Evolution of Multimodal Reasoning

1. GRIT: Teaching Models to Think with Images

The Core Problem

Traditional multimodal models primarily use images as input, generating only text responses. This raises a critical question: Are these models truly performing visual reasoning, or are they merely describing images verbally?

Technical Approach

1. Three-Stage Training Pipeline

• Stage 1 - Grounding: Learn to identify visual evidence (40K examples)
• Stage 2 - Thinking: Develop reasoning with grounded evidence (20K examples)
• Stage 3 - Integration: Combine both capabilities

2. Data Synthesis Strategy

Rather than manual annotation, GRIT uses GPT-4V to generate training data by:

1. Solving problems independently
2. Analyzing their reasoning process
3. Identifying which visual regions support each reasoning step
4. Converting image coordinates to bounding boxes

Results and Validation

On spatial reasoning benchmarks, GRIT-enhanced models show substantial improvements:

• VSR: 62.5% → 69.4% accuracy
• BLINK: 47.3% → 52.4% accuracy
• CV-Bench: Consistent gains across perception tasks

2. Video Models as Zero-Shot Learners

The Paradigm Shift

This paper proposes a radical idea: video generation models (like Veo 3) can function as zero-shot visual reasoners without any task-specific training.

Key Experiments

1. Visual Question Answering

Rather than answering "What color is the car?" with text, the model generates a video showing the car's color through visual transformation.

2. Future Prediction

Given initial frames, the model generates plausible continuations, demonstrating understanding of physics and dynamics.

3. Visual Reasoning

For problems requiring inference (e.g., "Which way will the object fall?"), the model generates video sequences showing the outcome.

Limitations and Challenges

• Computational Cost: Video generation is significantly more expensive than text generation
• Evaluation Ambiguity: Multiple valid videos can answer the same question
• Fine-Grained Control: Difficult to specify exact reasoning paths

3. ThinkMorph: Emergent Properties in Interleaved Reasoning

The Key Question

How should text and images be combined during reasoning? Should images simply illustrate textual reasoning, or should they provide distinct, complementary information?

Data Evolution Process

ThinkMorph introduces a systematic approach to creating high-quality training data:

Stage 1: Seed Generation

Human-designed prompts → GPT-4 generates text-only reasoning chains

Stage 2: Visual Injection

GPT-4V identifies where visual reasoning would be beneficial and inserts image generation prompts at strategic points

Stage 3: Image Generation

DALL-E 3 generates images that complement textual reasoning

Stage 4: Filtering

Models are trained incrementally, and data that improves performance is retained for the next iteration

Experimental Results

ThinkMorph models show superior performance on:

• Geometry: 12% improvement over text-only reasoning
• Graph Theory: 18% improvement through visual graph representations
• Complex Problem Solving: Sustained accuracy on multi-step problems

4. V-Thinker: Interactive Visual Thinking

The Critical Distinction

Previous approaches generated images as part of reasoning chains, but couldn't interact with them. V-Thinker introduces fully interactive visual thinking through end-to-end reinforcement learning.

Technical Architecture

1. Two-Module System

• Thought Generator: Language model decides when and what images to generate
• Image Generator: Stable Diffusion XL creates visual content from text prompts

2. End-to-End Reinforcement Learning

Unlike supervised learning, V-Thinker uses Direct Policy Optimization (DPO) where:

• Reward: Correctness of final answer
• Positive Examples: Trajectories leading to correct answers
• Negative Examples: Trajectories leading to incorrect answers

Progressive Training Curriculum

V-Thinker uses a carefully designed progression:

1. Stage 1: Simple visual reasoning (geometry basics)
2. Stage 2: Multi-step problems (spatial transformations)
3. Stage 3: Complex integration (combining multiple visual concepts)

Benchmark: VSTaR

The paper introduces VSTaR (Visual Self-Teaching through Reinforcement), evaluating:

• Visual Necessity: Can the problem be solved without images?
• Reasoning Depth: Number of visual steps required
• Feedback Utilization: Does the model adjust based on generated images?

Results

• Geometry: 23% improvement over text-only models
• Multi-step Reasoning: 35% improvement on complex problems
• Efficiency: Fewer reasoning steps to reach correct answers

5. Thinking with Video: A Unified Paradigm

The Unifying Framework

This paper synthesizes previous insights into a single paradigm: video generation as the natural medium for multimodal reasoning.

Vision-Centric vs. Text-Centric Reasoning

Vision-Centric Tasks

Example: "What will happen if this ball rolls down the ramp?"

• Generate video showing the physical simulation
• Visual representation is the reasoning process
• Answer is implicit in the video content

Text-Centric Tasks

Example: "Solve: 3x + 5 = 14"

• Generate video showing step-by-step algebraic manipulation
• Visual representation supplements textual reasoning
• Each frame shows mathematical transformations

Technical Implementation

1. Training Data Construction

• 40K video sequences with reasoning annotations
• Mix of vision-centric (physics, spatial) and text-centric (math, logic) tasks
• Verifiable ground truth for objective evaluation

2. Model Architecture

• Base: Diffusion-based video generator (CogVideoX)
• Enhancement: Added reasoning-aware conditioning
• Output: 4-8 frame sequences showing reasoning progression

Experimental Validation

The paper introduces comprehensive benchmarks covering:

• Physics Reasoning: Predicting motion and collisions
• Spatial Reasoning: Object relationships and transformations
• Mathematical Reasoning: Visual representation of equations
• Logical Reasoning: Diagram-based problem solving

Results Across Modalities

• Vision-Centric: 45% improvement over text-only approaches
• Text-Centric: 12% improvement through visual scaffolding
• Cross-Domain: Single model handles diverse reasoning types

6. DeepEyesV2: Toward Agentic Multimodal Intelligence

The Final Frontier: Agentic AI

DeepEyesV2 represents the culmination of this research trajectory: a system that autonomously integrates multiple tools within its reasoning process.

Multi-Tool Integration

Available Tools

1. Code Execution: Python interpreter for calculations, data analysis
2. Web Search: Real-time information retrieval
3. Image Operations: Cropping, filtering, transformations

Integration Challenge

How do you train a model to use multiple tools effectively without providing exhaustive examples of every possible combination?

Two-Phase Training

Phase 1: Cold-Start SFT

• Data: 40K expert trajectories showing correct tool usage
• Goal: Teach the model that tools exist and how to invoke them
• Coverage: Basic patterns for each tool individually

Phase 2: Reinforcement Learning

• Reward: Simple correctness signal (binary: right/wrong answer)
• Discovery: Model learns when and which tools to use
• Emergence: Novel tool combinations not in training data

Emergent Agentic Behaviors

1. Task-Adaptive Tool Invocation

• Perception tasks → Image operations (e.g., cropping)
• Reasoning tasks → Numerical analysis

2. Complex Tool Combinations

RL enables spontaneous behaviors not present in training data, such as:

• Using web search to find current information
• Executing code to analyze the retrieved data
• Generating visualizations of the analysis

3. Context-Aware Decision Making

Model learns to selectively invoke tools based on problem context, reflecting autonomous, agentic reasoning.

Benchmark: MA-Eval

The paper introduces MA-Eval (Multi-tool Agentic Evaluation) measuring:

• Tool Selection Accuracy: Does the model choose appropriate tools?
• Combination Efficiency: Are tool combinations minimal and effective?
• Error Recovery: Can the model adapt when initial tool use fails?
• Final Performance: Overall task success rate

Results

• Multi-Step Reasoning: 34% improvement over single-tool models
• Complex Integration: Successfully combines 3+ tools in 68% of appropriate cases
• Emergent Behaviors: 23% of successful strategies were not in training data

How These Papers Connect

The Research Trajectory

Common Themes

• Data Efficiency: All approaches emphasize learning from limited high-quality data
• Reinforcement Learning: Most use RL to refine tool use and reasoning behaviors
• Emergent Properties: Complex behaviors emerge from relatively simple training setups
• Tool Integration: Progressive movement toward integrated, multi-tool systems
• Evaluation Innovation: Each introduces new benchmarks to measure novel capabilities

Key Takeaways for Practitioners

1. Choose the Right Paradigm

• For static visual reasoning: Start with grounded reasoning (GRIT-style)
• For dynamic processes: Consider video generation approaches
• For complex tool integration: Build agentic systems (DeepEyesV2-style)

2. Training Strategy Matters

Recommended pipeline:

1. Cold-start SFT: Establish basic tool-use patterns with high-quality trajectories
2. Reinforcement Learning: Refine and enhance behaviors with simple rewards

3. Data Quality Over Quantity

All papers demonstrate success with relatively small datasets (20-40K samples), emphasizing:

• Appropriate difficulty (not too easy, not impossible)
• Diversity across tasks and visual distributions
• Verifiable formats for objective evaluation
• Evidence that tool use improves performance

4. Complementary, Not Redundant

When designing multimodal systems, ensure text and visual reasoning provide different, complementary information rather than describing the same content in different modalities.

5. Evaluation Must Evolve

Traditional benchmarks measuring single capabilities are insufficient. New systems require:

• Cross-capability integration tests
• Real-world scenarios combining perception, search, and reasoning
• Evaluation of tool-use effectiveness, not just final answers

Future Directions

Open Challenges

• Generalization: Models trained with limited data still struggle with out-of-distribution scenarios
• Tool Reliability: Preventing reward hacking and ensuring consistent, meaningful tool use
• Computational Cost: Video generation and multi-tool reasoning are expensive
• Safety and Alignment: Agentic models with tool access raise new safety concerns

Promising Research Directions

• Better Rewards: Designing reward functions that encourage genuine reasoning without hacking
• Efficiency: Reducing computational requirements for video-based reasoning
• Tool Expansion: Integrating more diverse tools (e.g., 3D rendering, simulation)
• Test-Time Scaling: Exploring self-consistency and ensemble methods in multimodal settings
• Human-AI Collaboration: Designing interfaces for interactive multimodal reasoning