acceptodds
Under review as a conference paper at ICLR 2027

READ: Combining Reader Evidence for Non-Intrusive Black-Box Hallucination Detection

Abstract

Hallucination detection supports reliable use of large language models (LLMs), but many existing methods require internal access or repeated interaction with the target model. We study non-intrusive black-box hallucination detection (NI-BHD), which assesses previously generated responses without either requirement. An accessible model, termed a reader, can provide detection signals through teacher forcing on the question and observed answer tokens. Our analysis reveals a reader suitability mismatch: model size, answering ability, and reader–target family match do not reliably guide reader selection. Reader rankings also vary across targets and benchmarks; readers provide complementary evidence. Accordingly, we derive an information-theoretic characterization of a reader’s contribution beyond the selected evidence via optimal population log-loss reduction. We therefore propose reader evidence aggregation for detection (READ), a supervised framework selecting and combining multiple readers. READ selects readers by cross-validated log-loss reduction and integrates their hidden-state probe scores and token-likelihood features using a fixed additive logistic model. Experiments on 6 target LLMs and 7 benchmarks achieve 0.911 cell-macro AUROC, improving over the strongest published baseline under the same access constraints by 0.116. Despite strictly less access, READ also exceeds the per-environment best of 7 published white-box detectors in 19 of 28 matched environments.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
hallucination detection, large language models, uncertainty quantification

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_s1hdOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies non-intrusive black-box hallucination detection. The authors show that intuitive reader-selection criteria, such as model size, answering ability, or reader-target family match, do not reliably predict detection performance, while different readers provide complementary evidence. Based on this observation, they propose READ, which selects multiple reader models according to cross-validated conditional log-loss reduction and combines their hidden-state probe scores and token-likelihood features using an additive logistic model. Experiments demonstrate strong performance, reaching 0.911 cell-macro AUROC and outperforming existing methods under the same access constraints.

Critical strengths
  1. Hallucination in LLMs is itself a worthwhile research problem, and the paper studies it under a realistic constraint that the detector cannot access target internals or query the target further. This setting reflects real deployment scenarios and is clearly formalized.

  2. The paper shows that larger, stronger, or same-family readers are not consistently better detectors, and that combining readers can outperform even oracle single-reader selection. It further motivates reader selection via conditional mutual information, decomposing a reader’s value into individual signal and redundancy, providing a principled criterion for selecting complementary readers.

Critical weaknesses
  1. The comparison mainly focuses on access constraints rather than computational cost. READ requires supervised probe training, cross-validated reader selection, and multiple reader forward passes at inference. It would be helpful to provide a more systematic comparison of training and inference cost against baseline methods.

  2. The main evaluation primarily measures generalization to unseen questions within seen target and benchmark distributions. Although the appendix includes target-family and benchmark holdout experiments, performance drops under these distribution shifts, and the advantage of conditional-gain selection becomes relatively small. In particular, under benchmark holdout, it achieves only a modest AUROC improvement over individual-signal selection and even yields worse log loss. Therefore, the evidence for robust out-of-distribution generalization is somewhat limited.

  3. It remains unclear whether READ is equally effective on stronger models such as GPT-5/6, where hallucinations may be rarer and more subtle.

Questions for authors
  1. What is the training and practical inference cost of READ relative to competing black-box methods?

  2. Does READ equally effective on stronger models?

3 comments
AuthorsOpenReview ↗

Response to Reviewer s1hd (1/2)

Thank you for the review and for noting the practical value of the setting.

W1, Q1. The comparison mainly focuses on access constraints rather than computational cost. What is the training and practical inference cost of READ relative to competing black-box methods?

We measured the cost of READ and of every baseline that queries the target or a critic. All numbers are per response, on our 4×H100 server, with mean input 186 and answer 61 tokens.

MethodExtra target callsModels runFLOPsWeights in GPU memoryLatency
READ (K=4)04 readers (12.6B text parameters), one forward pass each6.2 T26.8 GB84 ms on 1 GPU
SelfCheckGPT, Llama-3.3-70B10 samplestarget + NLI model115 T142 GB2.0 s on 4 GPUs
Prompted critic, Qwen3.5-27B027B critic, prefill + decoding23.5 T54 GB1.0 s on 1 GPU
Prompted critic, Qwen3.5-9B09B critic, prefill + decoding7.8 T18 GB0.44 s on 1 GPU
Prompted critic, Gemini 3.6 Flash0API call––1.6 s

FLOPs count dense forward passes, 2 × parameters × tokens. Latency is the median over 200 responses, bf16, one response at a time. SelfCheckGPT prefills the prompt once and decodes its ten samples as one batch. Memory counts all loaded weights, including the unused vision towers of two readers. READ runs its four readers one after another on one GPU. On four GPUs in parallel it takes 36 ms. SelfCheckGPT on a 70B target needs all four GPUs and spends most of its time decoding ten samples. For API targets it also needs ten paid generations per response. READ needs none.

Training is light. One pass of all 14 candidate readers over the 7,560 responses, saving hidden states at four depths, takes about 0.3 GPU-hours. Fitting probes at four depths and running greedy selection takes about 5 CPU-minutes per outer split. No reader is fine-tuned.

Changes in the paper. We added this table and the training cost as Table 29 in the new Appendix F.8, and a summary to Section 5 as the new paragraph "Computation Cost".

AuthorsOpenReview ↗

Response to Reviewer s1hd (2/2)

W2. The main evaluation primarily measures generalization to unseen questions within seen target and benchmark distributions. … Under benchmark holdout, it achieves only a modest AUROC improvement over individual-signal selection and even yields worse log loss.

You are right that the main evaluation uses unseen questions from seen targets and benchmarks. It is the standard setting for supervised detectors. Under benchmark holdout, READ drops by 0.049, close to the supervised text encoders (0.051). Its lead over ELECTRA-large grows from 0.116 to 0.118 (Table 16). Under target-family holdout, READ still reaches 0.875 (Table 15).

You are also right that the selection rule adds little under benchmark holdout. The new paired CIs (95%, source-question bootstrap) follow.

READ minusTarget-family holdoutBenchmark holdout
Individual signal+0.010 [0.004, 0.016]+0.005 [−0.001, 0.012]
Random set+0.014 [0.009, 0.019]+0.010 [0.005, 0.015]
Fixed set+0.013 [0.006, 0.021]+0.004 [−0.004, 0.012]

When the test task is new, inner-CV gains on other tasks are a weaker guide, and any four good readers do about as well. What still transfers is the use of several readers. K = 1 gives 0.751 and K = 4 gives 0.862 under benchmark holdout.

Changes in the paper.

  • Appendix C (Limitations). We added this text. "When the evaluation benchmark is absent from training, conditional-gain selection adds little over individual-signal selection or a fixed reader set, and its log loss is higher. The gain from combining several readers remains in this setting. Selecting readers for a new task may therefore need a small labeled sample from that task."
  • Section 5. New paragraph "Transfer Limits" with the same result. Appendix F.6. Table 27 (the CIs above).

W3, Q2. It remains unclear whether READ is equally effective on stronger models such as GPT-5/6, where hallucinations may be rarer and more subtle.

GPT-6 Astra was released on 3 September 2026, three weeks before the full-paper deadline, and its API access was staged at the time. We could not collect and label a new target in that window. Your point stands, and GPT-6 Astra is now available, so we added it as a seventh target. We generated responses for the 2,014 source questions of the primary dataset with all nine prompt forms and labeled them with the same judge. Hallucinations are rarer (1.5–21% per benchmark, median 2.4%, against 4–34% for the six original targets). Where fewer than 90 were found, we use all hallucinated responses and the same number of correct ones, 464 + 464 in total.

READ is applied without any GPT-6 label. Its readers, probes and combiner are trained only on the original six targets. Each GPT-6 response is scored by the outer-split model whose held-out fold contains its source question, so no detector has seen a labeled response to the same question. Two annotators checked 140 GPT-6 labels (20 per benchmark), with a third resolving disagreements, and the adjudicated labels agreed with the judge on all 140.

Method on GPT-6 Astra (7 benchmarks)AUROC
ELECTRA-large (trained on the six targets)0.692
LLM-Check0.809
READ0.897 [0.877, 0.915]
GPT-6 AstraTQAHaluTrivCoQAGSMMBPPxLAMAvg.
Hallucination rate (%)21.02.41.81.51.99.06.0
Hallucinated responses used90614738489090
READ AUROC0.8750.9140.8770.9390.9020.8560.9140.897

GPT-4.1 is in the training data, so this is a new target from a seen family. READ on GPT-4.1 and Claude-Sonnet-4, trained with their own labels, reaches 0.924 and 0.916 over the same seven benchmarks. The result, 0.897, lies below these and above the target-family holdout of Table 15 (0.875), as expected for an unseen target from a seen family with rarer errors. The gap over ELECTRA-large and LLM-Check stays large.

Changes in the paper. We added the full setup and both tables to the new Appendix F.7 (Table 28), and a summary to Section 5 as the new paragraph "Stronger Unseen Target".

Thank you for the review. Your questions were fair, and working on them helped us find the limits of our method more clearly. We hope the revised paper and our replies resolve your main concerns. If they do, we would appreciate it if you could update your score. If not, please tell us what is still missing, and we will do our best to address it.

Reviewer_gfHVOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper formulates non-intrusive black-box hallucination detectio: scoring an already generated answer from the question–answer pair alone, without target-model internals and without additional generations or queries to the target. Auxiliary open-weight “reader” LLMs read the fixed pair by teacher forcing, and from each reader the method extracts a supervised hidden-state probe score and mean/minimum answer-token log-likelihoods. Contributions: (i) an empirical analysis of 14 readers showing a “reader suitability mismatch”: reader size, answer-recovery accuracy and family match with the target do not reliably predict detection AUROC, reader rankings vary across environments, and combining readers beats even an oracle single reader; (ii) READ itself, which greedily selects four readers by cross-validated log-loss reduction and combines their evidence with a fixed regularized additive logistic model. Validation: 6 targets (4 open-weight, 2 API) × 7 benchmarks, with 90 + 90 labeled responses per pair and source-question-grouped nested splits. Ablations cover selection rules, repeated-view controls, fusion models and the reader budget (K = 1–14), plus target-family and benchmark holdouts, an answer-recovery-failure subset, reader-family removal and paired factual-edit tests.

Critical strengths
  1. Careful, well-controlled protocol that makes the central empirical finding credible. The paper uses nested source-question-grouped splits, paired cluster-bootstrap CIs, a 490-response label audit with human adjudication, and many controls (repeated views, random and fixed reader sets, fusion alternatives, transfer holdouts, reader-family removal). This matters because it lets the reader trust that the four-reader stack beats development-selected and oracle single-reader selection, and that four distinct readers beat four layers or resampled probes of one reader

  2. Analyses that probe what the detector actually uses. Detection stays at 0.907 pooled AUROC on 168 responses on which all 14 readers fail generation, recognition and verification. And scores rise for entity, numeral and negation edits but barely for paraphrases. This supports the claim that reader evidence carries more than answer recovery or surface cues, which is the main practical argument for the approach

Critical weaknesses
  1. The paper does not show what READ’s own contribution (reader selection and its information-theoretic motivation) adds beyond supervised multi-reader stacking. o Random four-reader sets reach 0.887 AUROC versus 0.911 for READ, while single-reader references are 0.816–0.839. READ with K=3 already beats random K=4. Under benchmark holdout, a fixed set of four small Qwen readers reaches 0.858 AUROC and 0.471 log loss versus 0.862 and 0.486 for READ. Selection is called “central to the performance–computation trade-off”, but no random-K curve is shown. o The quantities in Eq. 7–8 (individual signal, signed redundancy, synergy) are never estimated on the real readers. Implemented cross-validated score equals conditional information plus differences in the logistic model’s approximation gap and in fitting/validation error. The procedure is greedy forward selection by validation loss followed by logistic stacking, and Appendix C concedes that no new theory is introduced, which to me sits uneasily with “we prove” and “principled basis” in the introduction. o Why it matters: the abstract and contributions present selection and its theory as the central step, but the evidence strongly supports only the claim that several heterogeneous readers help. It supports conditional-gain selection only modestly.

  2. The comparisons with the closest prior method and with white-box detectors are not like-for-like, so the headline margins overstate what the method adds. o LLM-Check (Sriramanan et al., 2024) scores a single response from the hidden states, attention maps and output probabilities of an auxiliary LLM. Here it is run as a training-free scorer with “an auxiliary model” and the model and score used are not stated. The gap between LLM-Check (0.760 cell-macro) and READ (0.911) therefore mixes supervision, number of readers and feature family. The paper’s own supervised single-reader controls (0.816–0.839) show that part of the gap is supervision. A supervised version of LLM-Check-style features in the same combiner is missing. o “Strictly less access” than white-box detectors holds for target access only. READ uses labels pooled across all targets and the internals of four additional models, whereas white-box probes are fitted per target, and Definition 1 allows full reader access even though the title and abstract emphasize “black-box”. o Why it matters: the main claims are relative (+0.116 over the best NI-BHD baseline, 19 of 28 environments over white-box detectors), so the regime in which the baselines are run determines how much of the gain can be credited to READ.

Questions for authors

Clarification requests

  1. Which auxiliary model(s) and which LLM-Check score(s) (hidden-state, attention, logit) produced the LLM-Check results, and how were score direction and thresholds handled?

  2. How were the recovery distractors constructed (matched by type or difficulty?), and how does four-way recognition accuracy relate to the generation and verification accuracies in Table 13? From Tables 8 and 13 I get weak correlations between reader accuracy and detection AUROC (n = 14): about −0.09 (generation), −0.21 (recognition) and +0.13 (verification). Please report these with CIs in the main text.

  3. Were the baselines’ hyperparameters and capacity chosen on development data rather than on evaluation folds?

  4. What are the latency, FLOPs and memory of K=4 READ versus SelfCheckGPT (ten samples) and the prompted critics?

Additional experiments that would most affect my assessment

  1. Supervised LLM-Check-style features fed to the same logistic combiner, for K=1 and for READ’s four readers.

  2. Random-K versus READ-K curves (K=1–8) with CIs, and CIs for the transfer comparisons in Table 15, especially READ versus the fixed set.

  3. How does the CV-gain ordering compare with an independent measure, such as the Gaussian Schur increment of Prop. 10 or the gain under an MLP/GBM combiner? How does greedy selection compare with exhaustive search over the 1001 four-reader subsets?

  4. A noise floor for the variability analysis. With 90 + 90 responses per environment, the standard error of one AUROC near 0.8 is about 0.03 (Hanley–McNeil approximation), and 13 distinct winners among 42 environments is close to what random winners among 14 equal readers would give (it would be 13.4 actually). Please provide a bootstrap or permutation null for the win counts, the rank concordance and a mixed-effects version of Fig. 3d (its 3,822 points are 91 reader pairs × 42 environments on shared responses).

  5. White-box baselines trained on the same pooled data as READ, and a repeated-view control with equalized features (for example score-only stacks for both conditions).

6 comments
AuthorsOpenReview ↗

Response to Reviewer gfHV (1/5)

Thank you for the detailed reading. Your summary of the protocol is accurate. We answer your points in four comments.

W1. The paper does not show what READ's own contribution (reader selection and its information-theoretic motivation) adds beyond supervised multi-reader stacking. … no random-K curve is shown.

E2. Random-K versus READ-K curves (K=1–8) with CIs.

READ is selection and integration together. The abstract defines it as "selecting and combining multiple readers", and both parts are ablated (Table 2c, 2d). Your question is still fair. How much does the selection rule add once several readers are stacked? We ran four new analyses to answer it. They are random-K curves with CIs, an exhaustive search over all 1,001 four-reader sets, estimates of individual signal and redundancy, and ranking checks against two independent measures. For the random-K curves, random sets use 8 draws per outer split. Differences have 95% percentile CIs from 1,000 paired source-question cluster bootstraps, as in Table 6.

K1234567814
READ0.8160.8700.8960.9110.9220.9290.9340.9360.943
Random0.7860.8380.8710.8870.9010.9130.9220.9270.943
Difference+0.030+0.032+0.025+0.024+0.021+0.016+0.012+0.0090
95% CI[0.021, 0.040][0.024, 0.039][0.018, 0.031][0.018, 0.030][0.016, 0.027][0.011, 0.020][0.008, 0.015][0.005, 0.012]–

Random K = 1 is the exact mean over all 14 single readers (Table 8). Random K = 4 and its CI are from Table 2c and Table 6. K = 14 is the full pool for both.

From K = 3 on, READ with K readers matches random sets with about 1.5 to 2 more readers. READ at K = 4 (0.911) is close to random sets of six readers (0.913), so it saves about two reader passes per response. This is what we mean by the performance–computation trade-off.

E3. How does the CV-gain ordering compare with an independent measure, such as the Gaussian Schur increment of Prop. 10 or the gain under an MLP/GBM combiner? How does greedy selection compare with exhaustive search over the 1001 four-reader subsets?

Exhaustive search. We evaluated all 1,001 sets, each as a fixed set in all splits. Only 6 of them do better than READ on held-out AUROC, and finding them needs evaluation labels. The mean over all 1,001 sets is 0.887, the random-set value. Searching all 1,001 sets by the same inner-CV loss gives 0.910, which is not better than greedy (READ − exhaustive = +0.001 [−0.003, 0.004]). Greedy evaluates 50 candidate sets per split instead of 1,001.

Independent measures. At each greedy step we rank candidates three ways. The CV-gain ranking agrees with the plug-in Gaussian Schur increment of Proposition 10 (mean Spearman 0.93). It also agrees with CV gains under the MLP combiner (mean Spearman 0.95). Greedy selection under the MLP combiner picks the same four readers in 3 of 4 splits.

AuthorsOpenReview ↗

Response to Reviewer gfHV (2/5)

W1. The quantities in Eq. 7–8 (individual signal, signed redundancy, synergy) are never estimated on the real readers. … Appendix C concedes that no new theory is introduced, which to me sits uneasily with "we prove" and "principled basis".

We estimated these terms on the real readers. The cross-validated gain is the plug-in estimate of the additive-class gain in Proposition 8. Its value against the length-only model estimates the individual signal I(Y;Zr∣U)I(Y;Z_r\mid U). Length alone is uninformative here (AUROC 0.49, Table 1). The drop from that value to the gain given the selected set estimates the signed redundancy. One outer split follows.

ReaderIndividual signal Δ^(r∣{U})\widehat\Delta(r\mid\{U\})Gain given {Qwen3-4B}Gain given the first three READ readers
Qwen3-4B (chosen 1st)0.168––
Llama-3.2-1B (chosen 2nd)0.1540.073–
gemma-3-4b-it (chosen 3rd)0.1330.062–
Ministral-3-3B (chosen 4th)0.1210.0580.034
Qwen3-0.6B (not chosen)0.1650.0450.027
Qwen3-8B (not chosen)0.1590.0410.025

Gains are reductions of inner-CV log loss in nats, pooled over settings as in selection. The length-only loss is 0.692. Along the path the loss falls to 0.524, 0.451, 0.405 and 0.371. Redundancy is the first column minus the later ones. It is 0.12 for the two Qwen readers after Qwen3-4B, and 0.08 for Llama-3.2-1B. Qwen readers lose about three quarters of their value once Qwen3-4B is in the set. Llama-3.2-1B keeps about half. By Eq. 19–21, these estimates equal the terms of Eq. 7–8 up to the change in approximation gap and fitting error. We cannot remove those terms on finite data. The ranking checks above show they do not change the ordering of candidates.

On wording, the identity in Theorem 4 is standard (Brown et al., 2012), as Appendix C says. Our use of it is to turn reader choice into a measurable quantity for this setting. We changed "We prove" to "We show", right after the citation of Brown et al., and replaced "provides a principled basis" with "gives a criterion".

W1. Under benchmark holdout, a fixed set of four small Qwen readers reaches 0.858 AUROC and 0.471 log loss versus 0.862 and 0.486 for READ.

E2. CIs for the transfer comparisons in Table 15, especially READ versus the fixed set.

You are right that the selection rule adds little under benchmark holdout. The paired 95% CIs (1,000 source-question cluster bootstraps) for the differences in Table 15 follow.

READ minusTarget-family holdoutBenchmark holdout
Individual signal+0.010 [0.004, 0.016]+0.005 [−0.001, 0.012]
Random set+0.014 [0.009, 0.019]+0.010 [0.005, 0.015]
Fixed set+0.013 [0.006, 0.021]+0.004 [−0.004, 0.012]

Under target-family holdout, READ leads all three rules. Under benchmark holdout, its intervals against the fixed set and individual-signal selection include zero. When the test task is new, inner-CV gains on other tasks are a weaker guide. What still transfers is the use of several readers. K = 1 gives 0.751 and K = 4 gives 0.862 under benchmark holdout (Table 15).

Changes in the paper.

  • Section 1. The two wording changes quoted above.
  • Section 4.2. New paragraph "Estimated Terms", which states that the cross-validated gain estimates the individual signal and redundancy, and gives the Qwen and Llama example above.
  • Section 5. New paragraph "Selection and Reader Budget" with the random-K and exhaustive-search results, and new paragraph "Transfer Limits" with the benchmark-holdout result.
  • Appendix F.1. Tables 19 (random-K) and 20 (selection log), and the exhaustive-search and ranking checks. Appendix F.6. Table 27 (transfer CIs).
  • Appendix C (Limitations). We added this text. "When the evaluation benchmark is absent from training, conditional-gain selection adds little over individual-signal selection or a fixed reader set, and its log loss is higher. The gain from combining several readers remains in this setting. Selecting readers for a new task may therefore need a small labeled sample from that task."

AuthorsOpenReview ↗

Response to Reviewer gfHV (3/5)

W2. LLM-Check is run as a training-free scorer. The gap between LLM-Check (0.760) and READ (0.911) therefore mixes supervision, number of readers and feature family.

E1. Supervised LLM-Check-style features fed to the same logistic combiner, for K=1 and for READ's four readers.

The +0.116 margin is measured against ELECTRA-large, not LLM-Check. ELECTRA-large is a supervised classifier. It uses the same labels and splits as READ, with a capacity search (Section 5, Appendix D.1). Supervision is therefore already matched in the headline comparison. The four-encoder stack also matches the number of models, and READ still leads it by 0.086 [0.076, 0.097] (Appendix D.5). The abstract said "the strongest published baseline", which invites this reading. It now says "the strongest evaluated baseline".

For LLM-Check itself, we ran the experiment you suggested. We extracted its five scores (hidden, attention, perplexity, logit entropy, windowed entropy) and fed them to READ's combiner with the same splits.

FeaturesReadersAUROC
LLM-Check, training-freeQwen3-4B0.760
LLM-Check features, supervisedsame 10.785
LLM-Check features, supervisedREAD's 40.840
READ featuresREAD's 40.911
READ + LLM-Check featuresREAD's 40.910

Supervision adds 0.025 and four readers add 0.055. The largest step, 0.071 [0.062, 0.080], comes from READ's evidence, mainly the hidden-state probe. The likelihood-only stack (0.756, Table 2d) tells the same story. LLM-Check features add nothing once READ's features are present.

W2. "Strictly less access" than white-box detectors holds for target access only. READ uses labels pooled across all targets and the internals of four additional models, whereas white-box probes are fitted per target.

E5. White-box baselines trained on the same pooled data as READ.

"Strictly less access" refers to the target only, and we agree the phrase was imprecise. We replaced it with "without access to the target" in the abstract and "with no target access" in Section 5. Black-box in our title also refers to the target, as in LLM-Check's black-box setting with an auxiliary model. Definition 1 states this.

A probe on target hidden states cannot be pooled across targets. The four targets have different hidden sizes, layers and tokenizers (Section 2). The only pooled version is to let each target read the responses of all six targets and train its probe on all of them. We ran exactly this, with the READ single-reader setup (same depths, answer-span pooling, probe, likelihood features and splits). Evaluated on its own responses, such a target reaches 0.853 on the 28 open-weight cells, against 0.830 when the other targets read the same responses and 0.907 for READ. Per-target results are in our reply to nNU5.

Pooling labels across targets does not explain the gain either. Under target-family holdout, READ sees no label from the evaluated target's family. It still reaches 0.869 on the 28 open-weight cells, on par with SAPLMA (0.871), which is fitted on that target's own labels.

E5. A repeated-view control with equalized features (for example score-only stacks for both conditions).

We refit all three variants of Table 2b as score-only stacks. Each stack uses the probe scores, answer length and score dispersion, with no likelihood features.

Score-only stackAUROC
One probe, Qwen3-4B0.747
Four resampled probes, Qwen3-4B0.780
Four layers, Qwen3-4B0.815
Four distinct readers0.871

Over one probe, resampled probes add 0.033, four layers add 0.068 and four distinct readers add 0.124. The ordering is unchanged. Distinct readers lead four layers by 0.056 [0.048, 0.064], a little more than with full features (0.047).

Changes in the paper.

  • Abstract. "published baseline" → "evaluated baseline", and "Despite strictly less access" → "Without access to the target". Section 5. "with strictly less access" → "with no target access".
  • Section 5. New paragraph "Comparison with LLM-Check Features" with the decomposition above, and new paragraph "Label Pooling" with the target-family holdout result.
  • Section 3.2. New paragraph "Equal Features" with the score-only result.
  • Appendix F.2. Targets as readers (Tables 21–22). Appendix F.3. Tables 23 (LLM-Check features) and 24 (score-only views). Appendix F.6. The 28-cell holdout result.

Reviewer_nNU5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes READ, a supervised framework for non-intrusive black-box hallucination detection that does not require access to the target model's internal states or additional target queries. The authors find that reader size, answering ability, and model-family similarity do not reliably predict hallucination detection performance. Based on this observation, READ selects four complementary reader models using cross-validated log-loss reduction and combines their hidden-state probe scores and token likelihood features through logistic regression. Experiments across 6 target LLMs and 7 benchmarks achieve an average AUROC of 0.911, outperforming existing baselines under the same access constraints by 0.116.

Critical strengths

The paper provides a comprehensive empirical analysis of reader suitability and complementarity. I particularly appreciate the extensive experiments and ablation studies, which provide useful insights into hallucination detection.

Critical weaknesses
  1. The proposed ensemble method mainly combines existing techniques, including linear probing, greedy feature selection, and logistic regression. While this is not necessarily a weakness, it does not add much to the novelty of the paper.

  2. The authors show that model size, answering ability, and family match do not reliably predict detection performance. However, the paper provides limited explanation for why these mismatches occur. In particular, I would be interested in understanding how reader–target similarity affects detection performance. For example, how does using the target model itself as a reader compare with using other models under the same probing setup?

Questions for authors
  1. Could the authors provide more insights into why model size, answering ability, and family match do not reliably predict hallucination detection performance? What factors actually determine reader suitability?
  2. Could the authors further investigate how reader–target similarity affects detection performance? In particular, how does using the target model itself as a reader compare with using other models under the same probing setup?
2 comments
AuthorsOpenReview ↗

Response to Reviewer nNU5

Thank you for the positive assessment and the focused questions.

W1. The proposed ensemble method mainly combines existing techniques, including linear probing, greedy feature selection, and logistic regression.

Yes, by design. Each part is simple so that the gain can be traced to its source. A 4-unit MLP on the same evidence gives the same 0.911 as the additive model (Table 2d). The gain therefore comes from what the readers supply and which readers are combined, not from the classifier. The contribution we claim is the NI-BHD setting, the finding that common reader-selection rules fail while readers complement each other, and a selection rule that follows from this finding.

Q1. Why do model size, answering ability, and family match not reliably predict detection performance? What factors actually determine reader suitability?

Size and answering ability describe how well a reader generates answers. Detection asks something else. It asks whether the reader's states separate correct from hallucinated answers of a given target on a given task. We examined this gap with three analyses, one of them new.

  • Answering is not needed. On the 168 responses that all 14 readers fail to answer, recognize and verify, READ still reaches 0.907 pooled AUROC (Appendix E.2).
  • Suitability depends on the setting. The reader–setting interaction explains 69.9% of the variance of the single-reader matrix, and which reader is used explains only 10.4% (Appendix D.5). A fixed property such as size cannot capture a quantity that changes this much across settings.
  • Inside a set, what counts is new information. We analyzed the selection logs for this question. In one outer split, Qwen3-0.6B has the second-largest individual gain (0.165 nats of log loss). After Qwen3-4B is selected, its gain drops to 0.045. Llama-3.2-1B starts lower (0.154) but keeps 0.073, so it is chosen next. The two Qwen readers make similar errors. Llama-3.2-1B makes different ones.

So the factor that decides suitability is the label information a reader adds in a given setting, beyond the readers already chosen. This is the quantity READ estimates.

Changes in the paper. We added this explanation, with the Llama-3.2-1B example, to Section 3.1 as the new paragraph "What Decides Suitability". Llama-3.2-1B has the lowest recognition accuracy (Table 13) yet wins 12 of 42 environments.

Q2. How does using the target model itself as a reader compare with using other models under the same probing setup?

We ran a new experiment for this question. We used each of the four open-weight targets (24B to 70B) as a reader, with exactly the READ single-reader setup (same depths, answer-span pooling, probe, likelihood features and splits). Like READ's readers, each target reads all 7,560 responses and its probe is trained on all of them. Results on the 28 open-weight cells follow.

Reader(s) on the 28 open-weight cellsAUROC
Same-family small reader (Table 2a)0.767
Best small single reader, Qwen3-4B0.825
Other open-weight targets as readers (mean of 3)0.830
The target reading its own responses0.853
Oracle best small reader per cell0.872
READ, four small readers0.907
Target (mean over 7 benchmarks)Self-readingOther targets as readersQwen3-4BSAPLMA (target states)READ
Llama-3.3-70B0.8890.8480.8090.8850.905
Qwen3-32B0.8130.8350.8430.8610.907
Mistral-Small-24B0.8830.8190.8330.9040.915
Gemma-3-27B0.8280.8180.8170.8340.901

Being the generating model helps on average. Self-reading beats the other large models on the same responses by 0.023, but not for Qwen3-32B, whose responses the other targets read better. Family match does not carry this benefit, since same-family small readers reach only 0.767. Self-reading stays below SAPLMA on three targets, by up to 0.048 on Qwen3-32B. SAPLMA fits an MLP probe per target with a capacity search per dataset (Appendix D.1), while self-reading uses READ's linear probe pooled over all targets. Four small readers from outside the target exceed self-reading by 0.054, and they need no access to the target.

Changes in the paper. We added both tables to the new Appendix F.2 (Tables 21 and 22) and a summary of the 28-cell result to Section 3.1 as the new paragraph "Target as Its Own Reader".


Thank you for your support of the paper. Your feedback was helpful, and the paper has improved because of it. If our reply resolves your concerns, we would appreciate it if you could consider a higher score. We are happy to provide more details if needed.

Comments on the paper

6 comments
AuthorsOpenReview ↗

General response to all reviewers and the AC (2/2)

Where each concern stands

We ran twelve new experiments and analyses, including a new target model, GPT-6 Astra, and four large open-weight models as readers. Every concern is now resolved by direct evidence.

  • Reader selection adds little beyond stacking (gfHV). Resolved. From K = 3 on, READ matches random sets with about two more readers, so it saves reader passes at the same accuracy. Only 6 of all 1,001 four-reader sets beat it, and finding them needs evaluation labels.
  • The comparisons are not like-for-like (gfHV). Resolved. The +0.116 margin is over ELECTRA-large, a supervised classifier with the same labels and splits. With LLM-Check features in READ's own combiner and the same four readers, READ still leads by 0.071 [0.062, 0.080]. A target reading its own responses with pooled training reaches 0.853, against 0.907 for READ.
  • Reader variability may be sampling noise (gfHV). Resolved. Under a bootstrap-calibrated null without reader–setting interaction, mean rank concordance is 0.32 [0.24, 0.40], against 0.118 observed. Noise explains only 29% of the interaction variance.
  • The theory is not new, and its terms are never estimated (gfHV). Resolved. We estimated individual signal and redundancy on the real readers, and the wording in Section 1 now matches the scope stated in Appendix C.
  • Model attributes fail without explanation, and the target as reader is untested (nNU5). Resolved. We explain what decides suitability, with evidence, and ran all four open-weight targets as readers. Four small readers beat each target reading its own responses.
  • Cost is not compared (s1hd, gfHV). Resolved. READ needs no target call, uses about 1/18 of the FLOPs of SelfCheckGPT on Llama-3.3-70B, and takes 84 ms per response on one H100.
  • Stronger models may be harder (s1hd). Resolved. On GPT-6 Astra, with no GPT-6 label in training, READ reaches 0.897 AUROC, against 0.809 for LLM-Check.
  • Generalization beyond the training data is limited (s1hd, gfHV). Resolved, and the limit is now stated. Under benchmark holdout READ keeps its lead over all baselines, and its gap over ELECTRA-large grows from 0.116 to 0.118. Under target-family holdout it reaches 0.875. The selection rule itself adds little when the benchmark is new, and the paper now says so.

The central claims of the paper stand on direct evidence, and no concern from any reviewer remains open.

Revised PDF

  • The main text is now ten pages, with new paragraphs in Sections 3.1, 3.2, 4.2 and 5.
  • All new results are in the new Appendix F (Tables 19–29). Tables 1–18, all figures, equations and appendix labels keep their numbers.
  • Wording changed in the abstract, Section 1, Section 3.1, Section 5, Appendix C and Appendix D.1.
  • In Appendix E.3, the sentence "conditional-gain selection continues to make useful choices … under both forms of transfer" now ends "under target-family holdout. Under benchmark holdout, its advantage over individual-signal selection and the fixed set is small", to match the new intervals.

AuthorsOpenReview ↗

General response to all reviewers and the AC (1/2)

Thank you all for your careful reading. We have addressed every concern and every question that each of you raised. The table lists each point, what we did, and where it is in the revised PDF. Changes in the PDF are shown in blue.

ReviewerConcernWhat we didWhere in the revised PDF
nNU5W1. Novelty of the partsWe explain the design choice. Simple parts make the gain traceable, and a 4-unit MLP on the same evidence gives the same 0.911Table 2d
nNU5Q1. Why model attributes fail, and what decides suitabilityWe analyzed the selection logs and added an explanation with evidence§3.1 "What Decides Suitability"
nNU5Q2. The target as its own readerWe ran a new experiment with all four open-weight targets (24B to 70B) as readers§3.1, App. F.2 (Tables 21–22)
gfHVW1. What selection adds beyond stackingWe ran random-K curves with CIs, an exhaustive search over all 1,001 sets, estimates of signal and redundancy, and two ranking checks§4.2, §5, App. F.1 (Tables 19–20)
gfHVW1. "We prove" and "principled basis"We revised the wording§1
gfHVW2, E1. Comparison with LLM-CheckWe show the +0.116 margin is over a supervised baseline, and ran LLM-Check features in our own combiner§5, App. F.3 (Table 23)
gfHVW2. "Strictly less access"We revised the wordingAbstract, §5
gfHVW2, E5. White-box baselines on pooled dataWe ran the targets as pooled readers and re-scored target-family holdout on the open-weight pairs§5 "Label Pooling", App. F.2, F.6
gfHVE5. Repeated views with equal featuresWe refit all three variants as score-only stacks§3.2, App. F.3 (Table 24)
gfHVQ1. LLM-Check configurationWe report the full configurationApp. D.1
gfHVQ2. Distractors and correlations with CIsWe describe the construction and added CIs for all five correlations§3.1, App. F.5 (Table 26)
gfHVQ3. Baseline tuningWe state the tuning protocolApp. D.1
gfHVE4. Noise floor for reader variabilityWe built a bootstrap-calibrated null model and redid Fig. 3d by environment§3.2, App. F.4 (Table 25)
gfHV, s1hdE2, W2. CIs for the transfer resultsWe computed paired intervals and added the limitation§5, App. C, App. F.6 (Table 27)
gfHV, s1hdQ4, W1. CostWe measured FLOPs, memory and latency on our 4×H100 server§5, App. F.8 (Table 29)
s1hdW2. Generalization beyond training dataWe report transfer results and their boundary§5, App. C
s1hdW3. Stronger models such as GPT-5/6We added GPT-6 Astra as a seventh, unseen target, with new generations, labels and a human label check§5, App. F.7 (Table 28)
AuthorsOpenReview ↗

Response to Reviewer s1hd (follow-up)

Thank you for the follow-up. We measured both baselines under the same setting as READ. The setting is one H100, bf16, one response at a time, the same mean input of 186 + 61 tokens, and the median over 200 responses. We also measured READ at smaller reader budgets, so cost and accuracy can be read side by side. K = 1 to 4 add Qwen3-4B, Llama-3.2-1B, gemma-3-4b-it and Ministral-3-3B, in the order READ selects them. AUROC is the cell-macro value over the 42 target–benchmark pairs (Tables 1 and 9). FLOPs use the same nominal length for every model. Memory is the loaded weights, and peak memory during a pass adds at most 0.4 GB.

MethodAUROCFLOPsWeights in GPU memoryLatency (1 GPU)
ELECTRA-large0.7950.17 T0.7 GB6 ms
LLM-Check (Qwen3-4B)0.7602.0 T8.0 GB25 ms
READ, K = 10.8162.0 T8.0 GB26 ms
READ, K = 20.8702.6 T10.5 GB35 ms
READ, K = 30.8964.5 T19.1 GB71 ms
READ, K = 40.9116.2 T26.8 GB84 ms

The table shows three things.

  • At equal cost, READ is more accurate than LLM-Check. LLM-Check runs one forward pass of Qwen3-4B and reads its attention maps, so it costs the same as one READ reader. READ with one reader reaches 0.816 against 0.760. With two readers, at 2.6 T and 35 ms, it reaches 0.870.
  • The accuracy gain over ELECTRA-large costs little in absolute terms. ELECTRA-large is a 335M-parameter text encoder, so it is the lightest method. The gap to READ at K = 4 is about 6 T FLOPs and 78 ms per response on one GPU. For this, READ gains 0.116 AUROC in domain and 0.118 under benchmark holdout (Table 16), and it reaches 0.897 against 0.692 on GPT-6 Astra. With one reader, READ is still more accurate than ELECTRA-large on average (0.816 against 0.795), and the gap in latency is 20 ms. ELECTRA-large must also be fine-tuned, which took about 9 GPU-hours with its capacity search. READ keeps its readers frozen and needs about 0.3 GPU-hours.
  • All three methods run in real time. Each stays under 100 ms on one GPU and needs no target call. SelfCheckGPT on Llama-3.3-70B needs 115 T and 2.0 s on four GPUs.

The reader budget K sets the trade-off. On the cell-macro average, READ is above both baselines at every budget, and each added reader has a known cost.

Changes in the paper. In Appendix F.8, we added ELECTRA-large and LLM-Check to Table 29, and added the cost–accuracy table above as the new Table 30 with a short discussion.

You called our rebuttal thorough, and your follow-up raised one further question, which the table above answers. We believe every concern in your review is now resolved, and we would be grateful if you would raise your score. We wish you all the best in your work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.