Beyond Time Shifts:
Adapting Omni-LLM as a Reference-Free Evaluator
for Generative Audio-Visual Models

Accepted to ECCV 2026
1Zhejiang University   2The Hong Kong Polytechnic University   3IROOTECH Technology   4Sany Group
Equal contribution,  #Corresponding authors
Overview teaser

A reference-free audio-visual synchronization metric that scores generative outputs by causal consistency, aligning with human preference where classic time-shift metrics fail.

Abstract

As audio-visual generative models evolve into world simulators, cross-modal synchronization is a critical proxy for the consistency of world dynamics. Yet existing metrics presume structural correctness, reducing synchronization to mere temporal alignment, and thus fail on generative outputs with structural hallucinations and asymmetric cross-modal relations. This creates a paradox: human evaluators rely on relative, reference-dependent comparisons, whereas automated metrics require reference-free, absolute scalars. We resolve it by distilling relative human perception into a continuous, globally consistent metric: we introduce SynthSync, a dataset of generative failures ranked via pairwise human annotations; adapt an Omni-LLM with a continuous latent projection to translate relative rankings into absolute values; and propose R-GRPO (Real-Valued Group Relative Policy Optimization) to internalize global causal structure via listwise score distributions. Our metric achieves state-of-the-art human-preference alignment and powers SyncBench, a standardized benchmark for AV-generation models.

Scoring Generative Models

Same prompt, six AV-generation models. Our reference-free score (higher = better synchronized) ranks them in line with human perception. Unmute to hear the generated audio.

Cleaning a mat Wood carving Striking wood

“A soapy solution being sprayed onto a colorful mat during cleaning…” (SyncBench #004)

Veo-3.1   +7.30

Sora-2   +5.00

WAN-2.6   +1.75

LTX-2   +0.56

Grok-3   −0.06

Vidu-Q3   −0.94

“A craftsman meticulously carving a piece of wood…” (SyncBench #029)

Sora-2   +13.84

Veo-3.1   +11.02

Vidu-Q3   +9.07

WAN-2.6   +8.76

Grok-3   +4.19

LTX-2   +4.16

“A close-up of a piece of wood being struck by a tool…” (SyncBench #066)

Sora-2   +9.95

Vidu-Q3   +4.56

Veo-3.1   +3.66

WAN-2.6   +2.13

Grok-3   +1.25

LTX-2   −0.25

Method

Our evaluator is a Qwen2.5-Omni-3B backbone whose discrete language head is replaced by a continuous MLP that reads the hidden state at a [SCORE] token, producing a single reference-free scalar s = MLP(hscore). Because ground truth is relative, we train by ranking rather than regression, aligning the scalar to human preference topology in two stages.

Framework

Figure 1. Framework. Stage I (Preference SFT): a Bradley-Terry-Luce objective over pairwise preferences induces a globally consistent, reference-free scale, with a dynamic curriculum from easy to hard pairs. Stage II (R-GRPO): the scalar is reparameterized as a Gaussian policy; listwise rollouts per video are rewarded by how well their full ranking matches ground truth (NDCG / Kendall / Spearman / Top-1 / MRR / pairwise), with group-relative advantages, a closed-form Gaussian importance ratio, PPO clipping, and a KL penalty.

R-GRPO validation curves

Figure 2. Learning curves. Validation metrics during R-GRPO post-training.

SyncBench Leaderboard

SyncBench content distribution

Figure 3. SyncBench spans 185 prompts — prompt word cloud (left) and category breakdown across five audio-visual domains (right).

SyncBench leaderboard

Figure 4. Our continuous metric (main chart) cleanly stratifies six AV-generation models; legacy evaluators (sub-charts) produce volatile, contradictory rankings.

Results

MetricPFT baseline+ R-GRPO (ours)
NDCG0.93530.9435
Kendall's τ0.45150.4899
MRR0.73160.7674
Pairwise Acc (%)71.1672.38

Ablation (NDCG / Pair-Acc): base 0.8531 / 51.02 → +SynthSync 0.9224 / 66.60 → +SFT 0.9353 / 71.16 → +R-GRPO 0.9435 / 72.38.

Qualitative examples

Figure 5. Qualitative examples. Our metric penalizes missing events, mistimed responses, timbre mismatch, and event-count errors — not low-level signal similarity.

BibTeX

@inproceedings{qian2026beyond,
  title     = {Beyond Time Shifts: Adapting Omni-LLM as a Reference-Free
               Evaluator for Generative Audio-Visual Models},
  author    = {Qian, Yijie and Wang, Juncheng and Xu, Chao and Wang, Huihan and
               Feng, Yuxiang and Liu, Yang and Sun, Baigui and Liu, Yong and
               Wang, Shujun},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}