Tencent BAC’s ProLaViT pushes multimodal reasoning into a progressive latent chain

机器之心 · wechat · 2026-07-21

What ProLaViT changes

Tencent BAC proposes ProLaViT (Progressive Latent Visual Thought) for multimodal reasoning tasks where models can see an image but still struggle to reason correctly over space, geometry, and logic.

It replaces brittle one-shot latent prediction with a progressive latent chain: Locate → Focus → Isolate for spatial tasks, and Hypothesize → Critique → Verify for logical tasks.
Instead of relying on external vision teachers such as SAM, DINO, or DepthAnything, it uses endogenous self-distillation from the model’s own pretrained vision encoder.
Training uses a programmatic synthesis pipeline that automatically creates step-wise auxiliary images such as crops and segmentations, so no human labels are needed.
A distance-weighted diversity loss prevents latent collapse by keeping adjacent reasoning steps somewhat similar while pushing distant steps apart.

Results

Built on Qwen2.5-VL-7B-Instruct, the full method reports a 75.11% average accuracy across MMVP, VisPuzzle, VStar, ChartQA, BLINK, and CV-Bench.

Biggest gains show up on harder spatial reasoning tasks such as VisPuzzle (+2.25%) and BLINK-Jigsaw (+10.00%) versus one-shot latent prediction.
The paper argues the progressive chain is not just more accurate, but also more stable and more interpretable.
Ablations show that removing the progressive chain, shuffling the step order, or swapping in external teachers all hurt performance.

Overall, the work argues that structured latent reasoning can bridge the gap between expensive explicit image generation and unstable implicit one-step inference.

Original post →

Tencent BAC’s ProLaViT pushes multimodal reasoning into a progressive latent chain

More from Multimodal