When a language model links a name in text to an entry in a knowledge base such as Wikipedia, it usually picks the best candidate even when the right entity is not in the base at all. The failure is worst when an obscure person or company shares a name with a famous one. The model confidently links to the famous namesake, and the error flows into downstream facts as a hallucination. A standard defense against hallucination is to sample the model several times and trust answers it repeats. That defense breaks here, because a strong prior toward the famous entity makes the model repeat the wrong choice every time. We propose a different use of repetition. The model never chooses among candidates. It is asked several times to describe, from the surrounding text alone, what the mentioned thing is. We then measure two things separately: whether those descriptions agree with each other, and whether they agree with any candidate's knowledge-base entry.
The point is that agreement inside a sampled answer can come from true knowledge or from a shared bias. When the model is shown a list that includes the famous namesake, the bias attaches to the list. It attaches much less to a free description driven by the context, which says things like "a regional footballer born in 1994". If the description is stable but matches no entry, the entity is probably real but missing from the base. If the description is unstable, the mention is probably a phrase that names no entity, or something the model cannot pin down. If the description is stable and matches an entry, the model should link. This three-way split also answers a question that current hallucination detectors cannot: whether an ungrounded entity is missing from the base or was invented.
Core Claim
Asking the model to choose among candidates and checking whether it repeats its choice gives the wrong kind of confidence for detecting entities absent from a knowledge base. It measures how stable the choice is, and the pull of a famous name makes that choice stable. A better signal comes from sampling open descriptions of the mention without showing candidates, then scoring each candidate by how many samples a textual entailment check says are supported by that candidate's description. We claim this "profile consistency" score beats single-shot abstention and majority-vote abstention on the case that matters: true missing entities whose famous namesake remains in the base. It should also keep accuracy on entities that are present.
Where This Sits in Prior Work
Two lines of work meet here without covering this case. The first uses sampled agreement as a stand-in for truth. SelfCheckGPT scores sentences by whether resampled passages contradict them. FINCH-ZK extends this across several models and prompt rephrasings at the level of text blocks. Selectively Answering Ambiguous Questions shows that counting repeated answers beats token likelihood under ambiguity. We reuse this method but move it from sentences to single mentions of entities. None of these papers asks whether an unsupported claim concerns an entity missing from a knowledge base.
The second line handles entities missing from a knowledge base, usually called NIL entities. BLINKout adds a learned NIL token to a fine-tuned cross-encoder. CLINK penalizes candidates whose type conflicts with the mention's type, drawing on 187 Wikidata types. RAISED prompts an LLM to pick a retrieved candidate or answer "None of the candidates". Both trained systems need labeled NIL data and a fixed schema. RAISED is closest to our setting, and its authors say two things we build on. First, LLM selectors are pulled toward famous namesakes even when the right candidate has been retrieved. Second, the method was never tested on true NIL entities, only on cases where retrieval missed an entity that exists in the base.
Three results constrain the design. Rethinking Hallucinations shows that consistency detectors measure randomness rather than correctness, and that some errors are prompt-agnostic: the model repeats the same misconception under every wording. Do Language Models Know When They're Hallucinating References? is the nearest precedent for our mechanism. It checks whether a cited paper exists by asking indirect questions about its authors several times, and this beats asking directly whether the paper exists. We carry that indirect-probe idea from citation titles to entity linking against a knowledge base, where the question is not "does this exist" but "is this the entry in the base". Distinguishing Ignorance from Error shows that prompting fixes errors on facts the model knows but barely helps when knowledge is missing (<3%). This is why we do not expect instructions such as "abstain if unsure" to solve missing-entity cases.
Research Program
For each mention $m$ in context $x$, we draw $N$ short descriptions $d_1,\dots,d_N$ from the model at non-zero temperature. The prompt asks what $m$ refers to in $x$, with type, role, dates and affiliations, and it shows no candidates. A retriever proposes candidates $c$ with knowledge-base descriptions $\text{desc}(c)$. We use BM25, as in RAISED, so that retrieval matches that baseline. An entailment model or LLM judge decides whether each $d_i$ is supported by $\text{desc}(c)$. We then compute two quantities:
$$\text{support}(c) = \frac{1}{N}\sum_{i=1}^{N} \mathbb{1}[\text{desc}(c) \text{ supports } d_i], \qquad \text{stability} = \frac{|\text{largest cluster of mutually entailing } d_i|}{N}.$$
The system links to $c^* = \arg\max_c \text{support}(c)$ when $\text{support}(c^*) \ge \tau$. It predicts "missing entity" when stability is high and support is low, and "non-entity or unresolved" when stability is low. The thresholds are tuned on a held-out development split.
We add one diagnostic that targets the famous-name trap. We also sample descriptions from the name alone, with the context removed. If the descriptions made with context collapse onto the descriptions made from the name alone, the output is driven by the prior rather than the text, and we discount support for the matching candidate. This component is ablated separately, because it may add cost without benefit.
We build one new evaluation split, because the file reports no benchmark of true NIL mentions that also have famous namesakes. We take the long-tail mentions from ShadowLinks, which belongs to the ZELDA suite. We delete their gold entities from the knowledge base and keep the famous namesakes. This follows the pruning logic of MM-pruned in BLINKout. Every mention in this split is a true NIL with a strong wrong attractor. The existing NIL benchmarks give the natural cases: NEL from CLINK, with Missing Entities and Non-Entity Phrases labeled separately, and MM-2014AB and NILK-sample from BLINKout, which use time-split knowledge bases.
As a secondary application, we run the method on the entities in HalluEntity, which contains GPT-3.5 biographies with 18,785 annotated entities. We link each entity to Wikidata and test whether the "unstable" outcome tracks annotated hallucinations better than the "stable, no match" outcome does. HalluEntity does not label real entities that are missing from the base. This part therefore tests only whether the stability side of the score separates invented entities. It does not validate the three-way split.
We run everything black-box with two model families: one API model of the kind used in RAISED, and Qwen3 run locally. Running Qwen3 lets us also compute token-probability baselines.
Baselines
The naive reference is forced-choice linking: always take the top retrieved candidate, which gives zero NIL recall. Next to it is a fixed cutoff on retriever score, the brittle heuristic that BLINKout reports being fooled by near-miss entities.
The strongest existing black-box competitor is RAISED: BM25 retrieval plus an LLM selector with a "None of the candidates" option. It reports inKB micro-F1 of 86.3 and abstention-aware F1 of 90.7 on ZELDA with GPT-5.4 mini (Select, Don't Train). We rerun it with the same retriever, the same candidate lists, and the same models as our method. We first check that we reproduce the reported ZELDA numbers within about one point.
The key ablation is RAISED with majority voting: the same selector sampled $N$ times, with the fraction of samples voting for the winner used as confidence. This is vanilla self-consistency, and it isolates our one change: describing the mention instead of choosing among candidates. We match it to our method on total generated tokens, not on sample count. Our descriptions are longer, so the voting baseline receives more samples.
For supervised methods, we use CLINK, which reports 87.71% overall accuracy on NEL (Learn to Not Link), and BLINKout, which reports a 30–40% relative F1 gain in NIL discovery on MM-2014AB and NILK-sample (Reveal the Unknown). On these benchmarks they are trained in-domain. We report their published numbers and do not claim a matched budget. For the constructed shadow-pruned split there is no reported result on Lacuna, so all numbers there come from our own reruns. On HalluEntity, the reference is the Focus method, which reports AUROC 0.78 and AUPRC 0.40 on Llama3-8B (HalluEntity). We also run SelfCheckGPT restricted to sentences that contain entities.
To count as beating the baselines, our method must exceed majority-vote RAISED by at least 5 NIL-F1 points on the shadow-pruned split at matched tokens. It must also stay within 1 point of RAISED's inKB F1 on unmodified ZELDA and at least match RAISED on NEL overall accuracy.
Why not just…?
Why not just take a majority vote over RAISED's "None of the candidates" decisions? This is the real threat, and it is the ablation above. Our argument against it is mechanistic. RAISED reports that selectors prefer famous namesakes. Rethinking Hallucinations reports errors that persist across prompt variants. Together these predict that votes on the shadow-pruned split will consistently favor the wrong famous entry. If the voting baseline's confidence on those wrong links is not systematically high, and it ties our method, then the description step adds nothing. We would report the idea as reducing to self-consistency.
Why not just read token probabilities or ask the selector for a confidence number? Can LLMs Express Their Uncertainty? finds that instruction tuning makes stated and token-level confidence overconfident. HalluEntity finds that likelihood and entropy flag many rare but true entities, giving high recall and poor precision. Both baselines appear in the Qwen3 runs, so this will be measured rather than assumed.
Why not just fine-tune a NIL classifier like BLINKout or CLINK? Where labeled NIL data and a fixed knowledge base exist, these may win, and we expect them to be competitive on MM-2014AB. The case for our method is that it needs no retraining when the knowledge base changes or its type schema no longer fits, which is the limitation CLINK names. The NILK-sample time split tests that transfer directly.
Why not just verify existence with web search? Search, as in NEED4Tweet, answers whether something exists. It does not answer whether the knowledge-base entry is that thing, and an overshadowed mention returns results about the famous entity. Rethinking Hallucinations also finds that retrieval raises inconsistency across prompts. We include a search-then-select variant only as a secondary comparison.
Decisive Checks
The central experiment is the shadow-pruned split with matched token budgets. Our method beats majority-vote RAISED by at least 5 NIL-F1 points while keeping inKB F1, or the claim fails. Beyond F1 on each outcome, we report AUROC of the NIL score, overall accuracy, and accuracy against coverage as the threshold varies.
The second check is on NEL. There, the stability score should separate Non-Entity Phrases from Missing Entities better than a single NIL score can. If it cannot, the three-way split is not real.
The third check plots token cost against NIL-F1 for $N$ from 1 to 16. This tells whether a small $N$ is enough for practical use.
Finally, we report where the method fails. The most likely failure is a model that holds a confident false belief about the obscure entity itself, producing a stable description that matches nothing. That is exactly the kind of error consistency cannot see.
References
- SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models
- Zero-knowledge LLM hallucination detection and mitigation through fine-grained cross-model consistency
- Selectively Answering Ambiguous Questions
- Reveal the Unknown: Out-of-Knowledge-Base Mention Discovery with Entity Linking
- Learn to Not Link: Exploring NIL Prediction in Entity Linking
- Select, Don't Train: The Benefits of Modular Entity Disambiguation with LLM-Based Selection
- Rethinking Hallucinations: Correctness, Consistency, and Prompt Multiplicity
- Do Language Models Know When They're Hallucinating References?
- Distinguishing Ignorance from Error in LLM Hallucinations
- HalluEntity: Benchmarking and Understanding Entity-Level Hallucination Detection
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs
- NEED4Tweet: A Twitterbot for Tweets Named Entity Extraction and Disambiguation