Automatic factuality scorers now decide which language-model outputs count as "less hallucinated." They rank models, select among sampled answers, and serve as reward signals. These scorers are checked almost exclusively by correlating their outputs with human ratings on ordinary model text. That check says nothing about how a scorer behaves on the outputs an optimizer will push toward. A score that only measures the fraction of correct claims can be raised by padding a biography with trivially true statements, repeating a supported fact, or dropping the hard content. We propose to treat each hallucination type in existing taxonomies as a controlled edit with a known correct effect on the score. For example, "inject a fabricated entity" must not raise the score, and "duplicate a supported sentence" must not raise it either. Each scorer is then run against a suite of such tests, the way axiomatic metric validation has tested diversity and anomaly-detection metrics.
The claim has two parts. First, widely used hallucination scorers violate simple, taxonomy-indexed axioms at measurable rates, and those rates differ sharply between scorers that look equivalent under human correlation. Second, and this part is falsifiable, a scorer's violation profile predicts how badly it gets gamed when used to select outputs, beyond what its human correlation predicts. If the second part fails, the axioms are only of theoretical interest, and we will report that.
Where this sits
The three directions play different roles here. Hallucination taxonomies supply the edits. FAVA defines six span-level types (entity, relation, contradictory, invented, subjective, unverifiable) and injects them synthetically to train a detector-editor (Fine-grained Hallucination Detection and Editing for Language Models). HAD extends this to 11 types split into faithfulness and factuality and synthesizes about 90k injected examples (HAD). FG-PRM injects six reasoning-error types into math chains with over 94% injection success (FG-PRM). All three use injection to create training data for detectors. We reuse the same injection machinery for a different purpose: to build test cases for the scorers themselves.
Axiomatic validation supplies the method and the warning. Two lines of work show that standard metrics routinely violate monotonicity, uniqueness and continuity. In diversity measurement, none of roughly 15 popular measures satisfies all three axioms, and the satisfying ones are NP-hard (Measuring Diversity: Axioms and Challenges). In anomaly detection, Precision, Recall, F1, IoU and related metrics each fail at least four of eight properties (Formally Exploring Visual Anomaly Detection Evaluation Metrics). The 3D-reconstruction work shows these properties can be run as synthetic unit tests rather than proved one metric at a time (Explaining Human Preferences via Metrics for Structured 3D Reconstruction). That is the form we adopt, because LLM-based scorers cannot be analyzed in closed form.
Local inconsistency measures for symbolic knowledge bases show that requiring a measure to satisfy every desirable postulate can force exponential cost (Towards a Principle-based Framework for Assessing the Contribution of Formulas on the Conflicts of Knowledge Bases). This constrains us. We will not aim for a scorer that is perfect on every axiom. We aim for a cheap repair of the specific violations that turn out to matter.
The audit literature imposes a methodological constraint. Benchmark-validity audits can contain "silent no-ops," perturbations that never reach the scorer, which push no-op rates from near zero to 100% and change conclusions (Auditing the Audit). Our test suite must therefore verify that every edit actually changes what the scorer sees.
The closest prior work is FActScore, which splits long-form text into atomic facts and scores the supported fraction with under 2% error against human experts (FActScore). Its authors acknowledge it measures precision only. We do not propose a better FActScore. We propose a way to show, before deployment, which failure modes any scorer of this kind has, and to test whether those failure modes matter in practice.
Core Claim
Let $S(x, y)$ be a scorer's factuality score for output $y$ given prompt $x$, with higher meaning more factual. For each taxonomy-derived edit $e$, we write an axiom as a required sign of the score change $\Delta_e = S(x, e(y)) - S(x, y)$.
Some edits must not raise the score. Injecting an entity, relation, contradictory or invented error requires $\Delta_e \le 0$. Inserting a duplicate or paraphrased copy of a supported claim is a uniqueness axiom and also requires $\Delta_e \le 0$. Appending generic, trivially true claims unrelated to the request is a padding axiom and requires $\Delta_e \le \epsilon$, where $\epsilon$ is a small tolerance.
Other edits must lower the score. Deleting a request-critical fact is an omission axiom and requires $\Delta_e < 0$.
A graded severity axiom requires $\Delta_{\text{contradiction}} \le \Delta_{\text{unverifiable}}$ on the same span. We treat this ordering as a declared, adjustable normative choice rather than a fact, because Abramowitz and Mattei show that evaluation rules derived from values can produce outcomes that violate those very values (Axiomatic Choice and the Decision-Evaluation Paradox).
For each scorer we report two quantities per axiom: the violation rate, and the mean signed size of violations. Reporting magnitude as well as rate responds to the same paper's observation that binary axioms cannot handle graded satisfaction.
Research Program
We build the edit suite from three sources. The first is FActScore biographies, about 6,500 generations across 13 models. The second is FAVABENCH, about 1,000 human-annotated responses. The third is wiki_bio_gpt3_hallucination.
Error edits are produced with the FAVA and HAD injection prompts, then filtered by retrieval against Wikipedia. Duplication and padding edits are produced by rule plus paraphrase. Omission edits remove facts that annotators mark as central to the request. A human spot-check of a stratified sample confirms that each edit has its intended effect.
Following the audit gates of Auditing the Audit, we hash the exact input each scorer receives. Any edit whose scorer input is unchanged, for example because truncation or atomic-fact extraction dropped the edited span, is logged as a no-op and excluded rather than counted as a pass.
We then run every scorer through the suite. FActScore is predicted to fail padding and uniqueness, since it only measures precision. Span-detector scores derived from FAVA and HAD are predicted to be blind to omission. GPT-4-style judges and the temperature ensemble of Metric Ensembles For Hallucination Detection have no prediction in advance, which is exactly why they need testing. NLI and QA consistency metrics are included as weaker references. HaluEval's finding that judges change verdicts when shown contrasting pairs (HaluEval) motivates one additional invariance test: an output's score must not depend on which other output appears in the same prompt.
Finally, we build a minimal repair: FActScore with three additions. Supported claims are deduplicated by entailment. Generic, low-information claims are weighted down using FActScore's own nonparametric prior check. A recall term counts request-critical facts from a reference fact set. This repair is deliberately simple, so that any gain can be traced to specific axioms.
Baselines
The do-nothing reference is the scorer chosen the usual way, by human correlation alone. On wiki_bio_gpt3_hallucination, that is the GPT-4 temperature ensemble, which reports a Pearson correlation of 0.901 with human judgments, against 0.892 for GPT-4 at temperature 0.0 (Metric Ensembles For Hallucination Detection). We also include output length as a trivial scorer, since padding attacks raise length.
The strongest atomic scorer is FActScore, which reports under 2% error against human experts and a score of 58.3% for ChatGPT biographies (FActScore). The strongest span detector is FAVA, which reports 48.1% F1 on FAVABENCH, against 42.5% for GPT-4 (Fine-grained Hallucination Detection and Editing for Language Models). HAD-14B reports 89.10% binary accuracy on HADTest (HAD).
All scorers run on identical outputs, identical retrieval corpora (the Wikipedia snapshots used by FActScore and FAVA), and identical judge-call budgets. The temperature ensemble gets the same total number of judge calls as any repaired judge variant.
The key ablation is the repaired scorer with each addition removed in turn. A second ablation is a rubric-prompted GPT-4 judge whose instructions explicitly state each axiom. This isolates whether the gain comes from the axioms or from our particular implementation of them.
Beating the baselines means two things. The repaired scorer must stay within the confidence interval of the reported human correlations on unoptimized text. It must also degrade significantly less under the selection-pressure test described below.
Why not just…?
Why not just measure correlation with humans, as everyone does? This is the main threat. Correlation is measured on the natural distribution of model outputs, and padding and duplication are rare there. A scorer can therefore reach 0.9 correlation while being trivially exploitable. Whether this matters is an empirical question. The selection-pressure experiment settles it by checking whether violation rates predict gaming after controlling for correlation. If they do not, we will concede the point.
Why not just add recall and deduplication to FActScore and skip the axioms? Some fixes are obvious once a failure is named. But the suite is what names the failures, and it also covers judges and detectors, where no obvious patch exists. If the repaired FActScore wins only because of fixes anyone would have guessed, the ablations will show it. The contribution then narrows to the diagnostic suite and the predictive test.
Why not just prompt a strong judge with a careful rubric, or average it across temperatures? HaluEval shows that judges are sensitive to how inputs are framed. Ensemble gains disappear when weaker members are added. Neither result shows that a rubric judge violates our axioms, though. The rubric-judge ablation tests this directly. If it passes the suite and resists gaming, it becomes the recommended scorer, and that would itself be a useful finding.
Proposed Experiments
The decisive check is selection pressure. For each FActScore biography prompt, we sample N candidate outputs from Alpaca-, Llama2-Chat- and ChatGPT-class generators, with N ranging from 1 to 64. Each scorer picks its best candidate. We then judge the selected outputs with human annotators on a subsample, scoring both precision and coverage of request-critical facts. The quantity of interest is the gap between the true quality of selected outputs and their scorer-predicted quality, as N grows. The hypothesis is that a scorer's padding and uniqueness violation rates explain variance in this gap across scorers that human correlation does not.
A secondary check tests whether axiom violations explain ranking flips across models. We re-rank the 13 FActScore models under the original and repaired scorers and ask whether flipped models are the ones that write long, generic text.
Two further checks probe specific mechanisms. On FAVABENCH, we test whether "unverifiable" spans, which make up over 40% of human-annotated hallucinations, are penalized relative to contradictions in a consistent way across scorers. On HaluEval, we test the paired-input invariance of judges.
Failure modes would be visible. If violation rates do not predict gaming, the predictive claim is refuted. If the repaired scorer loses human correlation, the repair is too blunt.
References
- Fine-grained Hallucination Detection and Editing for Language Models
- HAD: HAllucination Detection Language Models Based on a Comprehensive Hallucination Taxonomy
- FG-PRM: Fine-grained Hallucination Detection and Mitigation in Language Model Mathematical Reasoning
- Measuring Diversity: Axioms and Challenges
- Formally Exploring Visual Anomaly Detection Evaluation Metrics
- Explaining Human Preferences via Metrics for Structured 3D Reconstruction
- Towards a Principle-based Framework for Assessing the Contribution of Formulas on the Conflicts of Knowledge Bases
- Auditing the Audit: Five Failure Modes in Benchmark-Validity Audits
- FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation
- Axiomatic Choice and the Decision-Evaluation Paradox
- Metric Ensembles For Hallucination Detection
- HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models