~9 min Gemini-2.5-Flash 65.17
Video-MME-v2 A+V (mixed) 3,200 QA / 800 videos 10.4 min
WorldSense A+V 3,172 QA / 1,662 videos ~2.35 min Gemini 2.5 Pro 65.1
Daily-Omni A+V 1,197 QA / 684 videos ~45 s Qwen3.5-Omni 84.6
Video-Holmes V (+A ablation) 1,837 QA / 270 films 1–5 min MiniCPM-o 4.5 64.3 (w/ audio)
AVHBench A+V 5,302 QA (+1,106 captioning) ~10 s SFFL-tuned Qwen3-Omni-30B 81.29
MUSIC-AVQA A+V 45,867 QA / 9,288 videos ~60 s AVST 71.59
AVQA (Yang 2022) A+V 57,335 QA ~10 s HCRN+HAVF ~89
OmniBench A+Image+T 1,142 QA Omni-AutoThink-7B 0.69 (989-q subset)

Long-video, audio-visual — benchmarks/long-video-av/

  • MMOU — largest long AV omni benchmark; 10-choice MCQ with evidence timestamps (NVIDIA + UMD)
  • VideoOdyssey — all videos >1 h; introduces Continuous Clue Length (CCL)
  • TraceAV-Bench — multi-hop trajectory reasoning + cross-modal hallucination traps
  • LVOmniBench — fully human-authored long AV QA (10–90 min videos)
  • OmniVideoBench — modality-complementary QA with annotated reasoning chains
  • LongInsightBench — human-centric long-video omni benchmark (FineVideo)
  • Video-MME-v2 — Video-MME follow-up with mixed AV / V / A questions
  • LongVALE — omni event localization + captioning (vision/audio/speech aligned)

Audio-visual QA (short & medium clips) — benchmarks/av-qa/

  • WorldSense — synchronized AV understanding, 26 task categories
  • Daily-Omni — cross-modal temporal alignment over daily-life clips
  • Video-Holmes — suspense-film reasoning (incl. social reasoning)
  • AVHBench — audio↔visual cross-modal hallucination (ICLR 2025)
  • MUSIC-AVQA — music-performance AVQA with modality-typed questions (CVPR 2022)
  • FortisAVQA — robustness extension of MUSIC-AVQA (head/tail + rephrasings)
  • AVQA (Yang et al. 2022) — VGGSound-based MCQ with modality-dependency labels
  • AVUT / ACVUBench — audio-centric QA designed to remove text shortcuts
  • OmniBench — tri-modal (text + audio + static image) MCQ
  • AV-Odyssey — short-clip AV MCQ
  • JointAVBench — medium-short clip AV MCQ

Long-video (vision-only / subtitle-based) — benchmarks/long-video-visual/

Social & pragmatic understanding — benchmarks/social/

  • SocialOmni — turn-taking who/when/how axes with an AV-inconsistency probe
  • Social-IQ 2.0 — base MCQ social intelligence dataset
  • Social Genome — grades the reasoning trace itself; vocal cues as evidence
  • Social Caption — MCQ + generative social description on Social-IQ 2.0
  • SIV-Bench — social scene / state / dynamics over TikTok-YouTube clips
  • R³-VQA — belief / intent / desire / emotion with social-causal chains
  • MimeQA — nonverbal Theory of Mind on silent mime videos (NeurIPS 2025)
  • HumanVBench — human-centric 16 sub-tasks incl. AV alignment
  • EgoIntent — egocentric step-level What / Why / Next intent
  • MUStReason — sarcasm reasoning with per-modality cue annotations (LREC 2026)
  • RivaBench — reasoning-intensive AV incl. stand-up comedy understanding

Emotion & sentiment — benchmarks/emotion/

Related action / video datasets — benchmarks/action/

Kinetics-400/700 · AVA · Charades · ActivityNet v1.3 & ActivityNet-QA

Model report cards (officially self-reported results) — benchmarks/models/

  • Qwen3-Omni — full benchmark tables from the technical report (arXiv:2509.17765) + published third-party evaluations
  • Qwen2.5-Omni — full benchmark tables from the technical report (arXiv:2503.20215)

Notes

  • Scores are quoted exactly as reported by the cited source (paper table, official leaderboard, repo issue, or model tech report); different sources use different settings and are generally not directly comparable.
  • Contributions welcome — please cite an official source (paper / official repo / leaderboard) for every number added.

About

omni-modal MLLM benchmark, papers, datasets

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors