Crosscheckgpt Eval

crosscheckgpt-eval

CrossCheckGPT: Universal Hallucination Ranking for Multimodal Foundation Models — Sun et al. (2024) (arXiv:2405.13684, 2024)

What this evaluates

This evaluation probes the ability of multimodal foundation models to generate factual content without hallucination across text, image, and audio-visual modalities. It measures how well reference-free ranking methods correlate with human judgments or gold-standard references to rank model outputs by hallucination severity.

Datasets

• WikiBio — total 238; splits: test (238)
• MHaluBench — total 270; splits: test (270)
• AVHalluBench — total ?; splits: test (-1)

Metrics

• System($ ho$) (primary) — range: [-1, 1]
  • Spearman’s Rank Correlation coefficient measuring the correlation between the hallucination ranking produced by the method and the overall ranking (from human annotation or gold-standard references) at the system level.
• Document($r$) / Image($r$) / Video($r$) — range: [-1, 1]
  • Pearson’s Correlation Coefficient measuring the correlation between the method's scores and reference-based scores (RefCheck or human annotations) at the document/image/video level.

Input / output format

Input: Input modality (text passages, images, or audio-visual clips) along with prompts for open-ended generation or description tasks.

Output: Generated text descriptions or passages from the target model, and evidence passages from multiple evidence models.

Scoring recipe

1. Generate outputs from target model and multiple evidence models.
2. Compute CrossCheckGPT scores (explicit or implicit) for each output.
3. Rank models by average hallucination score (lower score = less hallucination).
4. Compute Spearman's rank correlation (System($
ho$)) between the method's ranking and the ground-truth ranking.
5. Compute Pearson's correlation (Document($r$)) between method scores and reference scores at the instance level.

Common pitfalls

• Self-bias occurs when evidence models share the same base model as the target model, inflating scores and leading to contradictory rankings.
• CrossCheck-implicit performance heavily depends on task difficulty; it underperforms on challenging open-ended tasks where CrossCheck-explicit is preferred.
• Generation temperature significantly impacts SelfCheckGPT scores but CrossCheckGPT remains robust; failing to control temperature can skew rankings.

Evidence (verbatim from paper)

We report the system-level correlation between the hallucination ranking methods and the overall ranking measured by Spearman’s Rank Correlation coefficient (SRC), denoted as System($ ho$). In addition, as WikiBio contains reference texts, the references can be used as evidence texts, which can be considered an idealized fact-checking method. This method is referred to as RefCheck, and CrossCheckGPT and SelfCheckGPT scores also are compared against RefCheck at document-level using Pearson’s Correlation Coefficient (PCC), denoted as Document$(r)$.

Citation

@misc{sun2024crosscheckgpt,
  title={CrossCheckGPT: Universal Hallucination Ranking for Multimodal Foundation Models},
  author={Sun et al. (2024)},
  year={2024},
  note={arXiv:2405.13684}
}

• arXiv: 2405.13684