Why Naïve Visual Reasoning Fails and How ProLaViT Enables Step‑by‑Step Latent Reasoning in Multimodal LLMs

In the era of multimodal large language models (MLLMs), recognizing objects in an image is easy, but performing multi‑step spatial perception, geometric analysis, and logical inference remains challenging. Existing approaches either explicitly generate intermediate images (e.g., Anole, ThinkMorph) at high computational cost, or perform a single latent‑space prediction (e.g., Mirage, LVR) that often collapses in accuracy and attends to irrelevant background.

ProLaViT: Progressive Latent Visual Thought

To address these issues, Tencent BAC proposes ProLaViT (Progressive Latent Visual Thought). Its core principle is to avoid premature conclusions and instead conduct step‑wise deduction in a continuous latent space, following a causal chain of Locate → Focus → Isolate that gradually tightens visual attention until the target is precisely locked.

Endogenous Self‑Distillation Pipeline

Training multi‑step latent models faces the problem that intermediate steps lack ground‑truth images. Instead of relying on external visual experts (e.g., SAM, DINO, DepthAnything), ProLaViT uses endogenous self‑distillation:

For each training sample, code automatically synthesizes K auxiliary images corresponding to each reasoning step (e.g., crop(I, bbox), segment(I, target)) using geometric transformations; the code guarantees accuracy without manual labeling.

The frozen visual encoder V extracts teacher features from these auxiliary images; gradients are not back‑propagated, keeping the encoder fixed.

A cross‑modal attention module aligns the LLM‑generated latent thoughts with the teacher features, minimizing a distillation loss.

Distance‑Weighted Diversity Loss

Multi‑step latent inference often suffers from latent collapse, where later steps become overly similar. ProLaViT adopts a metric‑learning‑inspired loss that penalizes excessive similarity while allowing reasonable proximity for adjacent steps. The loss incorporates a causal distance weight, giving larger penalties to steps that are farther apart in the reasoning chain.

Training Objective and Stages

The overall objective combines language modeling with latent‑space supervision, and training proceeds in four stages: latent anchor initialization, endogenous knowledge distillation, structured chain evolution, and global capability integration. This schedule encourages stable emergence of progressive reasoning.

Empirical Results

Implemented on Qwen2.5‑VL‑7B‑Instruct, ProLaViT was evaluated on MMVP, VisPuzzle, VStar, ChartQA, BLINK, and CV‑Bench. The full model with distance‑weighted diversity loss achieved 75.11% overall high‑level accuracy, surpassing baselines. Notably, it improved VisPuzzle by +2.25% and BLINK‑Jigsaw by +10.00% over one‑step latent prediction.

Stability tests show 78.99% accuracy on ChartQA and a +16.67% jump on BLINK‑Jigsaw, confirming the effectiveness of the Hypothesize → Critique → Verify dialectical chain for logical tasks.

Ablation Studies

Removing the progressive decomposition (single‑step prediction) reduces performance by an average of 7.45% across benchmarks, demonstrating the necessity of multi‑step reasoning.

Replacing the native visual encoder with external experts (DINOv2, SDXL‑VAE) degrades fine‑grained tasks, suggesting that the native encoder provides naturally aligned supervision.

Randomizing the order of the four reasoning steps drops average accuracy by 6.24% (VisPuzzle –6.25%, CV‑Bench –8.52%), highlighting the importance of the causal sequence.

Future Directions

ProLaViT establishes a new paradigm of structured latent progressive deduction, bridging the modality gap between text‑based chain‑of‑thought and visual perception. Extending this approach to video temporal reasoning and more complex geometric transformations could further advance general and interpretable multimodal intelligence.