A probability-only hallucination detector cannot distinguish competing factual answers from paraphrases when their token probabilities match.
That blind spot matters for language model builders. Across three open-weight models and three question-answering benchmarks, GURI increased mean test ROC-AUC from 72.09 % for LOS-Net to 79.09 %. Shaked Goldshtein, Yitzchak Grunbaum and their coauthors achieved that gain by giving the detector candidate token identities alongside their probabilities.
The result points to an input problem. A more capable detector still cannot recover distinctions that its input has discarded.
What the probabilities leave out
Consider a language model choosing how to complete a sentence. Its candidate probabilities describe how it distributes uncertainty. Candidate token identities reveal what that uncertainty concerns.
The paper illustrates this with matching probability distributions.
For “The capital of France is,” the candidates include London, Paris, Berlin and Rome. For “The building is extremely,” the candidates include large, big, huge and enormous.
The probability-only detector receives the same numerical input in both examples. But the candidate token identities expose different alternatives. The first example contains competing factual answers. The second contains words with similar meanings.
The paper’s comparison figure makes this distinction visible. The probabilities match. The candidate token identities change what those probabilities can tell the detector.
This doesn’t mean candidate token identities establish whether an answer is correct. It means probability-only inputs erase a distinction that can help predict correctness.
What GURI adds
GURI combines candidate token identities with their probabilities, then aggregates those representations across the answer. The researchers train the fusion and prediction components jointly for answer-level hallucination detection.
GURI uses a shared external tokenizer to encode candidate token strings. That gives the same string a common representation across target models, without requiring their native tokenizer or embedding table.
The researchers use fixed random identity codes by default. Those codes match or outperform pretrained embeddings in their experiments.
That finding needs a narrow reading. The main gains do not require pretrained semantic geometry. The experiments do not establish that lexical structure or tokenization structure is irrelevant.
For a builder, this makes the representation worth examining before adding a more elaborate detector. The useful signal may already sit beside the probabilities.
Transfer helps, but access still matters
Among the three open-weight models, GURI outperformed transferred LOS-Net in 17 of 18 directed cross-model transfers without target fine-tuning.
That supports transfer within the tested setting. It doesn’t establish that GURI will transfer equally well to every model.
The access requirement also limits deployment. GURI needs candidate token strings and probabilities. Some commercial APIs do not expose them. Without those inputs, a builder cannot apply the method as described.
Gray-box access avoids the need for hidden states, but it still imposes an interface requirement.
What the benchmark leaves unresolved
The evaluation uses automatic string-match labels. Those labels can reward a correct substring inside an otherwise incorrect response. They can also penalize appropriate abstention.
So the reported ROC-AUC measures performance against those labels. It does not establish reliable detection of every factual error in a complete answer.
The open-weight experiments also included continuation after end tokens in detector inputs and labeling. The researchers did not evaluate the effect of stopping at the first end token. A builder whose generation stops there cannot assume the reported result carries over unchanged.
The retained NQ-WC data excluded the longest-context cases. That limits what the evaluation says about those cases.
The reported variation also needs care. The main comparison reports mean test ROC-AUC and standard deviation across 36 runs. Those runs vary both hyperparameter configurations and training seeds. The standard deviation therefore does not isolate seed variation.
I would test GURI with the candidate token access, stopping behavior and correctness labels of the intended application before trusting its score.
The paper gives builders a concrete reason to retain candidate token identities. Discarding them makes distinct language model decisions look identical to the detector. Keeping them improves detection in the reported experiments, while leaving the deployment and labeling limits to resolve.
Top comments (0)