NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
How Many Counterfactuals Does It Take? Probing VLM Hallucinations Through Circuits and Causal Effects
TL;DR
Visual Language Models (VLMs) often generate hallucinated outputs that lack visual grounding, creating a gap in understanding their reliability. A causal influence metric was developed to assess the robustness of these hallucinations under counterfactual perturbations.
✦ Why It Matters
Engineers can use this causal influence metric to better evaluate and enhance the robustness of VLMs against hallucinations.
Key Takeaways
- Utilize counterfactual examples to test VLM predictions.
- Apply causal influence metrics to assess model stability.
- Identify components responsible for hallucinations in VLMs.
- Determine the necessary sample size for reliable evaluations.
How It Works
The study defines a causal influence metric that measures the log-probability differences between factual and counterfactual outputs. This metric helps in understanding how changes in input affect the model's predictions, particularly focusing on hallucinations.
By analyzing the activation of specific model components through circuit discovery techniques, the authors can pinpoint which parts of the model contribute to unreliable outputs.