The Design Logic of Hallucination Rate Metrics in LLM Evaluation and Its Impact on Rankings

1. Starting From a Ranking Correction: Evaluation Indices Are Not “Objective Truth”

The large language model (LLM) community recently experienced a representative event: a third-party evaluation agency revised its comprehensive intelligence index by removing the “non-hallucination rate” from a knowledge-based sub-benchmark and republishing the model rankings accordingly. The ranking changes were immediate, sparking heated community discussions.

The true value of this event lies not in determining which model rose or fell in the rankings, but in how it intuitively reveals an often-overlooked fact: a comprehensive intelligence index is essentially a weighted aggregate of various sub-metrics. As long as the definitions, statistical scopes, or weights of these sub-metrics change, rankings can undergo massive reshuffling. Understanding this premise is essential for rationally utilizing any evaluation leaderboard.

2. The Non-Hallucination Rate Metric: Definition, Calculation, and Design Intent

To understand the controversy, we must first clarify what this metric actually measures.

2.1 What is a Hallucination?

A hallucination refers to a model generating content that is linguistically fluent and confidently toned, yet factually incorrect or unverifiable. Typical manifestations include fabricating non-existent literature citations, providing incorrect dates and data, and giving definitive answers to uncertain questions. The root cause lies in the probabilistic nature of generative models: they predict the text that “most resembles an answer” rather than text that is “verified to be true.”

2.2 Common Calculation Methods for Non-Hallucination Rate

In knowledge-based evaluations, a common industry approach is to calculate the proportion of factually correct answers among the questions the model chooses to answer. Questions where the model abstains or states “uncertain” are not counted as errors, nor are they included in the numerator. This can be formalized as:

Non-Hallucination Rate = Correct Answers / (Correct Answers + Incorrect Answers)

A simplified numerical example illustrates its incentive effect:

• Model A: Answers all 100 questions, gets 85 correct and 15 wrong. Non-hallucination rate = 85/100 = 85%.
• Model B: Abstains on 40 uncertain questions, answers only 60, gets 55 correct and 5 wrong. Non-hallucination rate = 55/60 ≈ 91.7%.

Even if Model B’s actual knowledge coverage is narrower, as long as it is “cautious enough,” its metric performance is better. This is the exact design intent of the metric: to reward the behavior of “preferring to abstain rather than answer incorrectly” and to encourage models to be honest about their uncertainties.

2.3 Why Do Some Models Score Extremely Low on This Metric?

According to public observations from the aforementioned evaluation, certain leading model series known for their “proactive answering” exhibited hallucination rates as high as 85%–92% on high-difficulty knowledge sub-benchmarks. This does not necessarily mean these models have “worse knowledge”; it primarily reflects their behavioral tendencies. After undergoing extensive preference tuning aimed at being “responsive and detailed,” these models tend to provide definitive answers to almost all questions, even those with high uncertainty. Under the calculation method where “abstaining is not penalized, but wrong answers are,” the point loss for such models is significantly amplified.

3. The Double-Edged Sword: The Tension Between Honesty and Usefulness

Looking solely at its design intent, the non-hallucination rate seems flawless. However, making it a high-weight sub-metric in a comprehensive “intelligence” index brings problems to the surface.

1. Susceptible to Strategic Gaming
In extreme cases, a model that abstains from almost all difficult questions and only answers common-sense questions could score near-perfect on this metric, yet its actual usefulness approaches zero. The boundary between what the metric measures—“honesty” or “退缩” (retreating/shrinking back)—becomes blurred.

2. Inherently Unfriendly to Proactive Models
A proactive answering tendency is actually an advantage in many real-world business scenarios. In customer service, content creation, and code completion, users expect high-quality output rather than frequent “I don’t know” responses. Penalizing proactive answering with a single metric standard is like measuring all body types with a single ruler.

3. Another Instance of Goodhart’s Law
When a measure becomes a target, it ceases to be a good measure. If model teams conduct targeted training around this metric—learning to abstain more on evaluation sets—the metric score rises, but true reliability may not improve. This is the latest case of Goodhart’s Law in the LLM domain, a recurring theme in evaluation methodology.

4. Post-Removal Reshuffling: A Textbook Demonstration of “Benchmark Sensitivity”

The republished rankings after the agency removed the non-hallucination rate essentially served as a public sensitivity analysis:

• With the index composition changed, the evaluation focus shifted to dimensions like reasoning, coding, math, and accurate knowledge answering. Models that “dared to answer and answered correctly” saw their scores rise.
• Model series that answer proactively, previously dragged down by this metric, saw their rankings rebound, while models relying on cautious strategies to gain points dropped.

Which of the two rankings is correct? Actually, both are; they simply answer different questions. The index including the non-hallucination rate answers “which model is more honest and trustworthy for high-risk tasks.” The index excluding it answers “which model has stronger comprehensive capabilities (regardless of whether it is willing to answer).” The developer’s task is not to argue which is more “correct,” but to identify the implicit value orientation behind each leaderboard and make trade-offs based on their own business requirements.

5. Developer Insights: Treat Leaderboards as References, Scenarios as the Yardstick

1. Define Business Fault Tolerance Before Selecting a Model
In high-risk scenarios such as healthcare, finance, and law, factual errors carry extremely high costs. You should prioritize examining the model’s factual accuracy, reasonableness of abstention, and answer traceability. In scenarios like marketing copywriting or brainstorming, you should focus more on generation quality, style control, and response speed. The same leaderboard offers vastly different reference values for different scenarios.

2. Build a Custom Evaluation Set and Conduct Regular Regressions
Third-party leaderboards use public question banks, which inherently carry the risk of being “trained upon.” It is recommended to sample from real business traffic to build a private evaluation set covering typical questions, boundary questions, and misleading questions, and perform regression testing after every model version update or prompt adjustment. A simple scenario-based weighted scoring approach looks like this:

# Illustrative code: Custom weighted scoring by business scenario, rather than directly applying third-party comprehensive indices
WEIGHTS = {
    "factual_accuracy": 0.45,  # Factual accuracy (higher weight for high-risk scenarios)
    "abstain_quality":  0.20,  # Abstention reasonableness: abstain when necessary, answer when appropriate
    "task_quality":     0.25,  # Task completion quality
    "latency":          0.10,  # Response latency
}
def scenario_score(dim_scores: dict) -> float:
    return sum(dim_scores[k] * w for k, w in WEIGHTS.items())

3. Look at Granular Dimensions, Not Just the Comprehensive Score
Comprehensive scores compress too much information. Examining the scores of each sub-dimension and their variation trends across different versions is much more valuable than memorizing a ranking position.

4. Focus on Calibration, Not Just Accuracy
The ability to “know what you don’t know” (calibration) is often more important in high-risk applications than single-point accuracy. You can require the model to output a confidence level for its answers and then plot a reliability curve for evaluation.

6. Application Layer Guardrails: Engineer Hallucination Mitigation Regardless of Model Choice

Evaluation solves the problem of “who to choose,” but hallucination governance requires continuous investment at the application layer. The following are widely validated engineering practices:

1. Retrieval Augmentation and Answer Traceability
Anchor generated content on retrieved evidence. Require the model to answer solely based on provided materials and cite sources, distinctly marking statements unsupported by evidence.

2. Tiered Response Strategy
Design differentiated responses based on confidence and evidence coverage, rather than using a “one-size-fits-all” approach:

# Illustrative code: Tiered response strategy based on confidence and evidence
def respond(query, confidence, evidence):
    if not evidence:
        return "The knowledge base does not currently cover this question. Please provide more background information or escalate to human support."
    if confidence >= 0.8:
        return generate(query, evidence, cite=True)
    if confidence >= 0.5:
        # Answer but explicitly state uncertainty, distinguishing from "silent hallucinations"
        return generate(query, evidence, cite=True, disclaimer=True)
    return "The confidence level for this question is insufficient. The following related information is for reference only: " + summarize(evidence)

3. Post-Generation Verification
Extract claims at the assertion level (claim-level) from the generated results, compare them individually against the knowledge base, and intercept statements conflicting with the evidence before outputting them to the user.

4. Interaction Design Guardrails
Explicitly present citations and uncertainty prompts in the UI. Guide users to supplement key information, transforming rigid “abstentions” into “clarifying follow-up questions,” thereby balancing honesty and usefulness.

Whether the model comes from open-source deployment or cloud platform services like Baidu AI Cloud, the above methods apply equally—hallucination governance is never a single-point issue, but a systematic engineering effort combining “model capability + retrieval quality + response strategy + interaction design.”

7. Conclusion

The addition or removal of a single metric can turn a leaderboard upside down, reminding us that evaluation benchmarks are not just technical tools; they carry the designer’s value judgment on “what makes a good model.” Honesty and usefulness are inherently a pair of goals requiring dynamic balance, and any single metric can only capture one facet. For developers, the real moat is not chasing a leaderboard ranking for a specific cycle, but establishing their own evaluation system, understanding the incentive directions behind each metric, and building an engineering defense line against hallucinations at the application layer. Leaderboards will change; methodology will not.