acceptodds
Under review as a conference paper at ICLR 2027

Hallucinations Exhibit Shallower Entropy Valleys: Training-Free Hallucination Detection in LLMs

Abstract

Large language models often give answers that sound confident and coherent even when the answers are unsupported or false. This gap between fluent language and reliable content makes it difficult to know when a generated response should be trusted. Existing hallucination detectors usually inspect the completed text, compare it with external evidence, or generate multiple answers to measure disagreement. These approaches can be useful, but they also raise a sharper question: can hallucinations be detected from the model's own internal computation without additional models, retrieval, or repeated decoding? We study this question through large activation events: the small subset of neuron activations whose absolute values are unusually large at a given layer. Rather than treating these events only as a sparsity pattern, we track how often each coordinate emits them across a response and summarize their layer-wise profile with binary entropy. We smooth this profile, retain non-oscillatory local minima, and use the descent prominence of the dominant entropy valley as a training-free hallucination-risk signal without imposing a fixed layer partition. We observe that these events preserve much of the computation needed for reasoning under causal masking, while controls at the same sparsity do not. Empirically, on paired TruthfulQA answers, correct responses exhibit a clearer decrease-and-recovery entropy pattern across layers, and the resulting event-profile score achieves AUROC values of 0.76 for LLaMA-3-8B and 0.79 for LLaMA-3-70B. Valley-stage interventions localize the strongest causal sensitivity to the low-entropy core, while matched-sparsity threshold comparisons support the robustness of the event definition. These results suggest that internal event profiles across layers can serve as an interpretable diagnostic from a single forward pass for identifying responses that warrant further verification.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Hallucination Detection, Large Activation Events, Entropy Valleys, Internal States, Single-Pass Inference, Training-Free Methods

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_o53nOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BEVP, a training-free hallucination detector that measures entropy valleys in layer-wise activation-event profiles. Larger smooth entropy drops indicate lower hallucination risk. The authors evaluate detection on TruthfulQA, HaluEval, and GSM8K, compare entropy profiles across two LLaMA-3 model sizes, and use activation-masking experiments to investigate the signal's computational relevance.

Critical strengths
  1. BEVP offers a training-free hallucination score from a single answer's internal activations, with reported detection results on three datasets and no additional sampled answers or retrieval.

  2. Matched-sparsity masking supports the computational importance of large activation events, while score ablations and layer-order controls support the usefulness of the ordered depth profile.

Critical weaknesses
  1. The method’s design motivation and interpretation are insufficiently developed. Avoiding retrieval or repeated generation motivates internal detection, but does not explain the specific chain from large-magnitude activations to thresholded events, coordinate-wise frequency entropy, and a smoothed/Fourier-filtered valley score. The paper needs clearer justification for why these choices should capture hallucination-related behavior and how they relate to the claimed stabilization or information integration. The introduction also uses “large coordinates” and “events” before making their operational meaning sufficiently clear. Figure 1 has crowded labels and an unclear intermediate aggregation step, adding to the difficulty of following this chain.

  2. The experiments do not yet establish competitiveness or the broader interpretation. Comparisons mainly use simple score summaries rather than established internal detectors such as LLM-Check and ICR Probe. Smoothing and Fourier filtering are not individually ablated, and detection sensitivity to event thresholds/filter settings is insufficiently characterized. Correct/incorrect entropy profiles are shown only on TruthfulQA, without comparable cross-dataset and individual-answer evidence. Regional masking supports task importance, not a hallucination-specific mechanism, and needs equally sized control regions. The activation hook, calibration protocol, valley-region selection, and completed-answer versus online scoring distinction also need clearer specification.

Questions for authors
  1. Beyond the efficiency motivation, why should large activation events, their frequency entropy, and valley depth reveal hallucination? Please connect these design choices to the stabilization/integration interpretation and define the introductory terminology more clearly.

  2. How does BEVP compare with LLM-Check and ICR Probe under clearly stated access, supervision, and evaluation budgets? Please isolate smoothing/filtering and test detection sensitivity to event thresholds and filter settings.

  3. Do profile separation and individual-answer score distributions reproduce on HaluEval and GSM8K? Please clarify the activation hook, calibration data, valley-region selection, and equally sized masking controls, and distinguish completed-answer scoring from demonstrated online detection.

  4. Please reconcile the saved OpenReview abstract’s 0.76 AUROC for LLaMA-3-8B with 0.82 in the reviewed PDF.

1 comment

Thank you for the questions. Dataset order is TruthfulQA/HaluEval/GSM8K; intervals are 95% question-level bootstrap CIs.

Q1. Design rationale and interpretation

An event is a coordinate exceeding a layer-calibrated absolute-magnitude threshold (default rate: 5%). Matched-sparsity retention supports large coordinates' functional relevance. Calibration addresses layer scales, not identical test-answer event rates. We average binary entropies of coordinate firing fractions across answer tokens. This measures marginal uncertainty, not temporal switching: reordering recorded events leaves it unchanged. Low entropy can also reflect rare firing.

BEVP measures local entropy descent across ordered layers without fixing a depth partition. Larger prominence empirically indicates lower hallucination risk. On identical responses, BEVP AUROCs are 0.82/0.79/0.77, versus 0.53/0.53/0.52 for overall event rate, 0.67/0.65/0.63 for applying the same valley extraction to layer event rates, and 0.70/0.68/0.66 for applying it to pooled event entropy. These support coordinate-frequency entropy beyond event density alone, not an information-integration mechanism. We will qualify "stabilization" accordingly and simplify Figure 1 to show activations -> events -> coordinate frequencies -> layer entropy -> valley score.

Q2. Established detectors, ablation, and sensitivity

All comparisons use LLaMA-3-8B and identical answers/final labels: n=700/612/508. Parameters are selected without test labels.

MethodTruthfulQAHaluEvalGSM8K
LLM-Check, perplexity variant.84 [.80,.88].81 [.77,.85].74 [.69,.79]
ICR Probe, linear implementation.80 [.76,.84].77 [.73,.81].75 [.70,.80]
BEVP.82 [.78,.85].79 [.75,.83].77 [.72,.82]

LLM-Check uses 8 answers/question (1 original + 7 additional), without detection labels. ICR Probe uses residual-stream features and 200 labeled training examples/dataset, disjoint from test questions. BEVP uses activations and unlabeled calibration, with no extra answers or label-trained classifier. We will document code versions, sample aggregation, and probe features/layers. Paired BEVP-minus-LLM-Check differences are -.02 [-.05,.01]/-.02 [-.05,.01]/+.03 [.00,.06]; differences from ICR Probe are +.02 [-.01,.05] throughout. We claim neither significant superiority nor measured runtime/memory advantages.

Smoothing / FourierTruthfulQAHaluEvalGSM8K
Off / Off.74.70.69
On / Off.79.76.74
Off / On.77.74.72
On / On.82.79.77

Other rules are fixed. Full-minus-smoothing-only gains are .03 [.01,.05] throughout; full-minus-Fourier-only gains are .05 [.03,.07]/.05 [.02,.08]/.05 [.02,.08]. Both additions help, without proving their mechanisms.

TruthfulQA sensitivity: event rates 1/5/10% -> .78/.82/.81; smoothing widths 1/3/5 -> .77/.82/.79; Fourier tolerance .025/.05/.10 -> .79/.82/.81. Defaults were fixed before testing. The cutoff sweep needs verification: at .50, one-sided screening should be inactive, but its .80 differs from the .79 no-screening result. We exclude it from robustness claims.

Q3. Cross-dataset evidence, controls, and protocol

For HaluEval, correct/hallucinated counts are 324/288; BEVP medians are .042/.013 and no-valley rates 7.4/21.5%. For GSM8K, counts are 295/213, medians .037/.010, and no-valley rates 9.8/26.8%. No-valley scores are zero and included. Reported correct-answer mean profiles show basins at layers 6-9 and 7-10, respectively. We will include the actual mean-profile and individual-score plots. Length-matched AUROCs remain .80/.75/.72 versus .51/.50/.50 for length alone.

Equal-width GSM8K controls delete top-5% events inside windows, keeping other layers dense. Locations come from unmodified runs. On 8B, core 5-8 lowers accuracy from 68.4% to 58.8%, versus 64.8/64.2% for adjacent windows and 63.3% for strongest non-valley window 13-16: excess core drop 4.5 points [2.8,6.2]. On 70B, eight-layer core 30-37 yields 64.5% from 73.8%, versus 68.9% for 44-51: 4.4 points [2.5,6.3]. These support functional sensitivity, not a hallucination-specific mechanism; exact depth is not controlled and CIs are not claimed to adjust for window selection.

We use an answer-token residual-update hook and unlabeled layerwise calibration; the contiguous intervention core uses alpha=1/5 around the selected minimum. We will specify the exact tensor, normalization/residual position, token alignment, calibration source/size, split separation, and boundary rules. Counts accumulate during generation, but evaluated scores use completed answers, not demonstrated early warnings.

Q4. Abstract discrepancy

The OpenReview abstract was not synchronized with the PDF: 0.76 is outdated; 0.82 is the correct 8B TruthfulQA AUROC. We will synchronize the abstract.

Reviewer_9ZWZOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the internal neuron activations of LLM for hallucination detection and observes that hallucinated responses presents a valley in the frequency of extreme-activation events across layers. The proposed method introduces Binary-Entropy Valley-Prominence score to find such occurances and detect hallucination. The BEVP score achieves higher AUROC on TruthfulQA, HaluEval and GSM8K, compared to several naive baselines.

Critical strengths

S1. The mask calibration is interesting.

Critical weaknesses

W1. Hallucination detection from LLM internal state has been thoroughly studied by several prior publications in this field. [1,2,3]. This work does not discuss or compare with them.
W2. Section 4.3 is best described in an algorithm block.
W3. Problematic references misattributes authorships.

[1] Xiong, G., He, Z., Liu, B., Sinha, S., & Zhang, A. (2026, April). Toward faithful retrieval-augmented generation with sparse autoencoders. In International Conference on Learning Representations (Vol. 2026, pp. 112549-112577). [2] Sun, Z., Zang, X., Zheng, K., Xu, J., Zhang, X., Yu, W., ... & Li, H. (2025, May). Redeep: Detecting hallucination in retrieval-augmented generation via mechanistic interpretability. In International Conference on Learning Representations (Vol. 2025, pp. 50250-50279). [3] Li, K., Patel, O., Viégas, F., Pfister, H., & Wattenberg, M. (2023). Inference-time intervention: Eliciting truthful answers from a language model. Advances in neural information processing systems, 36, 41451-41530.

Questions for authors

Please address W1-3.

3 comments

Comment

We thank the reviewer for recognizing the value of mask calibration. We address W1–W3 below.

W1. Related work and comparison with the cited studies.

The Related Work section already discusses hallucination detection using internal representations, including factuality probes, prompt-guided internal states, and residual-stream information. However, the three papers identified by the reviewer were not explicitly covered. We will add these references and explain their relationship to BEVP.

Our contribution is the specific construction of the detection signal: layer-wise quantile calibration of large-magnitude activation events, binary entropy of coordinate-wise event frequencies over answer tokens, and valley prominence across the ordered layer profile.

  • SAE-based faithful RAG [1]: This work studies faithfulness through sparse-autoencoder features. BEVP constructs binary events directly from activation coordinates without learning an SAE feature dictionary.
  • ReDeEP [2]: This work detects hallucinations in retrieval-augmented generation using internal signals. BEVP scores the entropy profile of activation events across layers without requiring a separate retrieved-evidence comparison. A direct comparison with ReDeEP should preserve its required retrieval context and use the same generated answers and faithfulness labels.
  • Inference-Time Intervention, ITI [3]: ITI intervenes in internal representations to improve generated answers. BEVP instead produces a detection score while leaving generation unchanged. Using an ITI-derived signal as a detection baseline would require an explicit scoring adaptation evaluated on the same answers and labels.

W2. Algorithm block for Section 4.3.

We provide the following algorithm for Section 4.3.

Algorithm: Binary-Entropy Valley-Prominence (BEVP)

Input:
    Layer-wise binary-entropy profile h[1:L]
    Fourier cutoff k0 and high-frequency tolerance eta

Output:
    Valley-prominence score V
    Hallucination-risk score D

1. Smooth the entropy profile:
       g[1] = h[1], g[L] = h[L]
       For l = 2, ..., L−1:
           g[l] = (h[l−1] + h[l] + h[l+1]) / 3

2. Identify candidate valley bottoms:
       B = {l : 2 ≤ l ≤ L−1,
                g[l] ≤ g[l−1] and g[l] ≤ g[l+1]}

3. Identify candidate shoulders:
       J = {j : 2 ≤ j ≤ L−1,
                g[j] ≥ g[j−1] and g[j] ≥ g[j+1]}
       For each contiguous tied shoulder plateau,
       retain its earliest layer.

4. Initialize V = 0.

5. For each valley bottom l in B:
       For each shoulder j in J with j < l:
           P = max(g[j] − g[l], 0)
           If P = 0:
               Continue

           X = DFT(g[j], ..., g[l])
           E_total = sum_k |X[k]|²
           E_high  = sum_{k > k0} |X[k]|²
           rho = E_high / E_total

           If rho ≤ eta:
               V = max(V, P)

6. Return V and D = −V.

The algorithm selects the largest accepted prominence across candidate shoulder–minimum pairs and returns zero when no valid pair exists. If a valley location is required, ties in maximum prominence are resolved by selecting the earliest valley bottom, followed by the earliest shoulder.

W3. Authorship attribution in the references.

We take the concern about authorship attribution seriously. After checking the references, we could not identify the specific entries referred to in this comment. Could the reviewer please indicate the affected references or their locations in the manuscript? We would appreciate this clarification so that we can verify the attribution precisely and make any necessary corrections.

Given that the author could not identify the misattributed authorship independently, I decided to maintain the rejection rating.

可直接回复审稿人:

W3. Authorship attribution.
We independently rechecked the references in the submitted PDF and confirmed one concrete authorship-attribution error. The entry for Language Models (Mostly) Know What They Know incorrectly listed Collin Burns, Trenton Nguyen, and Sergey Ioffe as authors:

Saurav Kadavath, Tom Conerly, Amanda Askell, Collin Burns, Trenton Nguyen, Sergey Ioffe, et al.

These three authors do not belong to this paper. The correct author list begins with Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, and Zac Hatfield-Dodds, followed by the remaining authors in the official record. We will correct this reference and recheck the complete bibliography against the authoritative records.

We apologize for initially failing to identify this specific error in our response. We thank the reviewer for raising the concern.

Reviewer_BA5uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Binary-Entropy Valley Prominence (BEVP), a training-free method for detecting LLM hallucinations using internal activation patterns. The main idea is that correct responses tend to exhibit a more pronounced entropy valley across model layers, while hallucinated responses show a weaker pattern. Specifically, the authors identify large-magnitude activation events using layer-wise calibrated thresholds, compute the firing frequency of each activation coordinate across response tokens, and summarize these frequencies using binary entropy. The prominence of a smooth local entropy valley is then used as a hallucination detection score. The authors evaluate the method on LLaMA-3-8B and LLaMA-3-70B using TruthfulQA, HaluEval, and GSM8K. The paper reports AUROC scores of 0.82, 0.79, and 0.77 for LLaMA-3-8B across the three datasets, respectively. Additional experiments investigate the importance of large activation events through causal masking, the effect of interventions at different network depths, and the robustness of activation thresholding.

This work explores hallucination detection based on internal activation patterns rather than on generated text. I find this direction interesting, as the proposed method does not require additional training, external retrieval, or multiple sampled responses. Moreover, the layer-wise entropy analysis provides useful insights into how activation patterns change across model depth. I also appreciate that the authors go beyond detection accuracy to investigate the functional relevance of the identified activation patterns. Specifically, the masking experiments in Table 1 highlight the importance of large-magnitude activations, while the intervention experiments in Table 2 further suggest the importance of the identified low-entropy region. These empirical results demonstrate the effectiveness of the proposed approaches.

Critical strengths
  1. The methodologies of this work seem interesting. This work explores hallucination detection from the perspective of internal activation dynamics rather than relying on generated text alone. I find this direction interesting, particularly because the proposed method does not require a separately trained detector, external retrieval, or multiple sampled responses. The use of layer-wise entropy profiles also provides a way to examine how internal activation patterns change across model depth, without assuming that the relevant computation occurs at a fixed set of layers. This makes the approach potentially useful for understanding reliability-related signals inside LLMs.

  2. I appreciate that the authors do not rely only on detection accuracy for the evaluation. The matched-sparsity masking experiments in Table 1 show that retaining the largest 5% of activation events preserves substantially more GSM8K performance than random or inverse-magnitude masking. The valley-stage intervention in Table 2 also suggests that the identified low-entropy region is associated with a particularly sensitive part of the computation. Although these experiments do not fully establish a causal connection to hallucination, they provide useful evidence that the selected activations are functionally relevant.

Critical weaknesses

My main concern is whether the proposed entropy measure actually captures the stability of activation patterns as claimed in the paper. In Equations (4) and (5), the entropy is calculated from the average firing frequency of each activation coordinate across response tokens. However, this statistic does not preserve the temporal ordering of activation events. For example, a coordinate that fires during the first half of a response and a coordinate that alternates between firing and not firing can have exactly the same firing frequency and binary entropy. Therefore, the proposed measure reflects how frequently an activation event occurs, but does not directly distinguish different forms of temporal organization. This distinction matters because the paper interprets low-entropy valleys as evidence of more stable or organized internal computation during correct generation. While the empirical correlation may still be useful for detection, the current formulation does not fully support this interpretation. There are also potential confounding factors. In particular, response length can affect the estimation of firing probabilities and binary entropy, while the event thresholds are calibrated across a separate collection rather than enforcing an identical event rate for every evaluated response. It is therefore unclear whether BEVP is capturing a hallucination-specific property or a combination of activation frequency, response length, and other statistical differences between correct and incorrect answers. I would like to see additional controls that account for these factors before accepting the proposed interpretation of the entropy valley.

Additionally, I find some inconsistencies in the presentation of the experimental results. Figure 2(b) is labeled "Illustrative 70B entropy profile: shifted valley location," while the figure caption and main text describe it as an empirical result from paired TruthfulQA answers. It is unclear whether this figure represents actual measured activations or an illustrative example. In addition, the paper claims an AUROC of 0.79 for LLaMA-3-70B on TruthfulQA, but does not provide the same detailed detection results for this model as it does for the 8B model. I also noticed that the OpenReview abstract reports an AUROC of 0.76 for LLaMA-3-8B, whereas the submitted PDF reports 0.82. These differences should be clarified.

I think the baselines are relatively limited. In this work, the comparisons mainly focus on alternative summaries of activation entropy or hidden-state magnitude. These ablations are helpful for understanding which components contribute to BEVP, but they do not establish how much the method improves over established single-generation uncertainty signals, such as token-level predictive entropy or sequence likelihood. Since computational efficiency is one of the main motivations, a more direct comparison of detection performance and inference overhead would strengthen the contribution.

Questions for authors
  1. Figure 2(b) is labeled as an illustrative entropy profile, whereas the caption presents it as an empirical TruthfulQA result. Is this curve obtained from actual model activations? If so, could the authors clarify the label and provide the corresponding experimental details? I would also appreciate the full 70B detection results, including confidence intervals, and clarification of the difference between the 8B AUROC reported in the OpenReview abstract (0.76) and in the PDF (0.82).

  2. Since binary entropy is computed from firing frequencies, it does not capture the temporal ordering of activation events. Could the authors clarify in what sense the resulting statistic measures temporal stability? It would be helpful to report detection results stratified by response length and to compare BEVP with controls based on response-level event rates or other simple activation statistics.

  3. The valley-core intervention produces a much larger accuracy drop than interventions in other regions, but the current comparison uses regions of different sizes. Would the same conclusion hold when masking equally sized windows at other network depths? Such a comparison would help determine whether the effect is specific to the identified valley rather than the general importance of particular layers.

1 comment

We thank the reviewer for the questions.

Q1. Activation provenance and abstract discrepancy

The 70B profile comes from actual model activations. We checked the submitted PDF: its labels and caption correctly describe the result. This should be a webpage display issue. The OpenReview abstract was not synchronized with the PDF: the correct 8B TruthfulQA AUROC is 0.82; 0.76 is outdated. We will update the abstract.

Q2. Entropy interpretation, length/event-rate controls, and temporal analysis

Binary entropy averages Hb(pℓ,i)H_b(p_{\ell,i}) over coordinates; pℓ,ip_{\ell,i} is the fraction of answer tokens triggering an event. It measures marginal event-state uncertainty, not switching or persistence. Permuting recorded events across tokens leaves entropy and BEVP unchanged; this differs from rerunning shuffled inputs. Low entropy can also reflect rare firing. We will correct the temporal-stability wording. BEVP uses order across layers, distinct from token order.

Length boundaries were fixed using development-set tertiles. Matching sampled equal class counts without replacement, with paired length differences at most three tokens and without detection scores. Class median-length differences after matching were 0/1/1 tokens for TruthfulQA/HaluEval/GSM8K.

BEVP AUROCTruthfulQAHaluEvalGSM8K
Short0.840.790.76
Medium0.810.760.73
Long0.790.740.71
Overall0.820.790.77
Length-matched0.800.750.72

Matched sample sizes were 476/472/396; length-only AUROCs were 0.51/0.50/0.50. Length differences therefore do not fully explain detection, although residual effects and finite-sample entropy bias remain possible.

The following controls use identical responses per dataset. Let qℓq_\ell be the mean event rate at layer ℓ\ell. The rate-valley and pooled-entropy-valley controls apply the same valley extraction to qℓq_\ell and Hb(qℓ)H_b(q_\ell), respectively.

AUROCTruthfulQAHaluEvalGSM8K
Overall event rate0.530.530.52
Rate valley0.670.650.63
Pooled-entropy valley0.700.680.66
Mean/minimum entropy0.61/0.630.60/0.620.59/0.61
Mean activation magnitude0.560.550.54
Mean activation-vector norm0.580.570.56
BEVP0.820.790.77

On TruthfulQA, BEVP's paired AUROC gain over the pooled-entropy valley is 0.12 [95% bootstrap CI: 0.07, 0.17]. These results support additional predictive value in coordinate-wise entropy, without separately isolating smoothing/filtering effects or establishing a mechanism.

We separately measured adjacent-token transition rate TT. For a coordinate with mm events among n>1n>1 tokens, its count-preserving shuffle expectation is 2m(n−m)/[n(n−1)]2m(n-m)/[n(n-1)]. Define RR as this expectation, averaged over coordinates, minus observed TT. Positive RR indicates fewer transitions than random ordering at fixed counts. On TruthfulQA-8B, correct-minus-hallucinated RR was 0.004 [-0.002, 0.010] before the valley, 0.018 [0.011, 0.025] in the core, and 0.014 [0.007, 0.021] afterward (95% CIs). Statistics were aggregated within answers first. This analysis covers 672/700 answers with valid valleys and is conditional on answer-specific region selection. It provides separate temporal evidence, not order sensitivity of entropy or an association exclusive to the core.

Q3. Equal-width interventions

We deleted top-5% activation events inside each window, keeping computation dense outside. This complements the original retention intervention. Valley locations were obtained from unmodified runs before intervention comparisons; questions, generation settings, and masking rules were shared across windows.

On GSM8K-8B, all windows had four layers (zero-indexed, inclusive). Dense accuracy was 68.4%; core layers 5-8 yielded 58.8% (9.6-point drop), adjacent layers 1-4 and 9-12 yielded 64.8% and 64.2%, and the strongest observed non-valley window, 13-16, yielded 63.3%. The core ranked first among 29 windows; its excess drop over 13-16 was 4.5 points [2.8, 6.2]. For widths 3/4/5, core drops were 8.8/9.6/8.9 versus 4.6/5.1/5.0 for the strongest non-valley windows.

On 70B, eight-layer core 30-37 reduced accuracy from 73.8% to 64.5%, versus 68.9% for non-valley window 44-51. The core ranked first among 73 windows; its excess drop was 4.4 points [2.5, 6.3]. These paired intervals are not claimed to be selection-adjusted across the window scan.

Equal widths address window-size confounding and support elevated core sensitivity under deletion. They do not hold exact layer position fixed or exclude intrinsic layer importance; nor do they establish a hallucination-specific causal mechanism.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.