Technical Summary: Vision-aligned Latent Reasoning for Multi-modal Large Language Model

1. Problem Statement

Despite significant advancements in Multi-modal Large Language Models (MLLMs) for tasks like image captioning and visual question answering (VQA), these models struggle with problems requiring extensive multi-step reasoning. The primary bottleneck is the progressive dilution of visual information during long-context generation. As the generated text sequence lengthens, the model's ability to retain and exploit visual signals diminishes, hindering the application of test-time scaling laws observed in pure language models.

Existing approaches to enhance reasoning often rely on:

• Text-centric methods: Supervised fine-tuning (SFT) or reinforcement learning that focuses on linguistic chains of thought (CoT) but suffers from visual signal attenuation over long sequences.
• Static visual re-introduction: Interleaving visual tokens or generating images. However, these methods typically treat visual features as a fixed initial context or rely on static single-view features, leading to a gradual loss of visual context as reasoning progresses.

2. Methodology: Vision-aligned Latent Reasoning (VaLR)

The authors propose Vision-aligned Latent Reasoning (VaLR), a framework designed to dynamically generate vision-aligned latent tokens before each reasoning step. This approach creates "visual checkpoints" that ground the reasoning process in perceptual cues within the latent space, preventing visual information decay.

Core Mechanism

VaLR operates by alternating between two modes during inference:

1. Language Mode: Generates standard text tokens for the response.
2. Latent Mode: Generates non-textual "latent reasoning tokens" that are not directly visible but serve as intermediate reasoning steps.

Crucially, during the latent mode, the model is trained to align its intermediate hidden states with dense visual representations extracted from external, pre-trained vision encoders (e.g., DINOv2/v3, CLIP, SigLIP, $\pi^3$).

Training Pipeline

VaLR employs a two-stage curriculum learning strategy:

• Stage 1 (Standard SFT): The model is fine-tuned on general CoT VQA datasets to establish foundational multi-modal reasoning capabilities. The vision encoder is frozen, and only the language backbone is trained.
• Stage 2 (Latent Token Training with REPA): The model is trained to insert $K$ latent tokens before each language reasoning step. A Representation Alignment (REPA) loss is introduced to align the intermediate features of the MLLM with features from external vision encoders.
  • Alignment Objective: The loss function maximizes the cosine similarity between the projected intermediate features of the MLLM and the patch-wise visual features extracted from the corresponding image frame by the vision encoder.
  • Multi-Encoder Alignment: The framework supports aligning with multiple vision encoders simultaneously to capture complementary strengths (e.g., semantic understanding from CLIP, fine-grained appearance from DINO, and 3D spatial structure from $\pi^3$).

Notably, the external vision encoders are only used during training. At test time, the model relies on the learned visual grounding within its latent tokens without requiring the external encoders.

3. Key Contributions

• Dynamic Visual Grounding: Unlike methods that use static visual features, VaLR dynamically allocates visual details at each reasoning stage, ensuring information preservation throughout long-context reasoning.
• Test-Time Scaling: The framework enables MLLMs to exhibit test-time scaling behavior, where performance improves with longer reasoning chains, a phenomenon previously unobserved in MLLMs due to visual signal attenuation.
• Model-Agnostic Design: VaLR can be adapted to various vision encoders (DINO, SigLIP, CLIP, $\pi^3$) and works effectively even with the standalone vision encoders of the original MLLM.
• Two-Stage Curriculum: A structured training pipeline that first establishes reasoning capabilities and then refines them with visual alignment, facilitating robust long-context understanding.

4. Experimental Results

Extensive evaluations were conducted on multiple benchmarks, including VSI-Bench (3D spatial reasoning), BLINK, MMVP, MMStar, and MathVista.

• Long-Context Reasoning (VSI-Bench): VaLR significantly outperformed existing baselines. Specifically, on VSI-Bench, VaLR improved the accuracy of the base model Qwen2.5-VL from 33.0% to 52.9%, a gain of 19.9 percentage points. It also surpassed strong reasoning models like R1-OneVision and Ocean-R1.
• Perception Tasks: VaLR demonstrated superior performance on perception benchmarks (BLINK, MMVP, V*), achieving state-of-the-art results among open models and outperforming API models like GPT-4o in specific configurations.
• Reasoning Length Analysis: Unlike baseline models that peak at intermediate reasoning lengths and then degrade (hallucination rates increase), VaLR showed monotonic performance improvement as reasoning length increased, confirming its ability to maintain visual grounding over extended chains.
• Ablation Studies:
  • Removing the representation alignment (REPA) component resulted in significant performance drops, confirming the necessity of visual alignment.
  • Using multiple encoders (VaLR-M) yielded synergistic effects, particularly for 3D spatial tasks, compared to single-encoder variants.
  • The method demonstrated high data scalability, achieving comparable performance to vanilla SFT models with significantly fewer training samples (faster convergence).

5. Significance and Claims

The paper claims that VaLR addresses a critical limitation in current MLLMs: the inability to sustain visual context during complex, multi-step reasoning. By injecting vision-aligned latent tokens, the method effectively mitigates the dilution of visual signals, enabling robust long-horizon planning and reasoning.

The authors emphasize that this work facilitates the deployment of MLLMs in complex applications requiring long-context understanding, such as Computer Use Agents (CUA) and Vision-Language-Action (VLA) models. The ability to achieve true test-time scaling in the multi-modal domain suggests a pathway toward more reliable and capable autonomous agents that can reason over extended sequences without losing sight of the visual environment.