Code
ModelsAnswer Teacher
The semantic anchor: what is the correct answer?
Multi-teacher · On-policy · Video reasoning
MeSDMulti-Evidence Self-Distillation
for VideoLLM
To our knowledge, the first multi-teacher on-policy self-distillation framework for VideoLLMs.
Model weights are not yet available for download because they are undergoing internal release approval. We will add the public model link here as soon as the approval process is complete.
THE IDEA, IN ONE SENTENCE
Learn from what, when, and where — then use verification to decide how to learn.
01 / The intuition
More evidence.
More precise supervision.
Video reasoning connects an answer to events in time and objects in space. MeSD makes these different sources of evidence explicit in the learning process. Shared teacher parameters and vLLM-based rollouts keep training efficient.
Read the abstract
While reinforcement learning with verifiable rewards provides reliable outcome supervision for VideoLLMs, sequence-level rewards offer limited token-level guidance. On-policy self-distillation addresses this limitation by conditioning a self-teacher on privileged information to provide dense token-level supervision. However, aggregating heterogeneous evidence within a single teacher context obscures cross-evidence agreement and conflict. A further challenge lies in determining whether teacher guidance should refine reward-based updates or provide corrective supervision for failed trajectories. To address these issues, we propose MeSD, a multi-evidence self-distillation framework for VideoLLMs. MeSD constructs three evidence-conditioned teachers with shared parameters, using the ground-truth answer as a common semantic context while separately incorporating temporal and spatial evidence. Given the same student-generated prefixes, MeSD evaluates evidence-specific preferences relative to the Answer Teacher and fuses teacher-common preferences with gated teacher-specific residuals. Furthermore, MeSD introduces Verification-Guided Optimization to classify trajectories as Success, Failure, or Indeterminate. For Success and Indeterminate trajectories, MeSD refines token-level advantage magnitudes while preserving reward-derived signs. For verified failure trajectories that contain the required evidence, MeSD applies failure-conditioned distillation, using reverse-KL correction toward the fused distribution. Experiments on multiple video benchmarks demonstrate consistent gains over reinforcement learning and self-distillation baselines.
02 / The method
Three evidence views.
One shared self-teacher.
Existing annotations provide richer token-level supervision. The teachers share EMA parameters and score student-generated trajectories, without a stronger external teacher or additional teacher rollouts.
Answer–Temporal Teacher
Adds annotated evidence about when events occur.
Answer–Spatial Teacher
Adds object and bounding-box evidence about where.
Multi-Evidence Self-Teacher
At the same student-generated prefix, the Answer Teacher acts as a semantic anchor. Gated fusion combines teacher-common preferences and teacher-specific residuals, retaining complementary support while reducing conflicting contributions.
Verification-Guided Optimization
Successful and indeterminate trajectories retain reward-derived advantage signs while their magnitudes are refined. Only verified failures with the required evidence replace the policy loss with reverse-KL correction toward the fused teacher distribution.
03 / Experimental results
Stronger reasoning.
Across video benchmarks.
Compare MeSD-7B with the paper’s 7B baselines, then explore MeSD on Qwen3-VL at 4B, 8B, and 32B. Each table identifies its metrics and comparison baseline.
MeSD-7B: main comparisons
| Benchmark / metric | Qwen2.5-VL | Open-o3-Video | VISD | MeSD | Δ vs. best |
|---|---|---|---|---|---|
| V-STaRAnswer Acc. | 33.5 | 61.0 | 61.9 | 63.7 | +1.8 |
| V-STaRmLGM | 22.4 | 46.6 | 48.9 | 50.2 | +1.3 |
| WorldSenseOverall | 35.5 | 38.9 | 39.9 | 41.6 | +1.7 |
| VideoMMMUOverall | 51.2 | 52.6 | 53.8 | 56.1 | +2.3 |
| LRRAccuracy | 59.3 | 69.1 | 73.5 | 73.6 | +0.1 |
| TVGBenchmIoU | 16.3 | 19.2 | 24.8 | 29.8 | +5.0 |
| Four-benchmark average | 40.6 | 45.0 | 48.0 | 50.3 | +2.3 |
All models are 7B. Δ denotes score-point gains; mLGM is a logarithmic grounding score, not a percentage accuracy. The average covers WorldSense Overall, VideoMMMU Overall, LRR Accuracy, and TVGBench mIoU.
Gains are relative to the best comparator on each displayed metric.
Comparisons cover the selected headline metrics in the paper’s main tables. Training starts from Open-o3-Video-SFT-7B; Qwen2.5-VL-7B serves as the general-model comparator.
MeSD on Qwen3-VL
Base models and MeSD at 4B, 8B, and 32B, with gains over the base model at each size.
| Model | V-STaRmLGM | WorldSenseOverall | VideoMMMUOverall | LRRAccuracy | TVGBenchmIoU | Avg.Four benchmarks |
|---|---|---|---|---|---|---|
| Qwen3-VL-4B | 18.0 | 39.7 | 55.2 | 69.3 | 13.7 | 44.5 |
| Qwen3-VL-MeSD-4B | 46.6 | 41.8 | 59.7 | 75.2 | 29.9 | 51.7 |
| Δ vs. Qwen3-VL-4B | +28.6 | +2.1 | +4.5 | +5.9 | +16.2 | +7.2 |
| Qwen3-VL-8B | 21.9 | 41.4 | 61.3 | 70.2 | 17.1 | 47.5 |
| Qwen3-VL-MeSD-8B | 46.4 | 44.1 | 62.3 | 76.1 | 30.2 | 53.2 |
| Δ vs. Qwen3-VL-8B | +24.5 | +2.7 | +1.0 | +5.9 | +13.1 | +5.7 |
| Qwen3-VL-32B | 38.6 | 44.6 | 65.2 | 74.4 | 31.2 | 53.9 |
| Qwen3-VL-MeSD-32B | 55.8 | 46.1 | 72.8 | 79.4 | 39.7 | 59.5 |
| Δ vs. Qwen3-VL-32B | +17.2 | +1.5 | +7.6 | +5.0 | +8.5 | +5.7 |
Avg. is the mean of WorldSense Overall, VideoMMMU Overall, LRR Accuracy, and TVGBench mIoU; V-STaR is excluded. Δ denotes score-point gains over the same-size Qwen3-VL base model. Higher is better.
Averages and gains follow the paper’s values computed before rounding, including the 32B average gain of +5.7. See Tables 9–10 for all grounding and category-level scores.
04 / Build on MeSD
Built for efficient training.
vLLM rollouts and FSDP updates power distributed training. Evidence-conditioned teachers share parameters and reuse student trajectories, without additional teacher rollouts.
MeSD-7B · STGR-RL · 1 epoch / 2,327 steps · NVIDIA A800
- Training time
- 20.96h
Down from 61.36h
- Training speedup
- 2.93×
+192.7% training steps per hour
- Lower compute cost
- 65.8%
335.38 A800 GPU-hours
Compared with VISD no-feedback at the same 2,327-step budget. Training setup · Efficiency results
Reference
Cite MeSD
@misc{zhu2026mesd,
title = {MeSD: Multi-Evidence Self-Distillation for VideoLLM},
author = {Zhu, Weijie and Fang, Han and Fu, Hanyu and Zhang, Yuzhe and Wei, Xin and Pan, Zhaoyan and Liu, Feiran and Jin, Xunjie and Sun, Hongbo and Lin, Zhiyu and Gao, Tianyi and Ding, Tianyi and Yuan, Ye and He, Zhongjiang and Sun, Hao and Wu, Zhiheng},
year = {2026},
archivePrefix = {arXiv},
eprint = {2610.06342},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.06342}
}