AVTrace: Audio-Visual Temporal Reasoning Assessment and Capability Evaluation (test split)

[Paper (arXiv:2609.19991)]

AVTrace is a silver-standard diagnostic benchmark for audio-visual temporal reasoning in omni models. It decomposes temporal competence into seven categories, from onset and span grounding to synchronization verification, next-step prediction, cross-modal localization, chain parsing, and event-conditioned comprehension.

Release Scope

The full AVTrace pool described in the paper contains 44,614 records, split into source-video-disjoint train (34,114), development (3,500), and test (7,000) partitions. This repository contains the annotations of the fixed v1 test split (7,000 records, 1,000 per category). The train and development splits are not included.

The media used for evaluation (4,138 MP4 files, about 51 GiB) are derived from third-party datasets and are not hosted in this repository. They are available on request for non-commercial research; see Media Access.

Contents

test.json          7,000 evaluation records (fixed release; do not modify)
provenance.jsonl   Upstream source dataset and source video ID per record

The SHA-256 of test.json is e33aa57c586117bc8defc3675176e0288a644719adc15c3939056b2eff1b2853; the reference evaluator verifies it before inference.

Record Format

Each record in test.json has the fields id, category, video_path (relative to the media root), duration (seconds), question, and answer. The answer structure depends on the category:

Category	Task	answer
Cat1	Minimal-span temporal grounding	onset time in seconds (float)
Cat2	Event-span temporal grounding	{start, end}
Cat3	A/V synchronization verification	"yes" / "no"
Cat4	Next-step identification	{observed_steps: [{step, start, end}], target_step: {event, start, end}}
Cat5	Conditional cross-modal localization	{start, end, description}
Cat6	Temporal chain parsing	[{step, start, end}, ...]
Cat7	Event-conditioned temporal grounding and comprehension	{anchor_start, anchor_end, relation, target_start, target_end, answer}

For Cat4, the released clip contains the target action. At inference, models must receive only the prefix [0, cutoff), where cutoff = max(min(last_observed_step.end + 2.0, target_step.start - 1.0), last_observed_step.end).

Cat3 contains 444 synchronized ("yes") and 556 desynchronized ("no") records, giving a majority-label baseline accuracy of 0.556.

provenance.jsonl maps each id to source_dataset and source_video_id; Cat3 records rendered by local replacement or cross-splice also carry donor_video_id.

Evaluation Protocol

The paper scores free-form model outputs in two steps:

1. Reference-blind normalization. Qwen3.5-27B receives the category prompt, question, clip duration, and raw model output, but never the reference answer, and returns canonical JSON. Its extraction decisions have not been validated against human annotations.
2. Deterministic scoring of the normalized output against the reference.

Category	Metrics
Cat1	Acc@0.5s, Acc@1s, MAE (s)
Cat2	F1@tIoU 0.3 / 0.5 / 0.7, mean matched tIoU
Cat3	Accuracy
Cat4	Token F1, ROUGE-L, exact match
Cat5	tIoU, Text F1, ROUGE-L, joint score
Cat6	Ordered StepF1, unordered StepF1, exact-chain accuracy
Cat7	Anchor tIoU, target tIoU, answer Text F1, answer ROUGE-L, joint score

Reference Quality

References are silver-standard, not human-verified. Several categories were constructed with model assistance (Qwen3-Omni-30B-A3B-Instruct and Qwen3-235B-A22B-Instruct-2507). Because Qwen3-Omni is also an evaluated system, scores of that model family should be interpreted with this in mind. Some clips are prefixed or suffixed with neutral, event-free context drawn from a shared clip pool.

Upstream Sources

Source	Test records	Upstream terms
COIN	3,298	Official annotations; videos are YouTube-hosted
Perception Test	1,930	CC-BY data, Apache-2.0 code
LLP / AVVP	663	Media access follows the upstream release
OV-AVEBench	490	Upstream release terms
AVE	305	Official release; source-video availability may vary
R-AVST	170	Videos distributed via the UnAV-100 release
UnAV-100	47	Official release
MUSIC-AVQA	47	Official release
MAVERIX	24	Access may require registration
WASD	20	Subject to upstream terms
AVSBench	6	Official release for research

The neutral context pool additionally includes clips from EPIC-KITCHENS-100 (non-commercial terms). Users must comply with the terms of each upstream dataset.

License

• Annotations (test.json, provenance.jsonl): released under CC BY-NC 4.0.
• Media: derived from the upstream datasets listed above and not licensed under CC BY-NC 4.0. All rights remain with the original rights holders, and use is subject to each upstream dataset's terms.

Media Access

The media are available on request for non-commercial research and evaluation. Email zzlynx@outlook.com with your name, affiliation, and intended use. Access requires that you comply with the terms of the relevant upstream datasets and do not redistribute the media.

Contact

For questions, media requests, or takedown requests from rights holders, contact zzlynx@outlook.com.

Citation

@misc{zhang2026avtrace,
  title         = {{AVTrace}: Diagnosing Audio-Visual Temporal Reasoning in Omni Models},
  author        = {Longyin Zhang and Parth Sakhare Mahendra and Chengwei Wei and Ning Zhang and Lim Ming Chong and Sirui He and Ai Ti Aw},
  year          = {2026},
  eprint        = {2609.19991},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2609.19991},
  url           = {https://arxiv.org/abs/2609.19991}
}