DeltaV: Thinking with Visual State Updates in Unified Large Multimodal Models

Problem Statement

Current Unified Large Multimodal Models (ULMMs) attempting interleaved multimodal reasoning typically generate intermediate visual states as full images. This paradigm introduces two critical inefficiencies:

1. Visual-Token Redundancy: Consecutive reasoning steps often share substantial visual content. Generating a full image for every step repeatedly models unchanged regions, leading to a massive overhead in visual tokens compared to textual reasoning steps.
2. Diluted Supervision: By reconstructing the entire state at every step, the learning signal is spread across both persistent content and the sparse, critical changes driving the reasoning. This biases optimization toward low-level reconstruction rather than the reasoning-critical state transitions.

Existing approaches either compress visual information into latent tokens (sacrificing interpretability) or rely on tool-augmented pipelines (lacking end-to-end coupling). Even within unified models, the computational cost of pixel-level generation often dominates semantic reasoning, restricting scalability.

Methodology: DeltaV

The authors propose DeltaV, a ULMM framework that reframes interleaved reasoning as a process of visual state transitions rather than independent full-image generations. Instead of predicting $Z_t$ (the full state) at step $t$, DeltaV predicts $\Delta Z_t$ (visual updates) conditioned on historical states.

1. Unified Autoregressive Modeling with Visual Updates
DeltaV decomposes the reasoning trajectory $Y$ into an initial visual state $Z_0$ and a sequence of variable-length visual update tokens $\Delta Z_t$:
$$Y = \{Z_0, X_1, \Delta Z_1, X_2, \Delta Z_2, \dots\}$$
where $X_t$ represents text tokens. This formulation shifts the learning target from state reconstruction to transition modeling, forcing the model to focus only on the visual evidence required to advance the reasoning trajectory.

2. TSIM-Tok: Variation-Aware Tokenization
To implement this, DeltaV introduces TSIM-Tok, a visual tokenizer comprising two key components:

• TSIM Router (Temporal Similarity Router): This module dynamically allocates the token budget for each update based on the magnitude of visual change. It calculates a Temporal Similarity (TSIM) score by comparing the current visual features against a weighted history of previous states. The weight considers temporal proximity (recent states matter more) and information capacity (states with more tokens provide richer references).
  • The router uses an offline calibration procedure to map TSIM scores to token budgets. It fits a curve relating token count to reconstruction fidelity (SSIM) and selects the budget where the marginal gain in reconstruction quality falls below a threshold. This ensures complex changes receive more tokens, while minor updates receive fewer.
• Visual Update Representation Module: This module encodes visual states into base tokens ($Z_0$) and variable-length update tokens ($\Delta Z_t$) using learnable queries. It conditions the generation of $\Delta Z_t$ on $Z_0$ and all preceding updates ($\Delta Z_{<t}$), ensuring temporal consistency without re-generating stable content.

3. StructCoT Dataset
Recognizing that existing interleaved reasoning datasets are limited in scale and domain diversity, the authors constructed StructCoT.

• Scale: 1.05 million samples.
• Coverage: 44 task domains, organized into seven reasoning categories: Strategic Planning, Spatial/Embodied Planning, Logical/Abstract Reasoning, Mathematical/Algorithmic Reasoning, Scientific/Domain Reasoning, Visual Search, and Image Restoration/Jigsaw.
• Purpose: To provide broad supervision for learning generalizable visual update patterns and reasoning structures.

Key Contributions

1. DeltaV Framework: The first unified multimodal reasoning framework that explicitly formulates interleaved reasoning as update-centric visual state modeling. It replaces full-image generation with compact, variation-aware visual updates.
2. TSIM Router: A novel mechanism that aligns token allocation with the actual magnitude of visual change, reducing redundancy by stopping token allocation once marginal reconstruction gains diminish.
3. StructCoT Dataset: A large-scale, diverse dataset designed to support generalizable update-centric multimodal reasoning, significantly expanding the coverage of existing benchmarks.

Experimental Results

Experiments were conducted on the Zebra-CoT dataset (for ablation and in-domain evaluation) and external benchmarks (for generalization).

• Efficiency: The visual-update paradigm reduces newly generated visual tokens by 55.6% on average compared to full-image generation, without compromising reconstruction fidelity.
• Reasoning Performance:
  • DeltaV improves multimodal reasoning by 3.3% over full-image generation baselines.
  • On in-domain evaluations (Zebra-CoT and StructCoT), the DeltaV-2B model outperforms substantially larger open-source models by an average of 8.4%.
  • On external multimodal reasoning and understanding benchmarks, DeltaV-2B surpasses the comparable-scale Qwen3-VL-2B by 5.9%.
• Ablation Studies:
  • Token Allocation: The TSIM Router (Slope-based) outperforms both random budget allocation and value-based thresholding, confirming that marginal utility is a superior criterion for budgeting.
  • Reconstruction Consistency: Qualitative analysis shows that full-image models suffer from "semantic drift" (e.g., identity or shape changes across steps), whereas DeltaV maintains consistent visual states by grounding updates on historical information.
  • Generalization: Visual-update training improves performance on general understanding benchmarks (e.g., MMBench, MathVista) compared to text-only baselines, suggesting the model learns transferable multimodal representations.

Significance and Claims

The paper claims that the bottleneck in current interleaved reasoning is not pixel-level generation itself, but the paradigm of treating every reasoning step as a full-image generation task. By shifting to a visual state transition perspective, DeltaV achieves:

• Higher Efficiency: Drastically reducing token overhead by modeling only changes.
• Better Reasoning Signals: Concentrating supervision on the sparse, critical changes that drive reasoning, rather than redundant content.
• Improved Consistency: Mitigating error accumulation and semantic drift across multi-step reasoning chains.

The authors conclude that DeltaV offers a more efficient and generalizable path toward unified multimodal intelligence, suggesting that future systems should adaptively select intermediate representations (compact updates vs. full images vs. external tools) based on task requirements.