Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMs

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that the cross-layer pattern of a model's hidden-state updates, captured in a new ICR Score, is a stable hallucination signal, and that a 16,000-parameter probe trained on it beats five baselines on most benchmarks tested.

desk verdict A genuinely new residual-stream signal for hallucination detection, but the paper's own appendix numbers break its transferability claim. read the letter →

arxiv 2507.16488 v1 pith:Q4VAYN5G submitted 2025-07-22 cs.CL cs.AI

Zhenliang Zhang , Xinyu Hu , Huixuan Zhang , Junzhe Zhang , Xiaojun Wan This is my paper · ORCID
classification cs.CLcs.AI
keywords ICRProbehallucinationdetectionresidualstreamhiddenstatedynamicsinformationcontributionJensen-Shannondivergencelayer-wisefeaturesLLMinterpretability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's claim is that the way an LLM's hidden states change from layer to layer, not the static hidden states themselves, carries a reliable fingerprint of whether the model is about to generate a hallucination. To capture that fingerprint it defines the ICR Score, which measures at each layer how much of the hidden-state update points along the context tokens' hidden states relative to how those tokens were weighted by attention; a score near the attention distribution means attention dominates the update, and a score far from it means the feed-forward network dominates. Pooling the scores across all layers into a per-layer vector and feeding it to a tiny 16,000-parameter neural network, the ICR Probe reports higher AUROC than five baselines on most of the twelve model-dataset combinations, and it transfers to unseen datasets with a smaller performance drop than its strongest competitors. A sympathetic reader would care because this would mean hallucination detection can run in one forward pass, without reference answers or multiple generations, using a classifier small enough to be nearly free.

What carries the argument

The central object is the ICR Score, defined for token $i$ at layer $\ell$ as $\mathrm{ICR}_i^\ell = \mathrm{JSD}(\mathrm{Proj}_i^\ell, \mathrm{Attn}_i^\ell)$, where $\mathrm{Proj}_i^\ell$ is the softmax over the projection lengths $p_{i,j}^\ell = (\Delta x_i^\ell)^\top x_j^\ell / \lVert x_j^\ell \rVert$ of the residual-stream update onto every context token's hidden state, and $\mathrm{Attn}_i^\ell$ is the attention-score vector averaged over heads, restricted to the top-$k$ tokens, with the paper fixing $k = 20$. The interpretation the argument rests on is that this scalar measures the relative dominance of the two components of the update $\Delta x_i^\ell = a_i^\ell + m_i^\ell$: the attention contribution $a_i^\ell$, which redistributes contextual information, and the FFN contribution $m_i^\ell$, which retrieves parametric knowledge. The ICR Probe then token-averages the $N \times L$ matrix of scores into one scalar per layer, and a four-layer MLP with roughly 16,000 parameters maps that $1 \times L$ vector to a faithfulness probability.

What would settle it

Causally remove the feed-forward contribution at the middle layers (set $m_i^\ell = 0$ during a forward pass) and recompute the ICR profile: the paper's account predicts the scores at those layers should collapse toward the attention distribution and the probe's detection signal should weaken once the model's normal computation is disrupted. Alternatively, run a causal mediation analysis that ablates attention and FFN separately at each layer and compare the resulting influence profile with the ICR layer curve; if the layers the ICR Score labels FFN-dominated do not match the layers where FFN ablation demonstrably changes the output, the metric is not measuring module contribution and the detection results would have to be explained another way.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the pattern of residual-stream updates separates hallucinated from faithful generations in a stable, transferable way. For each token and layer, the ICR Score is the Jensen-Shannon divergence between the softmax-normalized projection of the update $\Delta x_i^\ell$ onto all context-token hidden states and the multi-head-averaged attention-score distribution, restricted to the top-$k$ attended tokens. A small ICR Score indicates the update is aligned with the attention distribution, meaning the attention module predominantly drives the update; a large score indicates the feed-forward network dominates, injecting parametric knowledge. The paper shows empirically that this profile is consistent across four datasets with narrow variance bands, that single-layer detection on HaluEval with Qwen2.5 exceeds 0.7 AUROC in ten layers with a peak of 0.7690 at layer 11, and that a probe trained on the token-averaged $1 \times L$ vector of scores outperforms PPL, LN-Entropy, LLM-Check, SAPLMA, and SEP on most model-dataset combinations, with a cross-dataset AUROC drop of 8.61 percent on Gemma-2 versus 10.18 percent for SAPLMA and 11.67 percent for SEP.

Load-bearing premise

Everything turns on the assumption that the direction of the hidden-state update, compared against the averaged attention distribution, cleanly separates what the attention module contributed from what the feed-forward network contributed; if that separation is an illusion, the mechanistic story collapses even if the detector keeps working.

Editorial extensions

If this is right

  • Detection runs on a single forward pass: no reference answers and no sampling of multiple generations, so hallucinations can be flagged in real time where entailment-style or consistency-check methods cannot run at all.
  • The signal is transferable: on the Gemma-2 experiments the probe keeps AUROC at or above 0.66 on unseen datasets, and its average cross-dataset AUROC drop is 8.61 percent, below SAPLMA's 10.18 percent and SEP's 11.67 percent.
  • Parameter count is not what does the work: with roughly 16K parameters against SAPLMA's 110K, the ICR Probe still beats the five baselines on most of the twelve model-dataset configurations.
  • Middle layers carry the discriminative power: removing layers 15-28 degrades AUROC most (Table 3), matching the claim that FFN-dominated knowledge injection peaks in the intermediate layers.
  • At token level, key answer-bearing tokens drive the signal while frequent tokens behave as noise, which both explains the detector's failures and sets up future per-token refinements.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves mitigation to future work; the natural next test is to steer the residual stream at the layers the ICR profile flags as FFN-dominated (roughly layers 10-15 in Qwen2.5) and check whether generation faithfulness shifts in the predicted direction, turning the detector into an intervention target.
  • The per-layer ICR profile could function as a mechanistic fingerprint beyond hallucination: any failure mode that disturbs the attention-versus-FFN balance of information flow, such as contradictions, sycophancy, or prompt-injection susceptibility, might leave a signature in the same profile.
  • A check the paper itself does not run would settle the attribution claim: compare the ICR layer profile against causal mediation analysis that knocks out attention or FFN modules layer by layer; the ablation study validates detection performance, not the claim that the score accurately separates the two modules' contributions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces the ICR Score, defined as the Jensen-Shannon divergence between the projection of the residual-stream update onto context-token hidden states and the averaged attention distribution, and the ICR Probe, a small MLP trained on the layer-wise averaged ICR vector. The method is evaluated for hallucination detection on three LLMs (Llama-3-8B-Instruct, Qwen2.5-7B-Instruct, Gemma-2-9B-it) and four datasets, with AUROC comparisons against PPL, LN-Entropy, LLM-Check, SAPLMA, and SEP. The paper also reports cross-dataset generalization, ablations on the components and layers, and a token-level case study. The central empirical claim is that the ICR Probe, with roughly 16K parameters, outperforms prior methods on most model-dataset combinations and transfers across datasets better than the compared probes.

Significance. If the results hold, the paper would make a useful contribution: it proposes a compact, reference-free, single-forward-pass hallucination detector and identifies residual-stream update dynamics as a promising signal. The paper's strengths include a public code release, ablations on the metric components and layer groups, evaluation across three model families and two additional model scales, and a stated computational budget. However, the significance is currently conditional because the cross-dataset generalization claim is contradicted by the paper's own appendix tables, and the reported performance comparisons lack error bars or significance tests, making it impossible to assess whether the observed AUROC advantages are reliable. The mechanistic interpretation of the ICR Score is also asserted rather than validated, so the interpretability contribution is not yet established.

major comments (4)
  1. [Section 5.3 and Appendix C.4] The generalization claim in Section 5.3 that "Across unseen datasets, our ICR Probe maintains AUROC ≥ 0.66" is not supported by the paper's own tables. Appendix C.4 states that Tables 10 and 9 present Llama-3 and Gemma-2 results and that "AUROC values for both models exceed 0.6," but as printed Table 9 is labeled Qwen2.5 and Table 10 is labeled Llama-3. Table 10 contains off-diagonal AUROC values of 0.5658 (SQuAD→TriviaQA) and 0.5933 (HotpotQA→TriviaQA), both below the 0.66 floor claimed in Section 5.3 and below the 0.6 floor claimed in the appendix. This is an internal contradiction that directly affects the paper's transferability claim. The authors must correct the table labels, report a Gemma-2 table if that is the intended scope, and restate the generalization claim to match the actual numbers.
  2. [Section 5.1, Table 1, and Appendix B.4] The performance comparisons in Table 1 are reported without error bars, confidence intervals, or significance tests, even though Appendix B.4 states that "To obtain stable and reliable results, we perform multiple runs and take the average." Without variance information, it is unclear whether the reported AUROC advantages are statistically meaningful; some margins are small, for example Llama-3 on TriviaQA (0.7325 for ICR Probe vs. 0.7116 for SEP) and Qwen2.5 on HotpotQA (0.7917 vs. 0.7750). The central claim of outperforming baselines therefore needs either repeated-run standard deviations with significance tests or a clear statement of the number of runs and the variability.
  3. [Section 3.2, Equations 6 and 7] The paper repeatedly states that the ICR Score "quantifies the contribution of each module" and interprets small versus large ICR scores as indicating attention- versus FFN-dominated updates. However, the construction in Equations 6 and 7 measures the alignment between the overall residual-stream update and the attention distribution; it does not decompose the update into attention and FFN components, nor does it validate that this alignment tracks the actual relative contributions of the two modules. The ablation in Table 2 shows that the components are useful for detection, but it does not establish the mechanistic claim. The authors should either provide a validation of the proxy against known module contributions (for example, by intervening on attention or FFN outputs and measuring ICR changes) or explicitly reframe the interpretation as a heuristic, not a quantified mechanistic decomposition.
  4. [Appendix B.4, Algorithm 2] Algorithm 2 is inconsistent with the architecture described in the main text and in Appendix B.4. The algorithm initializes "probe weights θ (logistic regression)" and updates a linear model with sigmoid(θᵀf), while the paper describes a four-layer MLP with hidden sizes 128, 64, 32 and LeakyReLU activations. This makes the training procedure unclear and the reported parameter count difficult to verify. Please align the pseudocode with the actual architecture, or state clearly that Algorithm 2 is a simplified abstraction.
minor comments (5)
  1. [Appendix A.2 / Algorithm 1] The JSD in Equation 7 is computed on top-k truncated vectors Proj_i^l[S] and a_i^l[S], but the paper does not state that these sub-vectors are renormalized after truncation. If they are not renormalized, they are not probability distributions and JSD is not well-defined; please clarify the normalization step.
  2. [Section 5.4, Table 3] The layer ablation is described only as "remove a specific layer group," but it is not specified whether the ICR features for those layers are zeroed, the layers are bypassed in the LLM, or the probe input is truncated at those positions. Please provide the exact protocol.
  3. [Appendix C.4] The sentence "Tables 10 and 9 present the cross-dataset generalization evaluation results for Llama-3 and Gemma-2, respectively" should read "Tables 9 and 10," and the table captions in the appendix should be checked so that the model names match the content; as printed, one table labeled Qwen2.5 and one labeled Llama-3 are referenced as Gemma-2 and Llama-3.
  4. [Section 5.1] The text says the experimental setup for baselines "matches ours exactly," but several baselines have different operational requirements (e.g., LLM-Check uses an auxiliary LLM in black-box settings, SEP uses multiple generations). Please specify exactly what was matched, especially regarding training data, train-test splits, and generation settings.
  5. [Section 5.5] The token-level case study applies a sequence-level probe to individual tokens and interprets the output probability as token correctness, but the probe was not trained or calibrated for token-level decisions. This should be stated more prominently in the section, and the word "misdetection" for common tokens should be framed as an artifact of applying a sequence-level model at the token level.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ICR Score is an operational metric and the probe is evaluated against external hallucination labels.

full rationale

The paper's central claim is that a layer-averaged ICR-score vector, compressed to a 1×L representation and fed into a 16K-parameter MLP, carries transferable hallucination-detection signal. This claim is not made true by construction: the ICR Score is an operational definition (Equation 7, JSD(Proj_i^ell, Attn_i^ell)) rather than a parameter fitted to hallucination labels; the ICR Probe is a standard supervised classifier trained with binary cross-entropy on external dataset labels; and the headline AUROC numbers in Table 1, the cross-dataset heatmaps, and the ablation settings are all evaluated against ground-truth hallucination annotations outside the metric's own definition. No step reduces a 'prediction' to fitted constants or to the metric's definition: the interpretive statement that low ICR means attention-dominated updates and high ICR means FFN-dominated updates is an empirical/interpretive assignment, not an equation forced by Equation 7. The paper contains no load-bearing self-citation: the mechanistic motivations cite Stolfo et al., Elhage et al., and Geva et al., all external works, and no 'uniqueness theorem' is invoked to forbid alternatives. Hyperparameter choices such as k=20 and the four-hidden-layer probe architecture were selected by ablations on the same benchmark families, which is a mild selection/overfitting concern for generalization claims, but not circularity, because the reported numbers are measured rather than derived. Finally, Appendix C.4's statement that AUROC values for both models exceed 0.6 is contradicted by Table 10's off-diagonal cells (SQuAD→TriviaQA 0.5658 and HotpotQA→TriviaQA 0.5933 for Llama-3); this is an internal-consistency and correctness flaw in the transferability claim, not a circular-reasoning flaw.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central detection mechanism introduces one tuned hyperparameter (k) and one architecture choice, plus several domain assumptions about how residual stream deltas relate to module contributions. No new physical or model-level entities are postulated.

free parameters (2)
  • k (top-k tokens) = 20
    Number of highest-attention tokens used in the JSD computation; selected via ablation on Gemma-2 over the four test datasets (Appendix B.3) and then used for all main results.
  • Probe hidden architecture widths = 128-64-32
    Chosen by preliminary experiments on Gemma-2 and TriviaQA (Appendix B.4, Table 6); this affects the parameter count and the reported performance.
assumptions (3)
  • domain assumption The transformer residual stream update at each layer can be written as x_l = x_{l-1} + MHSA(...) + FFN(...), with layer normalization and other sublayer details ignored.
    Equation (1) in Section 3.1; the projection-based ICR Score depends on this simplified decomposition, whereas real models apply norms and sublayers.
  • domain assumption The consistency between the hidden-state update projection distribution and the averaged attention-score distribution is a valid proxy for the relative information contribution of attention and FFN modules.
    Section 3.2, Equations 6 and 7; no derivation or intervention experiment establishes this link, and the interpretability claims rely on it.
  • ad hoc to paper Averaging attention scores across heads before softmax represents the contextual information routing used by the model.
    Equations 3 to 5; standard transformers apply softmax per head before averaging weights, so pre-softmax averaging is a design choice not justified in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMs." pith.science (2026). https://pith.science/paper/Q4VAYN5G

@misc{pith2026250716488,
  author       = {Pith},
  title        = {Pith review of: ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q4VAYN5G}},
  note         = {Machine review of arXiv:2507.16488}
}
read the original abstract

Large language models (LLMs) excel at various natural language processing tasks, but their tendency to generate hallucinations undermines their reliability. Existing hallucination detection methods leveraging hidden states predominantly focus on static and isolated representations, overlooking their dynamic evolution across layers, which limits efficacy. To address this limitation, we shift the focus to the hidden state update process and introduce a novel metric, the ICR Score (Information Contribution to Residual Stream), which quantifies the contribution of modules to the hidden states' update. We empirically validate that the ICR Score is effective and reliable in distinguishing hallucinations. Building on these insights, we propose a hallucination detection method, the ICR Probe, which captures the cross-layer evolution of hidden states. Experimental results show that the ICR Probe achieves superior performance with significantly fewer parameters. Furthermore, ablation studies and case analyses offer deeper insights into the underlying mechanism of this method, improving its interpretability.

Figures

Figures reproduced from arXiv: 2507.16488 by the authors.

Figure 1
Figure 1. Overview of the ICR Score computation and ICR Probe detection process. We first probe each layer’s [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Mean ICR scores with standard deviation bands across four datasets. The narrow standard de￾viation bands indicate that the ICR score captures the residual stream patterns with strong cross-dataset con￾sistency and stability. mechanism contributing less to the informa￾tion stream. The ICR score captures the relative contribution of each module across L layers and N tokens, re￾vealing how attention and FFN modules sha… view at source ↗
Figure 3
Figure 3. ICR Score for hallucination detection: (a) Layer-wise AUROC and (b) probability density distributions, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Cross-dataset generalization heatmaps for ICR Probe, SAPLMA, and SEP. Each subplot displays the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Layer-wise mean ICR scores. Hidden layers 1 2 3 4 5 Performance 0.6410 0.7398 0.7889 0.8000 0.8006 [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Layer-wise AUROC. LLM HaluEval SQuAD HotpotQA TriviaQA Gemma-2 0.7850 0.7321 0.7465 0.7883 Qwen2.5 0.7592 0.6903 0.7345 0.7582 Llama-3 0.7458 0.7331 0.7340 0.7206 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. The Anatomy of a Truth Direction: Knowledge-Dependent Dimensionality, a Relational Law, and a Convergent Category Geometry in Small Language Models

    cs.LG 2026-07 conditional novelty 6.0 of 10

    In small LLMs, truth is a mostly one-dimensional, knowledge-gated internal direction built by attention and eroded by the feed-forward value stream; per-topic truth axes converge across model families.

Reference graph

Works this paper leans on

33 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  4. [4]

    Amos Azaria and Tom Mitchell. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.68 The Internal State of an LLM Knows When It `s Lying . In Findings of the Association for Computational Linguistics : EMNLP 2023 , pages 967--976, Singapore. Association for Computational Linguistics

    doi:10.18653/v1/2023.findings-emnlp.68 2023

  5. [5]

    Sky CH-Wang, Benjamin Van Durme, Jason Eisner, and Chris Kedzie. 2024. https://doi.org/10.18653/v1/2024.findings-acl.260 Do androids know they`re only dreaming of electric sheep? In Findings of the Association for Computational Linguistics: ACL 2024, pages 4401--4420, Bangkok, Thailand. Association for Computational Linguistics

    doi:10.18653/v1/2024.findings-acl.260 2024

  6. [6]

    Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. 2024 a . Inside: Llms' internal states retain the power of hallucination detection

    work page 2024

  7. [7]

    Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. 2024 b . https://api.semanticscholar.org/CorpusID:267499843 Inside: Llms' internal states retain the power of hallucination detection . ArXiv, abs/2402.03744

    arXiv 2024

  8. [8]

    Shiqi Chen, Miao Xiong, Junteng Liu, ZhengXuan Wu, Teng Xiao, Siyang Gao, and Junxian He. 2024 c . In-context sharpness as alerts: An inner representation perspective for hallucination mitigation. In Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, pages 7553--7567

    work page 2024

Show all 33 references
  1. [9]

    Glass, and Pengcheng He

    Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James R. Glass, and Pengcheng He. 2024. https://openreview.net/forum?id=Th6NyL07na Dola: Decoding by contrasting layers improves factuality in large language models . In The Twelfth International Conference on Learning Repres...

    2024

  2. [10]

    Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dari...

    2021

  3. [11]

    Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.751 Dissecting recall of factual associations in auto-regressive language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language ...

    2023 doi

  4. [12]

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.446 Transformer feed-forward layers are key-value memories . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484--5495, ...

    2021 doi

  5. [13]

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2025. https://doi.org/10.1145/3703155 A survey on hallucination in large language models: Principles, taxonomy, challenges, and o...

    2025 doi

  6. [14]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. https://doi.org/10.1145/3571730 Survey of hallucination in natural language generation . ACM Comput. Surv., 55(12)

    2023 doi

  7. [15]

    Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. https://doi.org/10.18653/v1/P17-1147 T rivia QA : A large scale distantly supervised challenge dataset for reading comprehension . In Proceedings of the 55th Annual Meeting of the Association for Computational...

    2017 doi

  8. [16]

    Malik, and Yarin Gal

    Jannik Kossen, Jiatong Han, Muhammed Razzak, Lisa Schut, Shreshth A. Malik, and Yarin Gal. 2024. https://api.semanticscholar.org/CorpusID:270703114 Semantic entropy probes: Robust and cheap hallucination detection in llms . ArXiv, abs/2406.15927

    2024 arXiv

  9. [17]

    Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv preprint arXiv:2302.09664

    2023 arXiv

  10. [18]

    Junyi Li, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.397 H alu E val: A large-scale hallucination evaluation benchmark for large language models . In Proceedings of the 2023 Conference on Empirical Methods in Natur...

    2023 doi

  11. [19]

    J. Lin. 1991. https://doi.org/10.1109/18.61115 Divergence measures based on the shannon entropy . IEEE Transactions on Information Theory, 37(1):145--151

    1991 doi

  12. [20]

    Andrey Malinin and Mark Gales. 2020. Uncertainty Estimation in Autoregressive Structured Prediction . In International Conference on Learning Representations

    2020

  13. [21]

    Potsawee Manakul, Adian Liusie, and Mark Gales. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.557 S elf C heck GPT : Zero-resource black-box hallucination detection for generative large language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural...

    2023 doi

  14. [22]

    Hadas Orgad, Michael Toker, Zorik Gekhman, Roi Reichart, Idan Szpektor, Hadas Kotek, and Yonatan Belinkov. 2024. https://arxiv.org/abs/2410.02707 Llms know more than they show: On the intrinsic representation of llm hallucinations . Preprint, arXiv:2410.02707

    2024 arXiv

  15. [23]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. https://doi.org/10.18653/v1/D16-1264 SQ u AD : 100,000+ questions for machine comprehension of text . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-...

    2016 doi

  16. [24]

    Jie Ren, Jiaming Luo, Yao Zhao, Kundan Krishna, Mohammad Saleh, Balaji Lakshminarayanan, and Peter J. Liu. 2022. Out-of- Distribution Detection and Selective Generation for Conditional Language Models . In The Eleventh International Conference on Learning Representations

    2022

  17. [25]

    Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed Chi, Nathanael Sch\" a rli, and Denny Zhou. 2023. Large language models can be easily distracted by irrelevant context. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org

    2023

  18. [26]

    Gaurang Sriramanan, Siddhant Bharti, Vinu Sankar Sadasivan, Shoumik Saha, Priyatham Kattakinda, and Soheil Feizi. 2024. LLM-Check : Investigating Detection of Hallucinations in Large Language Models . In Advances in Neural Information Processing Systems , volume 37, pages 3418...

    2024

  19. [27]

    Alessandro Stolfo, Yonatan Belinkov, and Mrinmaya Sachan. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.435 A mechanistic interpretation of arithmetic reasoning in language models using causal mediation analysis . In Proceedings of the 2023 Conference on Empirical Methods ...

    2023 doi

  20. [28]

    ZhongXiang Sun, Xiaoxue Zang, Kai Zheng, Jun Xu, Xiao Zhang, Weijie Yu, Yang Song, and Han Li. 2024. ReDeEP : Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability . In The Thirteenth International Conference on Learning Representations

    2024

  21. [29]

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Le Lan...

    2024 arXiv

  22. [30]

    Qwen Team. 2024. https://qwenlm.github.io/blog/qwen2.5/ Qwen2.5: A party of foundation models

    2024

  23. [31]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. https://doi.org/10.18653/v1/D18-1259 H otpot QA : A dataset for diverse, explainable multi-hop question answering . In Proceedings of the 2018 Conference...

    2018 doi

  24. [32]

    Yuheng Zha, Yichi Yang, Ruichen Li, and Zhiting Hu. 2023. Alignscore: Evaluating factual consistency with a unified alignment function. arXiv preprint arXiv:2305.16739

    2023 arXiv

  25. [33]

    Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.