IC-1408OpenFlamingo and Idefics models generate low-quality explanations in zero-shot, but ICL and model scale significantly improve explanation CIDEr

Mustafa Shukor, Alexandre Rame, Corentin Dancette, Matthieu Cord

SourceBeyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning

On VQA-X, LMMs are given an image, question, and correct answer, then asked to generate an explanation. Zero-shot CIDEr scores are significantly below SOTA: OFv2-9b scores 56.17 and Idefics-9b scores 67.43. ICL with more demonstrations substantially improves CIDEr: OFv2-9b rises to 80.51 at 32-shot, Idefics-9b to 90.18. There is a clear positive correlation between model size and explanation quality. Idefics-80b surpasses SOTA with ICL, but instruction tuning does not help the 80B model, possibly due to overfitting when training the LLM.

Evidence
correlational
Key metric
OFv2-9b CIDEr: 56.17 (0-shot), 61.43 (4-shot), 74.71 (8-shot), 80.41 (16-shot), 80.51 (32-shot); Idefics-9b CIDEr: 67.43 (0-shot), 80.13 (4-shot), 83.45 (8-shot), 88.59 (16-shot), 90.18 (32-shot)
Caveat
The paper notes that for 80B-parameter models, instruction tuning does not improve explanation quality, which might be due to overfitting when training the LLM. CIDEr measures syntactic similarity to human annotations, not semantic faithfulness of the explanation.
Model
OpenFlamingo, Idefics
Concepts
Explanation faithfulness, Scale-dependent behaviour
Datasets
VQA-X [eval]
Methods
In-Context Learning / In-context learning prompt [primary], CIDER (Compactness and Dispersion for OOD detection) [eval]
Related work
OpenFlamingo [context], Idefics [context]
Related findings
IC-1405, IC-1406, IC-1407
Extraction
automatic-extraction