IReasoningVLM<T>

Interfaces · AiDotNet.VisionLanguage.Interfaces

Interface for vision-language models with chain-of-thought reasoning capabilities.

Implements: IInstructionTunedVLM<T>

Known implementations: KimiVLThinking<T>, KimiVL<T>, LLaVACoT<T>, QVQ72B<T>, SkyworkR1V2<T>, SkyworkR1V<T>

How It Works

Reasoning VLMs extend instruction-tuned VLMs with explicit thinking/reasoning steps before producing a final answer. They are trained with reinforcement learning and/or chain-of-thought supervision to decompose complex visual reasoning tasks. Architectures include:

  • QVQ/Kimi-VL: MoE language models with visual reasoning alignment
  • Skywork R1V: Cross-modal transfer of text reasoning to vision
  • LLaVA-CoT: Chain-of-thought fine-tuning on visual instruction data

Properties

Property Summary
string ReasoningApproach { get; } Gets the name of the reasoning approach (e.g., “CoT”, “RL-Aligned”, “MoE-Reasoning”).

Methods

Method Summary
Tensor<T> ReasonWithChainOfThought(Tensor<T> image, string question) Generates a response with explicit chain-of-thought reasoning steps.