closed-book-hallucination-detection-training-small-supervised-classifier-over-many CLM-0030For closed-book hallucination detection, training a small supervised classifier over many heterogeneous uncertainty scores beats the best single score, but the gain comes from the diversity of sampling-based black-box signals: ensembles over only single-generation token-probability features add little, while black-box-only ensembles do nearly as well as the full set.
Capability: Stating false facts confidently
Observed on
Four closed-weight LLMs (Gemini 2.5 Flash/Pro, GPT-4o, GPT-4o-mini), nine datasets, short-form QA, claim-level long-form, and Python competitive-programming code; labels available .
Sources
- AUROC over 25 stratified splits per setting; best ensemble beat the best individual scorer in 30 of 32 settings and had lowest ECE in 29 of 32, with gains from ~100 labeled instances; ablation on five short-form datasets: black-box-only ensembles beat the best black- box scorer in 19 of 20 settings, white-box-only in 11 of 20. The single-scorer baseline is selected on the test set, so it is optimistic. Sample-efficiency finding replicates Bakman et al. Margins are often 0.01-0.02 AUROC and sometimes inside confidence intervals; [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For black-box hallucination detection using only sampled responses and token log-probabilities, semantic-entropy signals go to zero whenever the samples collapse into one semantic cluster — which happens for a large share of hallucinated queries, up to nearly all of them on tasks with confusable entities and figures — while token-level uncertainty still separates many of those cases, so the two signals fail on different queries and combining them detects more than either alone.Stating false facts confidently · reviewed by AI
- An abstention rule built from self-consistency, where the model samples several answers and itself scores how many agree with a candidate, can be thresholded by conformal prediction to bound the hallucination rate at a chosen level on closed-book QA; against log-probability confidence it abstains much less often on long-form answers (Temporal Sequences) for the same error bound and matches it on short answers (TriviaQA).Saying it does not know · reviewed by AI
- When multiple stochastic runs of several mid-tier LLMs each produce an unranked set of candidate security requirements from the same control templates, pooling the runs recovers far more valid items than any single run but also many hallucinations, and ranking the pooled candidates by how many runs produced them separates valid from hallucinated items well enough to beat every individual run's precision-recall and ROC area.Stating false facts confidently · reviewed by AI
- A ~2M-parameter classifier head that reads hidden states already computed during generation (tapping shallow, middle and deep layers of the generating model) detects response-level hallucination at least as well as specialized hidden-state hallucination detectors, and its detection quality rises with the scale of the base model it is attached to.Stating false facts confidently · reviewed by AI
- For detecting hallucinated text spans in vision-language model outputs, span localization and confidence calibration come from different architectures: a fine-tuned generative VLM proposes better span boundaries but scores them poorly, while a discriminative token tagger trained with a direct regression loss on human-consensus confidence gives better-correlated probabilities but conservative spans, so using the generative spans as proposals and the tagger's probabilities as scores beats either alone.Stating false facts confidently · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: An evaluation where an ensemble restricted to single-generation white-box token-probability features matches full or black-box ensembles, or where the full ensemble fails to beat a validation-selected single scorer in most settings. Proposed technique, not catalogued: Supervised ensemble of uncertainty-quantification scores.