RECALL_counterfactual-robustness.md

RECALL: Counterfactual Robustness

• Dimension: Generated Answer <-> GroundTruth Answer
• Reference: RECALL: A Benchmark for LLMs Robustness against External Counterfactual Knowledge
• Type: Exact Match

Counterfactual Robustness evaluates model's performance across two critical aspects.

• Response Quality examines whether models can continue to generate high-quality responses even when faced with interfering counterfactual information.
• Robustness Aspect measures the models' ability to resist incorporating counterfactual information from the provided contexts into their responses.

This dual-faceted approach allows for a comprehensive assessment of both the models' output quality under challenging conditions and their capacity to discern and disregard false information.

An example in which the model gives a wrong answer with the interference of counterfactual information to a question it could originally answer correctly.

Metric: Misleading Rate

Misleading Rate is the proportion of the queries that the model answers wrongly with edited contexts in all queries that the model can answer correctly without external knowledge.

Metric: Mistake Reappearance Rate

Mistake Reappearance Rate is the proportion of edited words appearing in the model's outputs in all edited words.