Rlaif V Trustworthiness Eval

rlaif-v-trustworthiness-eval

RLAIF-V: Open-Source AI Feedback Leads to Super GPT-4V Trustworthiness — Tianyu Yu et al. (2024) (arXiv:2405.17220, 2024)

What this evaluates

Evaluates the trustworthiness (hallucination reduction) and helpfulness of multimodal large language models across generative, discriminative, and free-format tasks.

Datasets

• Object HalBench — total ?; splits: (unstated)
• MMHal-Bench — total ?; splits: (unstated)
• MHumanEval — total 146; splits: (unstated)
• AMBER — total ?; splits: (unstated)
• RefoMB — total 360; splits: dev (99), test (261)
• MMStar — total 1500; splits: (unstated)

Metrics

• response-level hallucination rate (primary) — range: percent
  • Percentage of model responses that contain at least one hallucinated object or claim.
• mention-level hallucination rate — range: percent
  • Percentage of individual object mentions across all responses that are hallucinated.
• trustworthiness win rate (primary) — range: percent
  • Win rate calculated by comparing the model response with a GPT-4V reference response regarding trustworthiness.
• overall win rate — range: percent
  • Win rate based on an evaluation review comparing the model response with a GPT-4V reference response regarding both trustworthiness and helpfulness.
• accuracy — range: [0, 1]
  • Standard classification accuracy on the discriminative part of the AMBER benchmark.
• F1 — range: [0, 1]
  • F1 score on the discriminative part of the AMBER benchmark.
• overall score — range: other
  • Aggregate score on the MMStar benchmark covering 6 core capabilities and 18 detailed axes.

Input / output format

Input: Image and text instruction/prompt pairs.

Output: Text response generated by the MLLM.

Scoring recipe

def evaluate(predictions, gold, gpt4v_refs=None):
    resp_hall = sum(1 for p in predictions if is_hallucinated(p)) / len(predictions)
    men_hall = count_hallucinated_mentions(predictions) / count_total_mentions(gold)
    acc = accuracy_score(predictions, gold)
    f1 = f1_score(predictions, gold)
    if gpt4v_refs:
        trust_wins = sum(1 for p, r in zip(predictions, gpt4v_refs) if p_trustworthy(p, r))
        overall_wins = sum(1 for p, r in zip(predictions, gpt4v_refs) if p_overall(p, r))
        return resp_hall, men_hall, acc, f1, trust_wins/len(predictions), overall_wins/len(predictions)
    return resp_hall, men_hall, acc, f1

Common pitfalls

• RefoMB results in the main table use the dev split (99 instructions) to save evaluation costs, while the test split (261 instructions) is only reported in the appendix.
• Best-of-N (BoN) results are marked N/A for multi-choice and yes-no questions because these tasks only require a single token output, making response-level comparison infeasible.
• Hallucination rates are reported at two granularities (response-level vs. mention-level), which can be easily confused when comparing results across papers.

Evidence (verbatim from paper)

We evaluate models from two perspectives, including trustworthiness reflecting the hallucination degree, and helpfulness reflecting the general capability. For trustworthiness, we perform evaluation on five benchmarks: (1) Object HalBench is a widely adopted benchmark for assessing common object hallucination in detailed image descriptions. We follow[[66]] to use 8 diverse prompts to improve the evaluation stability. We report the response-level hallucination rate (i.e., the percentage of hallucinated responses) and the mention-level hallucination rate (i.e., the percentage of hallucinated objects).

Citation

@misc{yu2024rlaifv,
  title={RLAIF-V: Open-Source AI Feedback Leads to Super GPT-4V Trustworthiness},
  author={Tianyu Yu et al. (2024)},
  year={2024},
  note={arXiv:2405.17220}
}

• arXiv: 2405.17220