Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·25d ago·5 min read
EDCT-Bench: Uncovering Faithfulness Gaps in VLMs via Explanation-Driven Counterfactual Testing
TL;DR
EDCT-Bench introduces Explanation-Driven Counterfactual Testing (EDCT) to identify inconsistencies in Vision-Language Models (VLMs) by testing their responses against modified visual inputs. The findings reveal significant faithfulness gaps, indicating that many VLMs produce explanations that do not align with visual evidence.
✦ Why It Matters
Incorporate EDCT techniques to enhance the reliability of your VLMs in real-world applications.
Key Takeaways
- Implement EDCT to test VLM responses against modified visuals.
- Use EDCT-Bench for benchmarking VLM performance across various domains.
- Identify and address faithfulness gaps in your models.
- Leverage counterfactuals for fine-tuning VLMs to improve accuracy.
How It Works
EDCT operates by extracting visual concepts from a model's explanation, applying minimal edits to these concepts, and then evaluating whether the model's output remains consistent with the altered image. This method allows for a systematic assessment of the model's faithfulness to the visual evidence.