acceptodds
Under review as a conference paper at ICLR 2027

Max-pooling Network Revisited: Analyzing the Role of Semantic Probability in Multiple Instance Learning for Hallucination Detection

Abstract

Hallucination detection is essential for reliable large language models (LLMs). Hybrid methods such as HaMI combine semantic consistency with internal states through multiple instance learning (MIL), but repeated sampling and semantic comparisons impose substantial computational overhead. We analyze HaMI's decision margins and identify conditions under which semantic-probability scaling enlarges the average logit margin. Motivated by this analysis, we revisit classical sentence classification from a local margin dynamics perspective: a shared MLP extracts token-level features, feature-wise max pooling aggregates them, and a linear head scores the response. The detector requires no additional sampling or semantic consistency computation. Across five benchmarks and three LLMs, it is competitive with semantically weighted HaMI in detection AUC while sustaining several thousand QA detections per second. Its larger gains over fixed-position probes on LongFact support aggregating evidence across token positions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
hallucination detection, large language models, multiple instance learning, max pooling, margin analysis, semantic uncertainty, hidden state probing

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_P81fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper revisits feature-wise max pooling for hallucination detection using LLM hidden states. The authors analyze semantic-probability weighting in HaMI and provide theoretical results on margin improvement and pooling behavior. Experiments across five benchmarks and three LLMs show that max pooling outperforms mean pooling and achieves competitive detection AUC without additional sampling.

My main concern is that the theoretical analysis does not fully explain the observed detection improvements. The margin results rely on restricted conditions and do not establish better AUC after training. Moreover, the strong LongFact results rely on automatically generated factuality labels whose reliability has not been independently verified on that benchmark.

Critical strengths
  1. Strong empirical results with useful architectural ablations.
  2. Competitive detection performance without additional sampling.
Critical weaknesses
  1. Limited theoretical explanation for the empirical advantage of max pooling. Theorem 2 establishes a general property of logistic-loss optimization rather than a specific advantage of max pooling. Theorem 3 shows that max pooling avoids gradient dilution under a restrictive sparsity assumption, but the paper does not directly verify whether this assumption holds for the learned token features. More importantly, larger gradients do not necessarily lead to better hallucination detection AUC, and the experiments do not establish that the proposed mechanism accounts for the observed improvements. Given that max pooling is an established architecture, this limits the theoretical contribution of the paper.
  2. Insufficient validation of the LongFact evaluation labels. The strongest experimental results are obtained on LongFact, where factuality labels are assigned by GPT-5 mini without reference answers or external verification. Although the authors report 95.3% agreement with human annotations on TruthfulQA, this does not establish label reliability for long responses containing multiple factual claims. Furthermore, the 1024-token experiments contain only 69 and 35 hallucinated training examples for the 12B and 70B models, respectively. Without independent validation of the LongFact labels, the reported AUC improvements remain difficult to assess as evidence of genuine factual error detection.
Questions for authors

Is there empirical evidence that the assumed sparse-feature structure holds in the trained detectors and explains the AUC improvement?

Do the LongFact AUC improvements persist when evaluated against independently verified factuality labels?

Could the authors include comparisons with recent single-generation hallucination detectors, such as HaluNet, Simple Factuality Probes, HIDE and Seek, and SinkProbe? Evaluating detection AUC and inference cost under matched settings would help establish the competitiveness of the proposed method.

4 comments
AuthorsOpenReview ↗

Response to Reviewer P81f (Part1)

We thank the reviewer for the careful review and the useful suggestions. The main changes relevant to this review are:

  • The former Theorem 2 is now a supporting lemma (Lemma 2).
  • The sparsity assumption behind Theorem 3 is replaced by measurable quantities (Definition 1), and the theorem is verified on the trained detectors.
  • The long-form study now uses LongFact++, with retrieval-verified labels and class-balanced training data.
  • Four recent detectors are added, with matched AUC and end-to-end cost.

We address each point below.


W1.1

"Theorem 2 establishes a general property of logistic-loss optimization rather than a specific advantage of max pooling."

Thank you for this point; the result is indeed not specific to max pooling. It holds for any differentiable scorer trained with the logistic loss. Its role in the paper is only to show that pooling enters the first-order margin gain through the squared gradient norm ∥∇θzB∥2\|\nabla_\theta z_B\|^2, the quantity that Theorem 3 compares between max and mean pooling. In the revision we state it as a supporting lemma (Lemma 2) rather than a theorem. The pooling-specific statement is Theorem 3.


W1.2 & Q1

"Theorem 3 shows that max pooling avoids gradient dilution under a restrictive sparsity assumption, but the paper does not directly verify whether this assumption holds ... larger gradients do not necessarily lead to better hallucination detection AUC."

"Is there empirical evidence that the assumed sparse-feature structure holds in the trained detectors and explains the AUC improvement?"

Thank you for these questions. The sparsity assumption is removed and the result is verified on the trained detectors.

  • Definition 1. For feature jj of bag BB:
    • αB,j\alpha_{B,j} is the fraction of tokens on which the feature is active (uB,i,j>0u_{B,i,j}>0);
    • κB,j≥1\kappa_{B,j}\ge1 is the largest hidden-state norm among the active tokens, divided by the norm at the maximizing token.
  • Theorem 3 (revised). For any parameters and any bag, ∥∇θzBmax⁡∥2∥∇θzBmean∥2  ≥  (∑jπB,j αB,j2 κB,j2)−1,\frac{\|\nabla_\theta z_B^{\max}\|^2}{\|\nabla_\theta z_B^{\rm mean}\|^2}\;\ge\;\Big(\sum_j\pi_{B,j}\,\alpha_{B,j}^2\,\kappa_{B,j}^2\Big)^{-1}, where πB,j\pi_{B,j} is feature jj's share of the max-pooling gradient norm. A token on which a ReLU feature is inactive contributes neither to the pooled value nor to the gradient, so no assumption on the background tokens is needed.
  • Relation to the original theorem. The revised theorem generalizes the original one: under the original assumption (each feature active on at most sBs_B tokens, active-token norms in [g‾,g‾][\underline g,\overline g]), αB,j≤sB/TB\alpha_{B,j}\le s_B/T_B and κB,j≤g‾/g‾\kappa_{B,j}\le\overline g/\underline g, so the bound recovers Ω((TB/sB)2)\Omega\big((T_B/s_B)^2\big), now for any bag (Appendix B.5).
  • Verification (Appendix G, Table 36). We evaluated the bound with each bag's measured α\alpha, κ\kappa and π\pi on the trained detectors (twelve QA settings and LongFact++, five runs, all test bags).
QA (12 settings)LongFact++
Lower bound holdsevery bagevery bag
Median lower bound1.20–1.732.12–2.49
Measured gradient ratio (median)2.9–6.415–18
Median active fraction α\alpha0.71–1.000.46–0.54
  • Dilution by inactive tokens, measured directly (new). We compared mean pooling with mean pooling over only the tokens on which each feature is active. The ratio of the two squared gradient norms isolates the dilution by inactive tokens, which max pooling does not have. Its median is 1.4–2.4 on QA and 3.0–3.4 on LongFact++.

On AUC, we agree with the reviewer. A larger gradient does not by itself imply a higher AUC, and Theorem 3 does not claim it; AUC also depends on the test distribution. Lemma 2 links the gradient to the margin, since a bag's first-order margin gain is proportional to its squared gradient norm. We therefore checked each link from gradient to margin to AUC on max- and mean-pooling detectors for the four QA benchmarks.

LinkMax pooling vs. mean poolingSource
GradientSquared gradient norm 2.9–6.4× larger (median)Table 36
Margin gain per update (new)Batch-mean margin increases in 90–100% of mini-batches (mean pooling: 56–77%, epochs 5–50)One SGD step (n=128n=128) on 8 mini-batches per checkpoint; LLaMA-3.1-8B, 3 seeds
Margin after trainingLarger mean test margin in all twelve settings (by 13–16% on average per LLM)Table 5, Table 21
AUCHigher AUC in all twelve settings (average gains of 1.5–2.1 points)Table 2

Every link points in the same direction. We present this as empirical support for the mechanism, not as a proof that the gradient difference causes the AUC difference. The revision states this scope explicitly (Section 5.2).

AuthorsOpenReview ↗

Response to Reviewer P81f (Part2)


W2 & Q2

"Insufficient validation of the LongFact evaluation labels ... the 1024-token experiments contain only 69 and 35 hallucinated training examples."

"Do the LongFact AUC improvements persist when evaluated against independently verified factuality labels?"

Thank you for raising these concerns, which we share. Our original long-form study had two weaknesses that could not be fixed within its design: its labels came from a single LLM judge working from its own knowledge, and its longest cells had only a few dozen hallucinated training answers. We therefore replaced the study with LongFact++ (Obeso et al., 2026), whose labels are independently verified and which is large enough for class-balanced training.

Labels (Section 6.3, Appendix D.2).

  • Every entity in an answer is checked against web search by the dataset's annotation pipeline, not by our judge.
  • We label an answer hallucinated when at least 10% of its checked entities are not supported.
  • The dataset authors validated the pipeline (Obeso et al., 2026). Human annotators who checked a random sample of entity spans with web search agreed with it in 84% of cases (n = 50). On passages with injected factual errors, it detected 80.6% of the errors (729/904), with a 15.8% false-positive rate on unchanged entities.

Training data (Appendix D.2, Table 13). For each of LLaMA-3.1-8B-Instruct and LLaMA-3.3-70B-Instruct (median answer about 740 tokens), we use 1,000 class-balanced training answers (500 hallucinated) and an unweighted loss. Validation has 1,000 answers and test 1,500 at the natural class rates. All methods were retrained under the same protocol.

Answer to Q2: yes, the improvements persist on these labels (Section 6.3, Table 3).

Method8B AUC8B AUPRC70B AUC70B AUPRC
SEP0.4930.6650.5340.469
LLMs Know0.6230.7730.6330.558
SAPLMA0.6470.7740.6370.581
HaMI0.7110.8270.6650.602
HaMI (SP)0.7110.8280.6630.598
Mean0.7590.8640.7390.689
Max0.7830.8810.7390.691

8B: LLaMA-3.1-8B-Instruct; 70B: LLaMA-3.3-70B-Instruct. Mean of five runs. Chance AUPRC equals the hallucination rate (0.674 for 8B, 0.449 for 70B).

  • The four MIL detectors outperform the three fixed-position probes on both models.
  • Among the methods of Table 3, max pooling is best on LLaMA-3.1-8B-Instruct (AUC 0.783).
  • On LLaMA-3.3-70B-Instruct, max pooling ties mean pooling (0.739). It exceeds HaMI by 7.4 points and the fixed-position probes by 10.2–20.5 points.
  • Answer length alone reaches only 0.59 and 0.52 AUC (Section 6.3).
AuthorsOpenReview ↗

Response to Reviewer P81f (Part3)


Q3

"Could the authors include comparisons with recent single-generation hallucination detectors, such as HaluNet, Simple Factuality Probes, HIDE and Seek, and SinkProbe? Evaluating detection AUC and inference cost under matched settings ..."

Thank you for the suggestion. We added all four detectors: HaluNet (Tong et al., 2026), Simple Factuality Probes (SFP; Han et al., 2025), HIDE (Chatterjee et al., 2025) and SinkProbe (Binkowski et al., 2026) (Appendix E.4).

Setting.

  • Data and layer search. Same data, splits and labels as max pooling, and the same layer search: every stored layer, validation-selected, five runs.
  • SinkProbe and HIDE. SinkProbe uses all layers and HIDE the middle layer, as in their original definitions, with their hyperparameters selected on validation.
  • Inference cost. Measured end to end, from the prompt to the score including generation, on the same generated answers (Appendix E.1).
MethodQA AUC 8B / 12B / 70B (mean of 4 benchmarks)LongFact++ AUC 8B / 70BTime overhead vs. generation (median per response, QA)
Max pooling0.866 / 0.897 / 0.9000.783 / 0.739+0.5%
HaluNet0.852 / 0.887 / 0.8860.770 / 0.738+0.9%
SFP0.818 / 0.862 / 0.8850.635 / 0.632+0.4%
SinkProbe0.818 / 0.864 / 0.8780.765 / 0.734+30.8%
HIDE0.560 / 0.538 / 0.5400.500 / 0.534+8.1%
  • AUC. Among these four recent detectors, max pooling has the highest mean AUC on every generator and on both LongFact++ models (Tables 27 and 28). Per setting, it matches or exceeds every recent detector in 10 of the 12 QA settings; the exceptions are two settings where HaluNet is 0.6 and 1.2 points higher (Table 27).
  • Cost. Max pooling, SFP and HaluNet add under 1%. HIDE needs keyword selection with a sentence encoder (+8.1%). SinkProbe must generate with eager attention to obtain attention maps (+30.8% on QA, +15–28% on LongFact++).
  • Max pooling on long answers. It needs no stored states: the work after generation stays below 1 ms, while HaluNet and HIDE need up to 1.3 s and 0.9 s on LLaMA-3.3-70B-Instruct (Table 30).

References

  • Binkowski, J., Adamczewski, K. & Kajdanowicz, T. (2026). Attention sinks as internal signals for hallucination detection in large language models. ICML.
  • Chatterjee, A., Goel, Y. & Chakraborty, T. (2025). HIDE and Seek: Detecting hallucinations in language models via decoupled representations. arXiv:2506.17748.
  • Han, J., Band, N., Razzak, M., Kossen, J., Rudner, T. G. J. & Gal, Y. (2025). Simple factuality probes detect hallucinations in long-form natural language generation. Findings of EMNLP.
  • Obeso, O., Arditi, A., Ferrando, J., Freeman, J., Holmes, C. & Nanda, N. (2026). Real-time detection of hallucinated entities in long-form generation. arXiv:2509.03531.
  • Tong, C., Zhang, Q., Jiang, Z., Jiang, L. & Liu, Y. (2026). HaluNet: Learning hallucination risk from internal signals in LLM question answering. arXiv:2512.24562v2.
AuthorsOpenReview ↗

Following Up on Our Response to Your Review

Dear Reviewer P81f,

Thank you again for your careful review. In case it is helpful, here is a brief summary of our response.

  • W1 / Q1 (theory): Theorem 3 is restated without the sparsity assumption and verified on the trained detectors. We present the gradient-to-AUC link as empirical support, not proof of causation.
  • W2 / Q2 (long-form labels): We replaced the study with LongFact++ (retrieval-verified labels). Max pooling is best on LLaMA-3.1-8B and ties mean pooling on LLaMA-3.3-70B.
  • Q3 (recent detectors): We added HaluNet, SFP, HIDE, and SinkProbe. Max pooling matches or exceeds them in 10 of 12 QA settings.

We would be grateful to hear whether these address your concerns, or if anything needs further clarification.

Best regards,

The authors

Reviewer_WFo5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper revisits a max-pooling network for hallucination detection. It applies a learnable feature transformation to token-level hidden states, performs feature-wise max pooling across tokens, and predicts whether the response contains hallucinations. The authors analyze the role of semantic probability weighting in HaMI and the gradient properties of max pooling under sparse-signal assumptions, and evaluate the method on short- and long-form responses. The results show competitive in-distribution AUROC without additional sampling at inference time, along with low computational overhead for the detection module.

Critical strengths
  1. The method is simple and practically useful. The paper demonstrates that a lightweight feature transformation followed by max pooling can provide a competitive hallucination detection baseline, helping reassess the necessity of more complex components.

  2. The in-distribution comparisons and architectural ablations are reasonably thorough. Methods share the same generated responses, labels, and data splits. The ablations further show that the order of feature transformation and pooling substantially affects performance, providing direct evidence for the architectural choice.

  3. The limitations are discussed carefully. The authors distinguish local gradient analysis from actual training behavior and acknowledge that response length, content, and label distributions change simultaneously in the long-form experiments.

Critical weaknesses
  1. The incremental research contribution needs further clarification. The underlying architecture has a long history, and RAGLens also applies feature-wise max pooling to token features, although its feature source, training procedure, and task setting differ. The ablations support the design choices, but the paper should more clearly identify the new, testable insights it establishes beyond existing feature aggregation methods. The theoretical results primarily rely on specific assumptions or characterize local updates, and their connection to the observed detection gains remains limited.

  2. The validity of the long-form labels lacks direct verification. LongFact labels are generated by a single model without reference answers or external retrieval. The label validation on TruthfulQA does not directly establish the reliability of this procedure for long-form responses. It is therefore difficult to determine how much the reported gains reflect the detection of actual factual errors rather than agreement with the annotating model. The authors acknowledge this limitation, but it still weakens the evidence from the long-form experiments.

  3. The in-distribution results do not sufficiently establish the method’s scope of applicability. The current experiments primarily train and evaluate within the same data distribution, leaving unclear whether the method’s performance relative to HaMI carries over to new data sources. Since the paper emphasizes competitive performance with a simpler approach that requires no additional sampling, and HaMI already reports cross-dataset evaluations, the absence of corresponding comparisons limits the evidence for this trade-off.

Questions for authors

When training, layer selection, and hyperparameter selection use only source data, how does the proposed method compare with HaMI and HaMI(SP) across datasets? Under what conditions does it retain its favorable trade-off between detection performance and computational overhead?

6 comments
AuthorsOpenReview ↗

Response to Reviewer WFo5 (Part1)

We thank the reviewer for the constructive comments. The main changes relevant to this review are:

  • The related work now positions our detector against RAGLens.
  • Theorem 3 is restated without a sparsity assumption and verified on the trained detectors; the former Theorem 2 is now a supporting lemma (Lemma 2).
  • The long-form study now uses LongFact++, with retrieval-verified labels.
  • A leave-one-dataset-out evaluation is added.

We address each point below.


W1

"The incremental research contribution needs further clarification. ... RAGLens also applies feature-wise max pooling to token features ... the paper should more clearly identify the new, testable insights it establishes beyond existing feature aggregation methods. The theoretical results primarily rely on specific assumptions or characterize local updates ..."

Thank you for this comment. We agree that feature-wise max pooling is not new; the architecture goes back to max-over-time sentence classification (Kim, 2014). We now cite RAGLens (Xiong et al., 2026) in the related work (Section 2) and state the difference. RAGLens pools fixed sparse-autoencoder features, selects them by mutual information and classifies them with an additive model, for faithfulness to retrieved context. We instead learn the token features end to end from response-level labels, and we study factuality without a retrieved context. Our contribution is not the architecture but the following findings, each of which can be tested and is tested in the paper.

  1. Without sampling, feature-wise max pooling matches semantically weighted HaMI. On four QA benchmarks and three LLMs, a single-pass detector reaches the accuracy of HaMI (SP) (mean AUC 0.888 vs. 0.883; Table 2) while adding a median 0.5% to response time, against 7–34× for HaMI (SP) (Section 6.4).
  2. Where the features are pooled matters for max pooling, not for mean pooling. Pooling the raw hidden states instead of learned features lowers max pooling by 0.7–5.7 points, while mean pooling changes by at most 0.2 points (Table 6). The gain therefore comes from max pooling over learned features, not from max pooling.
  3. A measurable account of why max pooling keeps the learning signal. In the revision, Theorem 3 no longer assumes inactive background features: it is stated through measurable quantities (Definition 1, Section 5.2) and holds for any parameters and any bag. It lower-bounds the max/mean gradient ratio by (∑jπB,jαB,j2κB,j2)−1\big(\sum_j\pi_{B,j}\alpha_{B,j}^2\kappa_{B,j}^2\big)^{-1}, where αB,j\alpha_{B,j} is the fraction of tokens on which feature jj is active; this bound can be measured on a trained detector. The former Theorem 2, which concerned local updates, is now a supporting lemma (Lemma 2).
    • It generalizes the original Theorem 3 and keeps its motivation and claim: under the original sparsity assumption the bound recovers the original Ω((TB/sB)2)\Omega\big((T_B/s_B)^2\big) ratio, now for any bag rather than only positive ones (Appendix B.5).
    • On all twelve QA settings and on LongFact++ (five runs, all test bags), the lower bound holds for every bag.
    • Its median is 1.20–1.73 on QA and 2.12–2.49 on LongFact++, where features are active on only about half the tokens (Appendix G).
    • The prediction that longer answers make the bound larger is borne out.
  4. When semantic weighting helps HaMI. Theorem 1 gives a condition under which semantic-probability weighting enlarges HaMI's average margin, stated through the class ratio and a measurable sensitivity ratio γ\gamma. We measured γ\gamma in all twelve settings of the four QA benchmarks, and the condition holds in all of them (Appendix H).

On the scope of the theory: no pooling rule can be guaranteed to improve AUROC without assumptions on the test distribution, and we do not claim such a guarantee. The theorems concern the training signal: how pooling enters the margin gain, and how large that gain is for sparse features.

AuthorsOpenReview ↗

Response to Reviewer WFo5 (Part2)

W2

"The validity of the long-form labels lacks direct verification. LongFact labels are generated by a single model without reference answers or external retrieval ..."

Thank you for raising this concern, which we share. Our original long-form study had two weaknesses that could not be fixed within its design: its labels came from a single LLM judge working from its own knowledge, and its longest cells had only a few dozen hallucinated training answers. We therefore replaced the study with LongFact++ (Obeso et al., 2026), whose labels are obtained by external retrieval and whose size allows 500 hallucinated training answers per model.

Labels (Section 6.3, Appendix D.2).

  • Every entity in an answer is checked against web search.
  • We label an answer hallucinated when at least 10% of its checked entities are not supported.
  • The dataset authors validated these labels directly (Obeso et al., 2026). Human annotators who checked a random sample of entity spans with web search agreed with the pipeline in 84% of cases (n = 50). On passages with injected factual errors, the pipeline detected 80.6% of the errors (729/904), with a 15.8% false-positive rate on unchanged entities.
  • The labels therefore no longer reflect agreement with a single annotating model.

Setting (Appendix D.2, Table 13).

  • Answers of LLaMA-3.1-8B-Instruct and LLaMA-3.3-70B-Instruct, about 740 tokens at the median.
  • For each model, 1,000 class-balanced training answers, 1,000 validation answers and 1,500 test answers at the natural class rates.
  • All methods are retrained under the same protocol.

Results (Section 6.3, Table 3).

Method8B AUC8B AUPRC70B AUC70B AUPRC
SEP0.4930.6650.5340.469
LLMs Know0.6230.7730.6330.558
SAPLMA0.6470.7740.6370.581
HaMI0.7110.8270.6650.602
HaMI (SP)0.7110.8280.6630.598
Mean0.7590.8640.7390.689
Max0.7830.8810.7390.691

8B: LLaMA-3.1-8B-Instruct; 70B: LLaMA-3.3-70B-Instruct. Mean of five runs. Chance AUPRC equals the hallucination rate (0.674 for 8B, 0.449 for 70B).

  • The four MIL detectors outperform the three fixed-position probes on both models.
  • Among the methods of Table 3, max pooling is best on LLaMA-3.1-8B-Instruct (AUC 0.783; the paired bootstrap interval excludes zero against mean pooling and HaMI; Table 15).
  • Max pooling ties mean pooling on LLaMA-3.3-70B-Instruct (0.739), and exceeds the fixed-position probes by 10.2–20.5 points.
  • Answer length alone reaches only 0.59 and 0.52 AUC, so the gains are not explained by length.

W3 & Q1

"When training, layer selection, and hyperparameter selection use only source data, how does the proposed method compare with HaMI and HaMI(SP) across datasets?"

Thank you for the suggestion.

Leave-one-dataset-out on QA. We added a leave-one-dataset-out evaluation on the four QA benchmarks for all three LLMs (Appendix E.5, Table 31).

  • Protocol. Each benchmark is held out in turn, and the detectors are trained on the other three. Training, checkpoint selection, layer selection and hyperparameters use only source data. The held-out test split is the one of Table 2.
Mean AUC on the held-out benchmarkLLaMA-3.1-8BMistral-Nemo-12BLLaMA-3.3-70B
HaMI0.7570.8370.838
HaMI (SP)0.8570.8650.871
Mean0.7900.8500.848
Max0.8140.8660.873
  • Against HaMI and mean pooling. Max pooling exceeds both in all twelve held-out settings, by 1.8–9.0 points over HaMI.
  • Against HaMI (SP). Max pooling matches it on Mistral-Nemo-12B and LLaMA-3.3-70B (within 0.2 points on average), but trails it by 4.3 points on LLaMA-3.1-8B.
  • Why HaMI (SP) degrades least. Its semantic probability is computed from samples of the target response itself, so this part does not depend on the training distribution. Averaged over all settings, AUC falls relative to in-domain training by 1.9 points for HaMI (SP), 3.6 for max pooling, 4.0 for mean pooling and 4.6 for HaMI.

From QA to long-form answers (new). We also trained on all four QA benchmarks and tested on LongFact++, with the QA and LongFact++ answers read by the same model. The detectors that aggregate over all tokens transfer only partly: AUC is 0.604 for max pooling, 0.599 for mean pooling and 0.619 for HaMI on LLaMA-3.1-8B-Instruct, and 0.635, 0.646 and 0.634 on LLaMA-3.3-70B-Instruct, against 0.783 and 0.739 for max pooling trained on LongFact++. Fixed-position probes trained on a QA benchmark stay near chance (0.50–0.58). The two tasks differ in what the label measures: the correctness of one short answer versus the share of unsupported entities in a long answer. Detecting long-form hallucination therefore requires long-form training data, and aggregating over all tokens transfers better than reading a single position.

AuthorsOpenReview ↗

Response to Reviewer WFo5 (Part3)

Q2

"Under what conditions does it retain its favorable trade-off between detection performance and computational overhead?"

Thank you for this question. Max pooling scores a response during generation with a running maximum and adds a median 0.5% to the response time. HaMI (SP) generates four more answers and clusters them with LLM calls, at 7–34× the time of generation on QA (Section 6.4, Table 4). The trade-off is favorable in two conditions:

  1. In-domain: max pooling matches HaMI (SP) on average (0.888 vs. 0.883; Table 2).
  2. On long answers: semantic weighting brings no gain there (Table 3), while the extra sampling still costs more than five times the generation time (Table 20).

The trade-off is least favorable when a small generator meets a new dataset (LLaMA-3.1-8B in leave-one-dataset-out). There, the semantic probability adds information that the hidden states do not carry, and HaMI (SP) is worth its cost if latency and API calls are not a constraint.


References

  • Kim, Y. (2014). Convolutional neural networks for sentence classification. EMNLP.
  • Obeso, O., Arditi, A., Ferrando, J., Freeman, J., Holmes, C. & Nanda, N. (2026). Real-time detection of hallucinated entities in long-form generation. arXiv:2509.03531.
  • Xiong, G., He, Z., Liu, B., Sinha, S. & Zhang, A. (2026). Toward faithful retrieval-augmented generation with sparse autoencoders (RAGLens). ICLR.
Reviewer_xhsHOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper revisits feature-wise max pooling for hallucination detection from generated-token hidden states. Motivated by an analysis of semantic-probability weighting in HaMI, it applies a shared token-wise projection, pools each learned feature over token positions, and scores the resulting response representation. This avoids additional response sampling and semantic clustering at inference. The authors analyze margin behavior under per-bag updates and a sparse multiple-instance assumption, then report AUC, throughput, and ablations across four short-answer QA benchmarks and LongFact using three language models.

Critical strengths

The paper studies a meaningful accuracy–cost trade-off. Under the reported evaluation, max pooling exceeds matched mean pooling in all twelve short-answer model–dataset settings and leads the tested detectors in all six LongFact settings. It uses only the hidden states of the response being classified, while the semantically weighted HaMI variant requires additional sampling and comparison. The throughput measurements explicitly exclude initial response generation, which helps define their scope.

The feature-placement ablation is informative. Applying the token-wise transformation before max pooling improves AUC over pooling raw hidden states first, including a marked change on LLaMA-3.3-70B/TriviaQA (0.880 to 0.937; Table 6). This supports the particular architecture more directly than the capacity bounds alone.

Critical weaknesses

The theoretical analysis has a narrower scope than the empirical claim. The margin expansion result evaluates each bag after its own gradient update, not the shared detector after training, and the sparse-MIL result assumes inactive background features. The supplied detector additionally uses a positive-class-weighted loss, whereas the stated margin analysis uses an unweighted logistic loss. On LongFact, the 1024-token training sets for the two instruction-tuned models contain only 69 and 35 hallucinated responses, and factuality labels come from an automated judge rather than independent human annotation of these benchmarks. These factors limit how confidently the long-response results and proposed mechanism can be generalized.

Questions for authors

Was the positive-class weight of 0.3 in the supplied detector used for the reported results?

How sensitive are the LongFact AUC results to the small hallucinated training classes and automated factuality labels?

5 comments
AuthorsOpenReview ↗

Response to Reviewer xhsH (Part1)

We thank the reviewer for the careful review. The main changes relevant to this review are:

  • The sparsity assumption behind Theorem 3 is replaced by measurable quantities (Definition 1), and the theorem is verified on the trained detectors.
  • The former Theorem 2 is now a supporting lemma (Lemma 2).
  • The long-form study now uses LongFact++, with retrieval-verified labels and class-balanced training data.

We address each point below.


W1.1

"The margin expansion result evaluates each bag after its own gradient update, not the shared detector after training, and the sparse-MIL result assumes inactive background features."

Thank you for this careful assessment of the scope of our theory.

Assumptions in sparse-MIL
We agree that the assumption was stronger than necessary. We replaced it with a definition of measurable quantities and restated Theorem 3 without any assumption about background tokens (Section 5.2).

  • Definition 1. For a feature jj of bag BB:
    • αB,j\alpha_{B,j} is the fraction of tokens on which the feature is active (uB,i,j>0u_{B,i,j}>0).
    • κB,j≥1\kappa_{B,j}\ge 1 is the largest hidden-state norm among the active tokens, divided by the norm at the maximizing token.
  • Revised Theorem 3. For any parameters and any bag,
    ∥∇θzBmax⁡∥2∥∇θzBmean∥2  ≥  (∑jπB,j αB,j2 κB,j2)−1,\frac{\|\nabla_\theta z_B^{\max}\|^2}{\|\nabla_\theta z_B^{\rm mean}\|^2}\;\ge\;\Big(\sum_j \pi_{B,j}\,\alpha_{B,j}^2\,\kappa_{B,j}^2\Big)^{-1},
    where πB,j\pi_{B,j} is feature jj's share of the max-pooling gradient norm.
  • Why no background assumption is needed. A token on which a ReLU feature is inactive contributes neither to the pooled value nor to the gradient. With κ=1\kappa=1, the bound exceeds one as soon as some feature with vj>0v_j>0 is inactive on at least one token.
  • Relation to the original theorem. The revised theorem generalizes the original one and keeps its motivation and claim. Under the original assumption (each feature active on at most sBs_B tokens, active-token gradient norms in [g‾,g‾][\underline g,\overline g]), αB,j≤sB/TB\alpha_{B,j}\le s_B/T_B and κB,j≤g‾/g‾\kappa_{B,j}\le\overline g/\underline g, so the bound recovers Ω((TB/sB)2)\Omega\big((T_B/s_B)^2\big), now for any bag rather than only positive ones (Appendix B.5).
  • Validation (Appendix G, Table 36). The theorem gives a lower bound on the gradient ratio. We evaluated this lower bound with each bag's measured αB,j\alpha_{B,j}, κB,j\kappa_{B,j} and πB,j\pi_{B,j} on the trained detectors of all 12 QA settings and of LongFact++ (five runs, all test bags), and compared it with the measured ratio.
QA (12 settings)LongFact++
Lower bound holdsevery bagevery bag
Median lower bound1.20–1.732.12–2.49
Measured gradient ratio (median)2.9–6.415–18
Median active fraction α\alpha0.71–1.000.46–0.54

Activation is sparser on long answers, and the lower bound is larger there.

Per-bag updates
We agree that the margin result describes each bag under its own update, not the shared detector after training. Its role in the paper is to show that pooling enters the first-order margin gain through the squared gradient norm ∥∇θzB∥2\|\nabla_\theta z_B\|^2, the quantity that Theorem 3 compares between max and mean pooling, and we now state it as a supporting lemma (Lemma 2, Section 5.2). To see how far it carries over to shared training, we measured shared mini-batch updates directly (new):

  • Identity. Under a shared step on a mini-batch of nn bags, the margin of bag BB changes by ηn(aB+cB)+O(η2)\frac{\eta}{n}(a_B+c_B)+O(\eta^2), where aB=∥∇zB∥2/(1+emB)a_B=\|\nabla z_B\|^2/(1+e^{m_B}) is exactly the per-bag gain of Lemma 2 and cBc_B is the interference from the other bags in the batch.
  • Shared updates (new). One SGD step (n=128n=128, η=10−3\eta=10^{-3}) on eight random mini-batches per checkpoint (LLaMA-3.1-8B, four QA benchmarks, three seeds); ratios of medians over mini-batches:
EpochsPer-bag term ∑BaB\sum_B a_B, max / meanMini-batches with ∑BcB>0\sum_B c_B>0, max vs. meanBatch-mean margin gain, max / mean
0–22.5×86% vs. 82%2.5×
5–501.7×82% vs. 56%3.5×
1001.0×73% vs. 57%0.9×

For max pooling the interference mostly adds to the per-bag term (median ∑cB/∑aB\sum c_B/\sum a_B of 1.3 at epochs 5–50, against 0.3 for mean pooling). The gap closes at epoch 100, when max-pooling margins are already large and the factor 1/(1+emB)1/(1+e^{m_B}) of Lemma 2 is small. About one third of the individual bags lose margin in a typical step, so the identity is not a guarantee for every bag.

  • After training. Max pooling has a larger mean test margin than mean pooling in all twelve QA settings (Table 5, Table 21).

These results do not turn Lemma 2 into a statement about the trained detector, but they show that the per-bag term it isolates remains visible in shared training, and more so for max pooling.

AuthorsOpenReview ↗

Response to Reviewer xhsH (Part2)

W1.2 & Q1

"Was the positive-class weight of 0.3 in the supplied detector used for the reported results?"

Thank you for checking the supplied code.

The actual experiment used a positive-class weight of 1.0. The weight 0.3 found in the supplied detector was mixed in during an additional experiment. We uploaded a revised version of the supplementary materials; please refer to it for the details.

AuthorsOpenReview ↗

Response to Reviewer xhsH (Part3)

W1.3 & Q2

"How sensitive are the LongFact AUC results to the small hallucinated training classes and automated factuality labels?"

Thank you for this question.

Small hallucinated training classes. We agree that 69 and 35 hallucinated training answers were too few, and replaced the long-form study with LongFact++ (Obeso et al., 2026).

Training data (Appendix D.2, Table 13). For each of LLaMA-3.1-8B-Instruct and LLaMA-3.3-70B-Instruct (median answer about 740 tokens), we use 1,000 class-balanced training answers (500 hallucinated) and an unweighted loss. Validation has 1,000 answers and test 1,500 at the natural class rates. All methods were retrained under the same protocol.

Results (Section 6.3, Table 3).

Method8B AUC8B AUPRC70B AUC70B AUPRC
SEP0.4930.6650.5340.469
LLMs Know0.6230.7730.6330.558
SAPLMA0.6470.7740.6370.581
HaMI0.7110.8270.6650.602
HaMI (SP)0.7110.8280.6630.598
Mean0.7590.8640.7390.689
Max0.7830.8810.7390.691

8B: LLaMA-3.1-8B-Instruct; 70B: LLaMA-3.3-70B-Instruct. Mean of five runs. Chance AUPRC equals the hallucination rate (0.674 for 8B, 0.449 for 70B).

  • The four MIL detectors outperform the three fixed-position probes on both models.
  • Among the methods of Table 3, max pooling is best on LLaMA-3.1-8B-Instruct (AUC 0.783).
  • On LLaMA-3.3-70B-Instruct, max pooling ties mean pooling (0.739). It exceeds HaMI by 7.4 points and the fixed-position probes by 10.2–20.5 points.
  • Answer length alone reaches only 0.59 and 0.52 AUC (Section 6.3).

Automated factuality labels. The LongFact++ labels are not produced by our LLM judge. They come with the dataset: every entity in an answer was checked against web search, and an answer counts as hallucinated when at least 10% of its checked entities are not supported (Appendix D.2). The dataset authors validated this pipeline (Obeso et al., 2026): human annotators who checked a random sample of entity spans with web search agreed with it in 84% of cases (n = 50), and on passages with injected factual errors it detected 80.6% of the errors, with a 15.8% false-positive rate on unchanged entities. The long-form results therefore no longer rest on a judge that relies on its own knowledge.

Human-annotated long-form labels. We also trained and evaluated all detectors on long answers whose factuality was annotated by humans.

  • FActScore (Min et al., 2023): 493 biographies whose atomic facts were checked against Wikipedia. An answer is hallucinated when at least 10% of its facts are unsupported, the same rule as for LongFact++ (386 hallucinated).
  • SelfCheckGPT WikiBio (Manakul et al., 2023): 238 passages with sentence-level human labels. An answer is hallucinated when its mean sentence score exceeds 0.5 (125 hallucinated).

The answers were written by other LLMs and are read by the 8B and 70B models. We use stratified five-fold cross-validation (three folds for training, one for selection, one for testing; five seeds). SEP and HaMI (SP) need several sampled answers per prompt and are omitted.

AUCFActScore 8B / 70BWikiBio 8B / 70B
Max0.836 / 0.8440.870 / 0.831
Mean0.828 / 0.8400.867 / 0.857
HaMI0.814 / 0.8050.836 / 0.781
SAPLMA0.795 / 0.7940.722 / 0.733
LLMs Know0.789 / 0.7980.746 / 0.743
  • Max pooling outperforms HaMI and the fixed-position probes in all four settings. In a paired bootstrap over answers, its advantage excludes zero in all eight comparisons with SAPLMA and LLMs Know, and in three of four with HaMI.
  • It ties mean pooling except on WikiBio with 70B, where mean pooling is 2.2 points higher.

References

  • Manakul, P., Liusie, A. & Gales, M. J. F. (2023). SelfCheckGPT: Zero-resource black-box hallucination detection for generative large language models. EMNLP.
  • Min, S., Krishna, K., Lyu, X., Lewis, M., Yih, W., Koh, P. W., Iyyer, M., Zettlemoyer, L. & Hajishirzi, H. (2023). FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. EMNLP.
  • Obeso, O., Arditi, A., Ferrando, J., Freeman, J., Holmes, C. & Nanda, N. (2026). Real-time detection of hallucinated entities in long-form generation. arXiv:2509.03531.
Reviewer_DhitOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper revisits feature-wise max pooling for hallucination detection using LLM hidden representations. It first analyzes when semantic-probability scaling increases HaMI’s average signed logit margin. It then studies a simple detector consisting of a shared token-wise MLP, feature-wise max pooling, and a linear classification head. The theoretical analysis covers local margin dynamics under per-bag updates, gradient dilution under a sparse-MIL assumption, and Rademacher complexity bounds. Experiments span four short-answer QA datasets and LongFact using three LLMs. The detector consistently outperforms matched mean pooling on QA and achieves the highest reported mean AUROC in all six LongFact settings. The practical results and pooling-placement ablation are useful, but the theoretical explanation remains limited, and important evaluation and reporting issues need to be resolved.

Critical strengths

1.Relevant accuracy-cost trade-off. By aggregating hidden-state features from a single response, the method avoids the additional sampling, semantic clustering, and semantic comparison required by HaMI (SP), while retaining competitive detection performance. This makes the approach practically attractive for high-throughput hallucination detection. 2.Well-controlled comparison between mean and max pooling. The mean- and max-pooling models share the token-wise transformation, feature dimension, layer-search procedure, and tuning budget. Under this relatively controlled comparison, max pooling improves AUROC over mean pooling in all twelve QA model–dataset combinations, providing useful empirical evidence that feature-wise max aggregation can preserve sparse token-level signals.

Critical weaknesses

1.The theoretical results have limited explanatory power for the observed detection gains. Theorem 1 characterizes when a weighted difference of class-conditional logit shifts is positive. However, its sensitivity terms are measured from a trained network, and increasing the average signed margin does not establish improved ranking or AUROC. A uniform downward logit shift can increase the average margin in a factual-majority dataset while leaving AUROC unchanged. Furthermore, the fixed-network intervention analyzed by the theorem differs from comparing separately trained HaMI and HaMI (SP) models in Table 5. The integrated-sensitivity extension addresses changes in activation patterns, but does not close these conceptual gaps. Theorem 2 follows from a first-order expansion of logistic-loss descent for the same example that induces the update. It applies broadly to differentiable classifiers and is not specific to max pooling. The authors correctly state its limited scope, but its contribution to explaining the proposed architecture remains modest. 2.The reliability of the LongFact labels remains a major concern. LongFact is judged using GPT-5 mini’s internal knowledge without reference answers or retrieval. The 95.3% agreement on retained TruthfulQA examples is useful, but does not validate long-form factuality judgments, as the authors acknowledge. Long responses may contain a mixture of correct, incorrect, unverifiable, or incomplete claims, yet the precise response-level decision rule is insufficiently specified.

Questions for authors
  1. Theorem 2 analyzes each bag after its own gradient update. How should this be interpreted for actual training, where a shared parameter update is applied over minibatches and may improve some bags while hurting others? Can the authors provide evidence that the per-bag margin analysis predicts shared-model training dynamics? 2.Please clarify the train/validation/test split construction. Are questions disjoint across splits? When five responses per question are used in validation and test, are standard errors computed over random seeds only, or do they account for response correlation within the same question? 3.How robust are the results to the choice of automatic judge? For example, do the conclusions hold if labels are produced by a different LLM judge, retrieval-augmented verification, or human annotations on a subset of the main benchmarks? 4.Please provide paired, question-level bootstrap confidence intervals for the main AUROC differences. Given the class imbalance in LongFact, AUPRC and a threshold-based operating metric would also help assess practical usefulness? 5.Please report end-to-end detection cost separately from classifier-only throughput. Does the reported throughput include the cost of extracting hidden states from the generator, storing all token states, selecting the layer, and moving activations from GPU memory? 6.Since the method requires hidden states, how would it compare to other white-box supervised aggregators, such as attention pooling, learned TopK pooling, log-sum-exp pooling, DeepSets, Set Transformer pooling, or a small transformer over token hidden states? 7.How well does a detector trained on one dataset or one generator transfer to another dataset or generator? This would help assess whether the method learns general hallucination signals or dataset/model-specific artifacts. 8.Can Tables 24 and Table 25 be reconciled with Table 2 at (D=256)? Please specify the layer-selection protocol, seeds, checkpoints, and data configuration used for each table.
8 comments
AuthorsOpenReview ↗

Response to Reviewer Dhit (Part1)

We thank the reviewer for the detailed and precise review. The main changes relevant to this review are:

  • The scope of Theorem 1 is stated explicitly.
  • The former Theorem 2 is now a supporting lemma (Lemma 2).
  • The long-form study now uses LongFact++, with retrieval-verified labels and a stated decision rule.
  • New analyses: question-level bootstrap intervals, end-to-end cost by stage, set-aggregator baselines, a leave-one-dataset-out evaluation and a protocol-consistent dimension sweep.

We address each point below.


W1.1

"Increasing the average signed margin does not establish improved ranking or AUROC. A uniform downward logit shift can increase the average margin in a factual-majority dataset while leaving AUROC unchanged."

Thank you for this precise example. As it shows, an average margin is not a ranking measure, and any improvement in AUROC depends on the test distribution, so it cannot follow from a statement about margins alone. We therefore do not claim that Theorem 1 explains the AUROC of HaMI (SP), and the revision removes wording that suggested otherwise.

Theorem 1 is meant as an account of the mechanism by which semantic-probability scaling enters the classifier. The scaling lowers the logits, and the sensitivity ratio γ\gamma measures how unevenly it does so for factual and hallucinated bags. The theorem states when this class-dependent shift enlarges the classification margin on the training objective (Section 4.2). This is how semantic weighting acts on the margin, not a claim about detection accuracy.

W1.2

"The fixed-network intervention analyzed by the theorem differs from comparing separately trained HaMI and HaMI (SP) models in Table 5."

Thank you for pointing out this gap. Theorem 1 analyzes the scaling on a fixed network, and the separately trained models in Table 5 are related to it only approximately. The revision states this limitation with the proof (Appendix B.2). Table 5 is presented descriptively, as a comparison within each architecture family, and not as a test of the theorem.

W1.3

"Theorem 2 ... applies broadly to differentiable classifiers and is not specific to max pooling."

Thank you for this point; the result is indeed not specific to max pooling. Its role is only to show that pooling enters the first-order margin gain through the squared gradient norm ∥∇θzB∥2\|\nabla_\theta z_B\|^2, the quantity that Theorem 3 compares between max and mean pooling. In the revision we state it as a supporting lemma (Lemma 2, Section 5.2) rather than a theorem. Theorem 3, which compares this quantity, is also restated without a sparsity assumption, through quantities measured on the trained detector (Definition 1).


W2

"The reliability of the LongFact labels remains a major concern ... the precise response-level decision rule is insufficiently specified."

Thank you for raising this concern, which we share. Our original long-form study had two weaknesses that could not be fixed within its design: its labels came from a single LLM judge working from its own knowledge, and its longest cells had only a few dozen hallucinated training answers. We therefore replaced the study with LongFact++ (Obeso et al., 2026), whose labels come from external retrieval rather than from our judge, and which is large enough for class-balanced training.

Labels and decision rule.

  • Each answer comes with entity spans that the dataset's pipeline checked against web search and labeled Supported, Not Supported or Insufficient Information.
  • We label an answer hallucinated when at least 10% of its checked entities are Not Supported. Unverifiable spans count in the denominator but not as errors (Section 6.3, Appendix D.2).
  • The dataset authors validated the pipeline (Obeso et al., 2026). Human annotators who checked a random sample of spans with web search agreed with it in 84% of cases (n = 50). On passages with injected factual errors, it detected 80.6% of the errors, with a 15.8% false-positive rate on unchanged entities.

Setting and results (Appendix D.2, Table 13; Section 6.3, Table 3). For LLaMA-3.1-8B-Instruct and LLaMA-3.3-70B-Instruct (median about 740 tokens), we use 1,000 class-balanced training answers, 1,000 validation answers and 1,500 test answers per model.

Method8B AUC8B AUPRC70B AUC70B AUPRC
SEP0.4930.6650.5340.469
LLMs Know0.6230.7730.6330.558
SAPLMA0.6470.7740.6370.581
HaMI0.7110.8270.6650.602
HaMI (SP)0.7110.8280.6630.598
Mean0.7590.8640.7390.689
Max0.7830.8810.7390.691

Mean of five runs; chance AUPRC is 0.674 (8B) and 0.449 (70B). The four MIL detectors outperform the three fixed-positionprobes on both models, and answer length alone reaches only 0.59 and 0.52 AUC.

AuthorsOpenReview ↗

Response to Reviewer Dhit (Part2)

Q1

"How should Theorem 2 be interpreted for actual training, where a shared parameter update is applied over minibatches? Can the authors provide evidence that the per-bag margin analysis predicts shared-model training dynamics?"

Thank you for this question. The per-bag result (now Lemma 2) and shared training are linked by an exact first-order identity, and we measured both of its terms during training.

Notation. As in the paper, a bag BB has label yB∈{+1,−1}y_B\in\{+1,-1\}, logit zB(θ)z_B(\theta), margin mB=yBzBm_B=y_Bz_B and loss LB=log⁡(1+e−mB)\mathcal L_B=\log(1+e^{-m_B}). A shared step on a mini-batch B\mathcal B of nn bags is θ+=θ−ηn∑B′∈B∇LB′(θ)\theta^+=\theta-\frac{\eta}{n}\sum_{B'\in\mathcal B}\nabla\mathcal L_{B'}(\theta).

Identity. Since ∇LB′=−yB′∇zB′/(1+emB′)\nabla\mathcal L_{B'}=-y_{B'}\nabla z_{B'}/(1+e^{m_{B'}}), for every B∈BB\in\mathcal B mB(θ+)−mB(θ)=ηn (aB+cB)+O(η2),aB=∥∇zB∥21+emB,cB=∑B′∈B, B′≠ByByB′ ⟨∇zB,∇zB′⟩1+emB′.m_B(\theta^+)-m_B(\theta)=\frac{\eta}{n}\,(a_B+c_B)+O(\eta^2),\qquad a_B=\frac{\|\nabla z_B\|^2}{1+e^{m_B}},\qquad c_B=\sum_{B'\in\mathcal B,\,B'\neq B}\frac{y_By_{B'}\,\langle\nabla z_B,\nabla z_{B'}\rangle}{1+e^{m_{B'}}} .

  • aBa_B is exactly the per-bag gain of Lemma 2, and cBc_B is the interference from the other bags.
  • For n=1n=1, cB=0c_B=0 and the identity reduces to Lemma 2.
  • Whenever cB≥0c_B\ge0, the gain of bag BB under the shared update is at least ηnaB\frac{\eta}{n}a_B.

Measurements.

  • Setup. We trained max- and mean-pooling detectors with this loss (LLaMA-3.1-8B, four QA benchmarks, three seeds). At epochs 0, 1, 2, 5, 10, 20, 50 and 100 we applied one shared step (n=128n=128, η=10−3\eta=10^{-3}) to eight random mini-batches per run, 96 per checkpoint, and recomputed all margins.
  • The identity is accurate. The median relative error of the predicted change of the batch-mean margin is below 0.03%.
  • Some bags lose margin, as the reviewer anticipates. In a typical step about one third of the bags lose margin (33–38% for max pooling).
  • The lower bound holds when its condition does. cB≥0c_B\ge0 held for 59–65% of bags under max pooling at every checkpoint, and among them the lower bound held in more than 99.99% of cases.
  • At the batch level the per-bag term prevails for max pooling. The batch-mean margin increased in 90–100% of mini-batches for max pooling at every checkpoint, against 56–77% for mean pooling between epochs 5 and 50. The gain of max pooling exceeded that of mean pooling in 88–100% of mini-batches up to epoch 50.
  • Decomposition of the gain. Medians over the mini-batches of each checkpoint, max pooling against mean pooling:
EpochsPer-bag term ∑BaB\sum_B a_B, max / meanMini-batches with ∑BcB>0\sum_B c_B>0, max vs. meanBatch-mean margin gain, max / mean
0–22.5×86% vs. 82%2.5×
5–501.7×82% vs. 56%3.5×
1001.0×73% vs. 57%0.9×

For max pooling the interference mostly adds to the per-bag term (median ∑cB/∑aB\sum c_B/\sum a_B of 1.3 at epochs 5–50, against 0.3 for mean pooling). The gap closes at epoch 100, when max-pooling margins are already large and the factor 1/(1+emB)1/(1+e^{m_B}) of Lemma 2 is small.

Lemma 2 therefore identifies the term aBa_B that max pooling keeps large for sparse bags. The identity specifies when this term controls shared-update training. It is not a guarantee for every bag or every step.

Q2

"Please clarify the train/validation/test split construction ... are standard errors computed over random seeds only, or do they account for response correlation within the same question?"

Thank you for asking us to clarify these points.

Splits.

  • QA. Training questions come from each dataset's training split, and validation and test questions from its validation split, so no question appears in both training and evaluation (Section 6.1, Appendix D.1). Validation and test are split at the question level with a fixed seed, so all five responses to a question fall on the same side.
  • LongFact++. Every prompt has a single answer, and no prompt is shared across our splits (Appendix D.2).

Standard errors.

  • The standard errors in the paper (Tables 22–24) are over five training seeds with the test set fixed, so they reflect training variance only.
  • We now also compute question-level cluster-bootstrap intervals. These resample test questions together with all of their responses (B=1,000B=1{,}000), so they account for the correlation within a question.
  • On QA, the bootstrap standard error of AUC is 0.007–0.016, against at most 0.006 across seeds (Appendix E.3).
  • For LongFact++, where each prompt has one answer, the bootstrap is over answers (Table 15).
AuthorsOpenReview ↗

Response to Reviewer Dhit (Part3)

Q3

"How robust are the results to the choice of automatic judge? ... a different LLM judge, retrieval-augmented verification, or human annotations on a subset of the main benchmarks?"

Thank you for this question. For comparability with HaMI, the QA labels follow its protocol (Niu et al., 2025), which uses the judge prompt of Farquhar et al. (2024): an LLM (GPT-5 mini) judges each answer against the reference answers, and answers judged hallucinated are re-checked. We tested robustness to this choice in three ways.

1. A different LLM judge (Appendix F.3, Tables 33–34).

  • Setup. For each generator and QA benchmark we drew 300 test responses (3,600 in total). Claude Sonnet 5.5, a judge from a different model family, saw the question, the reference answers and the response, and applied the same criterion. The detectors are those of Table 2, trained on GPT-5 mini labels and not retrained; only the test labels change.
  • Agreement. The two judges agree on 94.1% of responses (mean Cohen's κ=0.86\kappa=0.86, 0.71–0.99 per setting).
  • AUC. Every method loses 0.013–0.024 AUC on average, and outside BioASQ the mean change per method is at most 0.014. On BioASQ, Sonnet marks fewer responses as hallucinated (24–42% vs. 34–49%), mostly partly correct list answers, and every method loses 0.04–0.06.

2. Retrieval-augmented verification (Appendix F.3, Table 35). The Sonnet judge re-checked, with one web search per response, every response it could not settle from the answer alone: 2,013 of the 3,600, where the answer does not contain a reference answer verbatim or the two judges disagree. This changed 2.9% of the Sonnet labels and brought them closer to GPT-5 mini (mean κ=0.89\kappa=0.89, 0.79–0.99).

Mean test AUC over the twelve settings on the same 3,600 responses ("Best": settings where the method is highest):

MethodGPT-5 miniSonnetSonnet + web searchBest (Sonnet / web)
Max0.8970.8790.8835 / 6
Mean0.8820.8690.8702 / 1
HaMI0.8670.8520.8530 / 0
HaMI (SP)0.8920.8770.8763 / 3
SAPLMA0.8570.8400.8431 / 1
LLMs Know0.8580.8410.8421 / 1
SEP0.8110.7870.7870 / 0

The order of the top four methods (max pooling, HaMI (SP), mean pooling, HaMI) is the same under all three label sets, and max pooling has the highest mean AUC under each.

3. Human annotations.

  • QA judge (Appendix D.3, Table 16). On 1,853 TruthfulQA responses with human labels, GPT-5 mini agrees with the annotators on 95.3% and misses almost no human-identified error (FNR 0.46%); most disagreements are answers marked incorrect by the judge but correct by humans (8.5%).
  • Long-form labels. The LongFact++ labels are retrieval-verified by construction, and their pipeline was validated against human annotation by the dataset authors: 84% agreement with human annotators who checked a random sample of spans with web search (W2).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Revision and Responses to Shared Concerns

Dear Area Chair and Reviewers,

Thank you again for your comments. We have uploaded a revised paper and replied to each review. The main changes are listed below; section and table numbers refer to the revised version.


1. Long-form labels (all reviewers)

We replaced the original long-form experiment with LongFact++ (Obeso et al., 2026). This fixes the two problems of the old setup:

  • Enough data. Each model now has 1,000 class-balanced training answers (500 hallucinated, where the old setup had only 69 and 35), plus 1,000 validation and 1,500 test answers.
  • Labels independent of our judge. The dataset's pipeline checks every entity in an answer against web search, and its authors report 84% agreement with human annotators.

Results (Section 6.3; Tables 3 and 15; Appendix D.2):

  • All four MIL detectors beat the fixed-position probes.
  • Among the methods in Table 3, max pooling is best on LLaMA-3.1-8B-Instruct and ties mean pooling on LLaMA-3.3-70B-Instruct.

2. Theory (all reviewers)

  • Theorem 3 no longer assumes inactive background features. It is stated in terms of quantities that can be measured on a trained detector (Definition 1) and contains the original bound as a special case.
  • The bound holds for every test bag of the trained detectors (Section 5.2; Appendix G, Table 36).
  • The former Theorem 2 is now a lemma, and we say explicitly that the analysis is about the training signal, not about AUC.

3. Choice of judge (Dhit, P81f, WFo5)

  • We relabeled QA test responses with a second LLM judge, with and without web search.
  • Max pooling has the highest mean QA AUC in both cases (Appendix F.3, Tables 33–35).

4. Statistics (Dhit)

  • We added question-level paired bootstrap intervals, AUPRC and threshold-based metrics (Tables 15, 25 and 26).

5. Cost (Dhit, WFo5)

  • Timing now runs from the prompt to the score and includes generation.
  • Max pooling adds a median of 0.5% to the response time; HaMI (SP) takes 7–34 times as long on QA (Section 6.4; Appendix E.1, Tables 18–20).

6. More baselines (Dhit, P81f)

  • We added seven set aggregators (ABMIL, gated ABMIL, log-sum-exp pooling, DeepSets, Set Transformer, learned top-k pooling and a Transformer encoder) and four recent detectors (HaluNet, SFP, SinkProbe and HIDE), with both AUC and cost (Appendix E.4, Tables 27–30).

7. Generalization (Dhit, WFo5)

  • We added a leave-one-dataset-out evaluation, in which training and model selection see only the source datasets (Appendix E.5, Table 31).

Further results in the individual responses

We also ran the following experiments during the discussion and will add them to the appendix:

  • how much inactive tokens dilute the mean-pooling gradient;
  • measurements of shared mini-batch updates;
  • results on two human-annotated long-form datasets;
  • transfer from QA to LongFact++.

If anything is still unclear or unresolved, please let us know and we will respond during the rest of the discussion period.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.