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