Multimodal-CoT

A two-stage framework that separates rationale generation and answer inference by incorporating vision and language modalities.

Neighborhood — ranked by edge-count

thinker

method

gated fusion
implements
Multimodal fusion technique combining language and vision representations via learnable gating parameters.
Vision Transformer (ViT)
implements
Vision feature extraction model used to extract patch-level features from images in Multimodal-CoT.

claim

framework

Chain-of-Thought (CoT)
extends
A prompting technique that elicits intermediate reasoning steps before final answer inference in language models.
T5
implements
Encoder-decoder architecture used as backbone for Multimodal-CoT; available in Base (200M) and Large (700M) variants.
two-stage separation of rationale generation and answer inference
implements
Architectural design principle that decouples rationale generation (stage 1) from answer inference (stage 2) in Multimodal-CoT.

finding

dataset

A-OKVQA
about
Knowledge-based visual question answering benchmark with 25k questions requiring commonsense and world knowledge.
ScienceQA
about
Large-scale multimodal science question dataset with 21k questions, annotated lectures and explanations across 3 subjects and 26 topics.

artifact

mm-cot
about
Publicly released code repository for Multimodal-CoT implementation.

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

CoT Promptingmethod0.747
Chain-of-thought prompting baseline used for comparison in creative writing and other tasks
VS-CoT and VS-Multi push the Pareto front of the diversity-quality trade-off across creative writing tasksclaim0.734
Empirical finding from creative writing experiments showing VS variants achieve higher diversity without sacrificing quality
Multilayer Perceptronframework0.731
VS-CoTmethod0.717
A VS variant that adds chain-of-thought reasoning before generating the distribution of responses with probabilities
Cross-Modal Samplingmethod0.714
Technique used to demonstrate that the self-prior captures visual–proprioceptive associations by recovering visual appearance from proprioception alone
Multimethod (Generic Function) Dispatchconcept0.705
n-dimensional association model can express binding mechanisms for multimethods by letting values range over methods of arity n and applying appropriate α and β transformations.
CoT Monitormethod0.699
Named method for monitoring chain-of-thought text to detect when the model signals its answer, compared against activation probes
Multi-Turn Promptingmethod0.696
A prompting baseline that elicits N responses across N sequential conversation turns