---
title: 'PhantomBench: Evaluating LLM Non-Existence'
url: https://www.emergentmind.com/topics/phantombench
type: topic
---

# PhantomBench: Evaluating LLM Non-Existence

Searching arXiv for the exact PhantomBench and PhantomWiki papers to ground the article with fresh citations.
PhantomBench is a benchmark and data generation pipeline for evaluating whether language models recognize that a queried concept does not exist and abstain rather than hallucinating plausible but ungrounded content. In its primary sense, the term denotes the benchmark introduced in "PhantomBench: Benchmarking the Non-existential Threat of Language Models" [2606.11105], which comprises more than 60K non-existent terms and entities derived from real concepts across diverse domains. In a separate, narrower usage within "PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation" [2502.20377], "PhantomBench" denotes a possible benchmark configuration built from PhantomWiki’s on-demand synthetic-universe generator rather than a fixed benchmark artifact. The dominant contemporary usage, however, is the hallucination benchmark centered on non-existent concepts [2606.11105].

## 1. Definition, scope, and motivation

PhantomBench is designed to test a specific capability of language models: recognizing the limits of their knowledge when presented with linguistically plausible but non-existent concepts, and abstaining instead of fabricating content [2606.11105]. The benchmark targets both non-existent **terms** and non-existent **entities**, and treats any informative factual answer about such a concept as hallucination by definition, because the queried concept has been constructed to have zero attestation in a massive web-scale corpus [2606.11105].

The benchmark addresses a failure mode that conventional hallucination benchmarks do not isolate cleanly. Standard factuality evaluations typically depend on existing entities, known references, or explicit ground truth. That setup is less suitable for probing knowledge boundaries, especially for rare or highly specialized concepts. PhantomBench reframes the problem: if the concept does not exist, the only fully correct behavior is abstention. This makes evaluation operationally simple and semantically sharp, because the positive class does not require adjudicating among multiple partially correct free-form answers [2606.11105].

The paper formulates several research questions around this setup: abstention reliability across models and domains, the effect of prompt presupposition, the influence of model size and reasoning fine-tuning, the role of domain specialization, and whether behavior on non-existent concepts serves as a proxy for behavior on rare real concepts [2606.11105]. A central claim is that non-existent concepts provide a controlled analogue of rare or unknown concepts, and that this analogue is useful precisely because hallucination is definitionally identifiable [2606.11105].

A distinct but related usage appears in PhantomWiki. There, the authors describe PhantomWiki as an on-demand generator of synthetic Wikipedia-style corpora and logically grounded QA pairs, and note that something called "PhantomBench" would effectively be "PhantomWiki with a particular configuration (universe size, question difficulty, task format, etc.) and a fixed evaluation protocol" [2502.20377]. This suggests two different benchmark-design philosophies associated with the same name: one centered on non-existence and abstention [2606.11105], and one centered on configurable reasoning and retrieval evaluation over synthetic corpora [2502.20377].

## 2. Benchmark construction and generation pipeline

PhantomBench is built through a multi-step pipeline: generate candidate non-existent concepts from real concepts, filter them against a massive corpus to ensure non-existence, create prompts targeting different properties, and evaluate responses with an LLM judge [2606.11105]. The pipeline is explicitly intended to be reusable, so that researchers can construct domain-specific variants rather than relying only on a fixed public split [2606.11105].

The benchmark distinguishes between **terms** and **entities**, because they require different generation strategies [2606.11105]. Seed concepts come from 17 datasets spanning medicine, science, law, and several categories of named entities such as festivals, conferences, holidays, sport events, elections, social issues, natural disasters, accidents, historical events, and PopQA "Creative Work / Place" with person entities removed [2606.11105].

For non-existent **terms**, the pipeline first constructs blended words by combining parts of existing words while preserving plausible affixes. Affix detection extracts prefixes and suffixes, excludes single-character ones, retains those appearing more than 3 times, and augments them with English prefix/suffix lists from Wiktionary. Each word is split at the position of the longest matching prefix or suffix, and blended words are formed by combining the first segment of one word with the last segment of another; the paper gives "enteric" + "macromolecule" $\rightarrow$ "entermolecule" as an example [2606.11105]. The system then synthesizes new terms by replacing half of the words in an existing seed term with randomly sampled words from the union of existing and blended vocabularies, preserving the overall compositional shape of the source term [2606.11105].

For non-existent **entities**, the pipeline decomposes names into a structural pattern and a lexical core. It extracts frequent bigrams and trigrams as productive patterns, filters them according to frequency and token constraints, then gathers lower-frequency lexical items under separate thresholds for single-word and multi-word candidates [2606.11105]. New entities are generated by attaching lexical items to structural patterns at plausible boundary positions, such as articles or prepositions. The paper gives examples including "Methods in Intelligent Human" and "Battle for Wagner" [2606.11105]. Numeric placeholders are later mapped back to digits with controllable distributions [2606.11105].

After generation, candidate concepts undergo **non-existence verification** using Dolma v1.7, an open 3T-token English corpus spanning 15 sources, together with Infini-gram for efficient exact-match search [2606.11105]. Because Infini-gram is case-sensitive, the pipeline approximates case-insensitive matching by summing matches under four casings: original, UPPER, Title Case, and lower. A candidate is retained only if the total number of matches is zero [2606.11105]. The authors characterize this as a robust heuristic rather than a formal guarantee.

After filtering, PhantomBench contains **62,411 non-existent concepts**, consisting of **36,901 terms** and **25,890 entities** [2606.11105]. The paper also defines targeted subsets including **Phantom-T**, **Phantom-E**, **Phantom-Med**, and **Phantom-Legal** for more controlled analyses by domain and concept type [2606.11105].

Human validation supports the plausibility of the generated concepts. Two fluent English annotators rated generated and rare real concepts on 5-point plausibility and specificity scales. For terms, there were no significant differences in plausibility or specificity between generated and rare real terms. For entities, plausibility was similar, while specificity was significantly lower for generated entities [2606.11105]. This indicates that the pipeline yields realistic-enough names for stress-testing models, while also identifying a concrete weakness in entity generation quality.

## 3. Task formulation and evaluation protocol

For each non-existent concept $X$, PhantomBench asks the model questions about different properties of $X$, with prompt templates that vary in how strongly they presuppose that $X$ exists [2606.11105]. This distinction is methodologically central. **Existence** prompts do not assume reality, whereas **meaning**, **date**, **place**, and several term-specific prompts implicitly do [2606.11105].

The benchmark uses the following property families [2606.11105]:

| Property | Applies to | Characterization |
|---|---|---|
| Existence | Terms and entities | Low presupposition |
| Meaning | Terms and entities | Presupposes existence |
| Date | Terms and entities | Presupposes existence |
| Place | Terms and entities | Presupposes existence |
| Etymology | Terms only | Presupposes existence |
| Application | Terms only | Presupposes existence |
| Relation | Terms only | Presupposes existence |

The operational definition of hallucination is simple: because $X$ is guaranteed to be non-existent by construction, any response that treats $X$ as real and provides specific factual content about it is a hallucination [2606.11105]. Abstention is defined broadly and includes explicit uncertainty, inability to find information, requests for more context, or reformulations that avoid making unsupported claims about $X$ [2606.11105].

To classify responses, PhantomBench uses an **LLM-as-a-judge** protocol with **Gemini 2.5 Flash** as the judge model [2606.11105]. The judge receives the concept, the user input, and the model answer, and returns JSON of the form:

```json
{"is_abstention": true}
```

or

```json
{"is_abstention": false}
```

The instruction explicitly labels as abstention responses that express uncertainty, point out problems in the query, ask for more context, or answer a revised question instead of hallucinating about the original concept [2606.11105].

The primary metric is **Hallucination Rate (HR)**, defined as

$$
\text{HR} = \frac{1}{N} \sum_{i=1}^{N} \mathbf{1}[\text{Abstain}(r_i) = \text{False}]
$$

where lower values are better [2606.11105]. HR can be computed by model, subset, or prompt type. The judge itself was validated via the **Alternative Annotator Test** on 120 prompt-response pairs with four human annotators; the reported results are Gemini’s winning rate $\omega = 1.00$ and advantage probability $\rho = 0.98$, which the paper interprets as reliability comparable to a human annotator for this task [2606.11105].

A further layer of analysis refines abstention into six answer properties, labeled A-F: **Uncertainty**, **Alternative**, **Context**, **Decompose**, **Presume**, and **Factual answer**, where only F is non-abstention [2606.11105]. This yields a more granular picture of how models fail or hedge when confronted with non-existent concepts.

## 4. Models evaluated and principal empirical findings

PhantomBench evaluates **21 models** in total, with six core models on the full benchmark and additional models on subsets [2606.11105]. The core models are **Llama 3.1-8B**, **Gemma 2-9B**, **Gemma 3-12B**, **Qwen 2.5-7B**, **Qwen 3-8B**, and **Mistral 7B v0.3** [2606.11105]. Additional analyses include larger and smaller Qwen and Llama variants, proprietary Gemini models, the base model **OLMo-7B**, reasoning models such as **DeepSeek-R1-Distill-Qwen-32B** and **GPT-OSS-20B**, and domain-specialized models including **BioMistral-7B**, **MedGemma-4B**, and **SaulLM-7B** [2606.11105].

The main empirical result is that hallucination on non-existent concepts is common. The paper reports average hallucination rates as high as **86.7%** in some settings, notably **Gemma 3-12B on meaning prompts** [2606.11105]. Across core models, **meaning** prompts are the hardest, with average HR around **33.4%**, whereas **existence** prompts are easier, with average HR around **16.2%** [2606.11105]. This gap indicates that many models can signal uncertainty when directly asked whether something exists, yet hallucinate when the prompt presupposes existence and requests a semantic or factual attribute.

Among core open-source models, **Llama 3.1-8B** and **Qwen 2.5-7B** are described as relatively strong abstainers, while **Gemma 3-12B** shows very high HR across several prompt families [2606.11105]. On subset results reported in the paper, **Gemma 3-12B** reaches **87.26%** HR on Phantom-T meaning prompts and **85.99%** on Phantom-E meaning prompts, with date and place attributes also above 70-80% [2606.11105].

Prompt presupposition has a large effect. Existence queries such as asking whether $X$ is real often elicit some uncertainty or refusal. By contrast, meaning queries such as "What does $X$ mean?" frequently cause models to generate plausible-looking definitions. Date, place, etymology, application, and relation prompts also produce substantial hallucination, especially because these prompts offer many possible continuations that are locally coherent even when globally false [2606.11105].

The paper also finds that **reasoning models hallucinate more**. Reported examples include **GPT-OSS-20B** with **54.77%** HR on Phantom-T meaning, **55.01%** on date, and **72.68%** on place, and **DeepSeek-R1-32B** with **66.96%** on meaning, **43.08%** on date, and **58.97%** on place [2606.11105]. The authors relate this to recent findings that chain-of-thought or reasoning fine-tuning can exacerbate hallucination on questions outside the training distribution [2606.11105]. A plausible implication is that mechanisms that improve elaboration and problem decomposition can also increase commitment to fabricated latent explanations when knowledge is absent.

Model size does not improve abstention monotonically. Within the **Qwen 3** family, HR generally decreases from **1.7B** to **14B**, but **Qwen 3-32B** regresses on several settings, including Phantom-E meaning and place [2606.11105]. Within **Llama 3**, the **70B** variant shows substantially higher HR than the **8B** variant on meaning and place prompts; the paper gives **26.42%** versus **50.94%** on Phantom-T meaning as one example [2606.11105]. This undermines the assumption that larger models are necessarily more reliable at recognizing knowledge boundaries.

Domain specialization yields mixed and sometimes adverse results. In the biomedical setting, **MedGemma-4B** substantially improves over **Gemma 3-4B**, but **BioMistral-7B** performs worse than **Mistral 7B v0.1** [2606.11105]. In the legal setting, **SaulLM-7B** performs substantially worse than **Mistral 7B v0.1**, with HR **89.69%** versus **63.79%** [2606.11105]. The paper interprets this as evidence that domain fine-tuning can amplify overconfident answer behavior rather than calibrated abstention.

## 5. Non-existent concepts as a proxy for rare knowledge

A major contribution of PhantomBench is the argument that model behavior on non-existent concepts is a useful proxy for behavior on **rare real concepts** [2606.11105]. To test this, the authors construct parallel datasets of non-existent concepts, rare existing concepts, and common existing concepts using the same Dolma plus Infini-gram frequency pipeline [2606.11105].

The analysis compares abstention rates across these sets. An ideal model would abstain often on non-existent concepts while rarely abstaining on common real concepts. The paper reports that many models do abstain more on non-existent than on common concepts, but the best abstainers also show relatively high abstention on existing concepts, indicating that low HR can sometimes reflect generalized caution rather than precise boundary detection [2606.11105].

The central quantitative result is a Pearson correlation analysis over 64 settings. The reported correlation between abstention rates on **non-existent** and **rare** concepts is $\rho = 0.755$ with $p < 0.001$, whereas the correlation between **non-existent** and **common** concepts is $\rho = 0.322$ with $p = 0.009$ [2606.11105]. The paper takes this as evidence that PhantomBench is a useful proxy for model behavior on rare knowledge rather than merely a specialized artifact of synthetic non-existence.

The fine-grained abstention taxonomy reinforces this point. For most models, the distribution over categories A-E is similar for non-existent and rare real concepts, while common concepts differ more substantially, especially by showing a larger share of **Presume** responses [2606.11105]. This suggests that non-existent and rare concepts elicit comparable uncertainty-management patterns, even when their ontological status differs.

Another significant qualitative finding concerns **hallucinations inside abstentions**. In a manual inspection of 64 abstaining responses, about **35.9%** contained specific factual claims about related concepts, and **47.8%** of those contained factual errors or referred to non-existent entities, corresponding to about **17.2%** of all inspected abstentions [2606.11105]. This means that top-level abstention is not sufficient as a reliability signal if the surrounding explanation still contains unsupported facts.

## 6. Relation to PhantomWiki, limitations, and benchmark-design significance

The term "PhantomBench" also has a second, benchmark-design-oriented meaning in PhantomWiki [2502.20377]. PhantomWiki is an on-demand generator of synthetic Wikipedia-style universes consisting of fictional people, relations, and short biography-like articles, with questions generated by a context-free grammar and answers computed by a Prolog world model [2502.20377]. The paper argues that a benchmark called "PhantomBench" could be defined as a concrete PhantomWiki configuration with fixed universe size, question difficulty, task format, and evaluation protocol [2502.20377]. In that sense, PhantomBench denotes a standardized suite built from a public generator rather than a single static dataset.

This alternative usage is conceptually distinct from the non-existence benchmark of [2606.11105]. PhantomWiki is aimed at disentangling reasoning, retrieval, and tool use through explicit control over reasoning depth and corpus size [2502.20377]. The 2026 PhantomBench, by contrast, isolates abstention and hallucination on non-existent concepts [2606.11105]. The shared naming reflects a family resemblance in benchmark philosophy: both reject static, easily saturated datasets and favor scalable generation procedures that can produce fresh evaluation instances. This suggests a broader design principle in current LM evaluation: robust benchmarking increasingly depends on controllable generation pipelines rather than immutable test sets.

The limitations of PhantomBench [2606.11105] are substantive. Non-existence verification is imperfect because a concept may exist yet be absent from Dolma v1.7, and Dolma has a knowledge cutoff around 2023, so newly emerged concepts may be misclassified as non-existent [2606.11105]. The pipeline is English-only and text-only, and the infrastructure requirements for Dolma plus Infini-gram are significant [2606.11105]. Generated entities, while plausible, are less specific than rare real entities [2606.11105]. The benchmark measures behavioral abstention, not calibrated internal uncertainty [2606.11105]. It also does not directly resolve how models would behave when coupled to live web search, although the paper notes that many real deployments cannot use web search because of privacy, offline, or high-stakes constraints [2606.11105].

The broader significance of PhantomBench is methodological rather than purely empirical. It provides a clean stress test for knowledge boundaries, a reusable generation pipeline for domain-specific evaluation, and evidence that strong instruction-tuned, reasoning, and domain-specialized models remain far from reliable abstention on non-existent concepts [2606.11105]. In conjunction with PhantomWiki’s generator-centered view of benchmarking [2502.20377], it exemplifies a contemporary shift toward evaluation frameworks that are scalable, refreshable, and explicitly designed to resist benchmark saturation while isolating particular failure modes.

Source: https://www.emergentmind.com/topics/phantombench