- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes DLIPS, a white-box method that detects whether a given answer is a hallucination from how sensitive the model is while reading the answer. Specifically, at each prefix of the answer, DLIPS measures how much the selected layers amplify a small change to the last token by computing the spectral norm of the layer's Jacobian. It then summarizes how these values change across prefixes to build the features for a simple classifier.
The paper claims three main contributions.
- It reveals that the sensitivity changes along the answer carry information for hallucination detection beyond per-layer and global statistics of the sensitivity.
- A detection framework that computes this sensitivity without forming the Jacobian, and classifies an answer with a frozen model.
- An evaluation across general QA, RAG, and medical QA tasks.
For validation, DLIPS with logistic regression reports better F1 and AUROC than the compared baselines on HaluEval, RAGTruth, and MedHallu with Qwen3.5-0.8B.
Critical strengths
The paper computes the Jacobian norm of each complete decoder block without forming the Jacobian, and checks the estimates against exact SVD (Appendix D). This supports the accuracy of the measured sensitivity values.
Critical weaknesses
- The performance advantage of the proposed method may still come from indirectly using the length of the LLM’s answer. Prior work shows that a simple detector using the length of response as feature alone can match complex hallucination detectors [1]. Although the paper provides an ablation study showing that the performance drop of removing the Global group that includes the length-related features is small, the Temporal features still contain length-related information:
- For answers of 1–3 tokens, some features are always zero, e.g., all ten Temporal features are zero for a one-token answer.
- For answers of 4–20 tokens, some features, such as the dynamic range, tend to increase along with the length.
- For answers longer than 20 tokens, the 20 prefixes are spread across the answer, so a longer answer has larger gaps between prefixes. According to Appendix B, the sensitivity differences between neighboring prefixes are not divided by the gap, so features such as the interval-averaged total variation may also change with the length.
- The main results lack uncertainty estimates. Tables 1 and 2 report point estimates without variability across seeds or splits or confidence intervals, which makes it unclear whether the improvements are stable.
[1] Janiak, Denis, Jakub Binkowski, Albert Sawczyn, Bogdan Gabrys, Ravid Shwartz-Ziv, and Tomasz Jan Kajdanowicz. "The illusion of progress: Re-evaluating hallucination detection in llms." In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.
Questions for authors
- How well does a classifier detect hallucinations if it uses only the prompt and answer lengths as input? Please compare it with DLIPS on the same test set. The released code in evaluation.py already includes a similar comparison. Please report its results.
- For each dataset, which subsets did you use, and how many examples were in the training, validation, and test sets?
9 comments
Response to Reviewer aJsY (2/3)
W2: The main results lack uncertainty estimates. Tables 1 and 2 report point estimates without variability across seeds or splits or confidence intervals, which makes it unclear whether the improvements are stable.
A2: We thank the reviewer for this comment. To assess statistical reliability, we repeated the experiments in Tables 1 and 2 using five random seeds (244–248). For each seed, we randomly sampled class-balanced responses per dataset and applied identical seeded train/test splits across methods. All results are reported as mean ± sample standard deviation (SD). Due to space constraints, Table R16 presents only DLIPS variants; complete baseline comparisons will be provided in the updated manuscript Table 1.
The repeated experiments confirm the consistency of our main findings. On Qwen 3.5-0.8B, DLIPS-Logistic achieves mean F 1 scores of 0.861/0.725/0.748 and AUROCs of 0.929/0.789/0.808 on HaluEval/RAGTruth/MedHallu, respectively, maintaining the highest mean F1 among the compared methods. The 8B/9B results also exhibit relatively small variation across seeds.
The ablation findings in Table R17 remain consistent across repeated runs. Removing temporal features reduces mean F 1 by 0.127/0.032/0.042 across HaluEval/RAGTruth/MedHallu, confirming their contribution to detection performance. Separating global statistics from token-length features further clarifies their respective contributions. The middle-layer configuration consistently achieves the strongest or comparable performance, while the relatively small SDs support the stability of these observations.
We will update Table 1 with the complete five-seed results for all methods. For Table 2, we will retain the compact mean-performance comparisons and summarize the repeated-run protocol and variability in the accompanying text to accommodate space constraints.
Table R16: DLIPS detection performance across five random seeds (Acc / F 1 / AUROC, mean ± SD, ↑).
| Backbone | Method | HaluEval | RAGTruth | MedHallu |
|---|---|---|---|---|
| Qwen 3.5-0.8B | DLIPS-Threshold | .745±.012 / .761±.010 / .823±.011 | .663±.010 / .660±.011 / .579±.012 | .546±.021 / .689±.020 / .541±.022 |
| DLIPS-RF | .857±.011 / .840±.012 / .936±.010 | .712±.014 / .712±.011 / .769±.012 | .677±.017 / .679±.016 / .734±.015 | |
| DLIPS-SVM | .815±.013 / .829±.014 / .935±.012 | .719±.008 / .722±.007 / .785±.006 | .678±.025 / .702±.022 / .736±.024 | |
| DLIPS-Logistic | .860±.006 / .861±.009 / .929±.007 | .731±.008 / .725±.007 / .789±.009 | .749±.011 / .748±.012 / .808±.013 | |
| Qwen 3.5-9B | DLIPS-Threshold | .766±.012 / .775±.011 / .827±.013 | .511±.010 / .657±.011 / .428±.009 | .640±.018 / .737±.019 / .713±.020 |
| DLIPS-RF | .891±.011 / .891±.014 / .937±.013 | .788±.010 / .766±.008 / .817±.009 | .696±.015 / .683±.014 / .756±.013 | |
| DLIPS-SVM | .890±.010 / .890±.011 / .939±.013 | .744±.011 / .745±.010 / .827±.009 | .717±.011 / .707±.010 / .739±.009 | |
| DLIPS-Logistic | .911±.011 / .917±.014 / .960±.009 | .782±.012 / .783±.011 / .864±.013 | .684±.013 / .678±.011 / .735±.014 | |
| Llama-3.1-8B | DLIPS-Threshold | .746±.010 / .772±.012 / .770±.009 | .489±.006 / .632±.008 / .490±.007 | .525±.010 / .652±.012 / .498±.011 |
| DLIPS-RF | .908±.007 / .904±.008 / .971±.006 | .811±.009 / .796±.007 / .856±.008 | .646±.023 / .632±.021 / .707±.022 | |
| DLIPS-SVM | .901±.006 / .894±.009 / .973±.007 | .801±.009 / .765±.008 / .841±.011 | .699±.013 / .717±.011 / .728±.010 | |
| DLIPS-Logistic | .920±.010 / .921±.008 / .981±.011 | .772±.006 / .754±.008 / .839±.007 | .706±.020 / .715±.021 / .733±.016 |
Table R17: Feature and layer ablations across five random seeds (F 1 / AUROC, mean ± SD, ↑).
| Type | Configuration | HaluEval | RAGTruth | MedHallu |
|---|---|---|---|---|
| Feature | Full Features | .861±.009 / .929±.007 | .725±.007 / .789±.009 | .748±.012 / .808±.013 |
| w/o Temporal | .734±.007 / .866±.008 | .693±.012 / .749±.011 | .706±.012 / .769±.011 | |
| w/o Spatial | .860±.013 / .920±.010 | .720±.014 / .801±.012 | .716±.011 / .762±.017 | |
| w/o Global Statistics | .828±.012 / .901±.013 | .711±.013 / .778±.008 | .744±.015 / .795±.014 | |
| w/o Token Lengths | .793±.006 / .887±.009 | .703±.009 / .768±.010 | .715±.017 / .778±.012 | |
| Layer | Full Layers | .860±.011 / .928±.009 | .724±.009 / .773±.010 | .731±.018 / .807±.020 |
| Bottom Layers | .508±.009 / .512±.010 | .492±.011 / .529±.012 | .534±.021 / .585±.024 | |
| Middle Layers | .861±.009 / .929±.007 | .725±.007 / .789±.009 | .748±.012 / .808±.013 | |
| Top Layers | .805±.006 / .887±.007 | .678±.010 / .731±.011 | .686±.011 / .771±.012 |
Response to Reviewer aJsY (3/3)
Q1: How well does a classifier detect hallucinations if it uses only the prompt and answer lengths as input? Please compare it with DLIPS on the same test set. The released code in evaluation. py already includes a similar comparison. Please report its results.
A3: We thank the reviewer for this suggestion. To isolate the contribution of response-length information, we evaluate w/o Global Statistics and w/o Token Lengths in Table R17. Removing explicit length features reduces AUROC by 0.042/0.021/0.030 on HaluEval/RAGTruth/MedHallu, respectively, while removing global statistics yields reductions of 0.028/0.011/0.013. These results indicate that length information contributes to detection but does not fully account for DLIPS's performance.
To further examine potential length-related shortcuts, we conducted length-only and length-matched evaluations, comparing a length-based Logistic classifier with DLIPS. Table R18 compares full DLIPS, length-only classification, and length-matched DLIPS.
Table R18: Length-only and length-matched controls (F 1 / AUROC, ↑).
| Dataset | Full DLIPS | Length-only | Length-matched DLIPS |
|---|---|---|---|
| HaluEval | 0.861 / 0.929 | 0.601 / 0.672 | 0.790 / 0.884 |
| RAGTruth | 0.725 / 0.789 | 0.566 / 0.631 | 0.706 / 0.770 |
| MedHallu | 0.748 / 0.808 | 0.585 / 0.648 | 0.713 / 0.776 |
The length-only classifier achieves AUROCs of 0.672/0.631/0.648, substantially below full DLIPS (0.929/0.789/0.808), corresponding to improvements of 0.257/0.158/0.160. Furthermore, on 100/120/80 length-matched pairs, the length-only AUROC drops to near chance (0.512/0.507/0.504), whereas DLIPS retains AUROCs of 0.884/0.770/0.776, correctly distinguishing hallucinated from faithful answers in 85.0%/80.8%/81.2% of pairs.
These results demonstrate that Jacobian sensitivity dynamics provide substantial discriminative information beyond response-length differences. We will incorporate the revised ablations and length-controlled comparisons into the manuscript.
Q2: For each dataset, which subsets did you use, and how many examples were in the training, validation, and test sets?
A4: We evaluated DLIPS on the QA subset of HaluEval, RAGTruth, and MedHallu, covering general question answering, retrieval-augmented generation, and medical hallucination detection, respectively. For each of the five random seeds (244–248), we independently sampled 800 class-balanced responses (400 hallucinated and 400 faithful) from each dataset. In each run, the newly sampled responses were divided into training, validation, and test sets at a ratio of 60%/20%/20% (480/160/160 examples). Splits were grouped by question or source to ensure that related responses and faithful–hallucinated pairs remained within the same partition, preventing data leakage. The same sampled responses and seeded splits were used across all compared methods within each run, and results were reported as mean ± standard deviation over the five runs. For HaluEval (QA), answers shorter than five tokens were excluded during the initial preprocessing. We will clarify the dataset selection, repeated-sampling protocol, and splitting details in the revised manuscript.
We thank you again for the thoughtful suggestions. We believe that these additional experiments and clarifications have strengthened our work and addressed the concerns raised. We would greatly appreciate any further feedback and welcome suggestions on any remaining issues.
Response to Reviewer aJsY (1/3)
We sincerely thank you for your thorough review and constructive feedback. Our detailed responses are provided below.
A1: We agree that temporal features can, in principle, encode length through the three mechanisms described, but our results show that DLIPS's performance does not depend on them. (i) Answers of 1–3 tokens are excluded from our evaluation for all methods. (ii) On a cohort where length, checkpoint count, and checkpoint spacing are identical within each faithful/hallucinated pair, DLIPS still achieves AUROCs of 0.884/0.770/0.776 on HaluEval/RAGTruth/MedHallu. (iii) Detection errors show no monotonic trend with answer length. We detail each point below.
Answers of 1–3 tokens. DLIPS models how sensitivity evolves during generation, and temporal statistics such as first- and second-difference variability need a minimum trajectory length to be well defined. As Table R14 shows, such answers occur almost only in HaluEval (QA), mostly as single-word or yes/no answers. To align HaluEval with the length regime of the other two benchmarks, we filter out its answers shorter than five tokens before sampling (Appendix E.1). This filter was fixed before any experiments; all baselines use exactly the same filtered samples and seeded splits. We target multi-token generation, where hallucinations develop during decoding, rather than single-token answers.
Answers of 4–20 and >20 tokens. Here the effective checkpoint count (T ≤ N) and the checkpoint spacing (T > N) vary with T. To isolate these effects, we use the length-matched cohort (Appendix H.3): each hallucinated answer is paired with a faithful answer from the same QA record and prompt, with identical token and word counts and a difference of at most three characters. Because T is identical within each pair, the effective checkpoint count, checkpoint positions (also verified empirically), and inter-checkpoint gaps are identical by construction, and length is identically distributed across the two classes. These factors therefore cannot separate the classes, and any remaining separability must come from sensitivity values. Two analyses support this:
Length-controlled results. As shown in Table R18 (our response to Question 1), length-only detection reaches AUROCs of 0.672/0.631/0.648 on the full set, well below DLIPS (0.929/0.789/0.808). On the matched cohort (100/120/80 pairs), DLIPS retains 0.884/0.770/0.776, a drop of only 0.045/0.019/0.032, showing that it relies on sensitivity dynamics rather than length.
Error rates across length intervals. Table R15 reports error rates by length (Qwen3.5-0.8B, 24-feature logistic regression, up to five checkpoints, prompt-grouped folds, threshold 0.5, five seeds; Appendix H.7). Errors are not monotonic in length; they peak at 21–50 tokens for HaluEval and MedHallu and at 51–100 for RAGTruth, while the longest RAGTruth interval (>100), with the largest checkpoint gaps, has one of the lowest error rates (20.47%), contrary to a spacing-driven pattern.
Table R14: Answer-length distributions before filtering (tokens).
Table R15: Detection error rates (%) across length intervals, averaged over five seeds (↓).
Reference
[1] Janiak, D., Binkowski, J., Sawczyn, A., Gabrys, B., Shwartz-Ziv, R., and Kajdanowicz, T. J. "The illusion of progress: Re-evaluating hallucination detection in LLMs." EMNLP 2025.