pith. sign in

arxiv: 2605.17028 · v1 · pith:XNYIUZVZnew · submitted 2026-05-16 · 💻 cs.CL · cs.AI

PARALLAX: Separating Genuine Hallucination Detection from Benchmark Construction Artifacts

Khizar Hussain , Murat Kantarcioglu This is my paper

Pith reviewed 2026-05-19 20:06 UTC · model grok-4.3

classification 💻 cs.CL cs.AI
keywords hallucination detectionLLM benchmarksbenchmark artifactstext similarityhidden state probesDRIFTSAPLMAlarge language models
0
0 comments X

The pith

Benchmark artifacts explain most reported success in LLM hallucination detection.

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

The paper shows that four of six widely used corpora for hallucination detection embed the ground-truth answer directly inside the input prompt. A simple text-similarity baseline called TxTemb exploits this leakage to reach near-perfect detection scores without ever inspecting model internals. Once those embedded answers are removed, the majority of established detection methods fall to near-chance performance across twelve models and twenty-two techniques. Only supervised probes that read upper-layer hidden states, such as SAPLMA and the newly introduced DRIFT, maintain consistent results under the controlled conditions.

Core claim

Much of the field's reported progress on hallucination detection is substantially explained by benchmark construction artifacts in widely used corpora. Four of the six corpora embed the ground-truth answer directly in the input prompt. A naïve text-similarity baseline called TxTemb exploits this to achieve near-perfect detection scores without any access to model internals. Under controlled conditions without these artifacts, the majority of established baselines perform near chance; the consistent exceptions are SAPLMA and DRIFT, both supervised probes on upper-layer hidden states.

What carries the argument

The TxTemb text-similarity baseline, which measures overlap between the input prompt and the model output to flag hallucinations and thereby exposes the ground-truth leakage present in four of the six corpora.

If this is right

  • Most current detection methods will not transfer to settings where prompts contain no embedded answers.
  • Reliable detection will likely require methods that inspect internal model states rather than output similarity alone.
  • DRIFT provides one concrete supervised probe over inter-layer hidden-state transitions that works for live generation.
  • Benchmark creators must prevent ground-truth leakage when building future evaluation sets.
  • Published performance numbers on the original corpora should be discounted until re-tested without the artifacts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Once benchmarks are cleaned of these leaks, the field may discover that unsupervised or output-only detection is even harder than current numbers suggest.
  • This finding directly affects safety claims for LLMs in medical, legal, or scientific use where undetected hallucinations carry real cost.
  • The same construction artifacts could exist in other LLM evaluation tasks that rely on prompt-based ground truth.
  • Extending hidden-state probes like DRIFT to additional model families could yield practical real-time detection tools.

Load-bearing premise

Removing the embedded ground-truth answers from the prompts creates a fair test of genuine hallucination detection capability rather than simply creating a harder or differently biased evaluation setup.

What would settle it

Construct new hallucination detection corpora that explicitly avoid placing any ground-truth answers in the prompts, then re-run the full suite of twenty-two methods and check whether only the hidden-state probes retain high performance while all others stay near chance.

Figures

Figures reproduced from arXiv: 2605.17028 by Khizar Hussain, Murat Kantarcioglu.

Figure 1
Figure 1. Figure 1: DRIFT architecture. Hidden states are tapped at four upper-layer positions (60–85% depth), mean-pooled over T tokens, and differenced across all C(4, 2)=6 pairs to form ϕab ∈ Rd+2, then concatenated into z ∈ R49,164. An L2-regularised logistic probe maps z to P(hallucination); its weight vector wE doubles as a hallucination contrastive direction. Dashed box: baseline comparisons. twelve instruction-tuned m… view at source ↗
Figure 2
Figure 2. Figure 2: Two evaluation regimes. Left: Teacher-forced benchmarks embed the answer in the prompt; TXTEMB achieves AUROC 0.69–0.98 and explains 81% of variance in method AUROC (R 2=0.81), a formatting artifact. Right: On live-generation benchmarks TXTEMB is at chance (≈0.50, R 2=0.08), isolating genuine internal-state signal. 0.69, which still requires factual knowledge to distinguish correct from incorrect answers) … view at source ↗
Figure 3
Figure 3. Figure 3: AUROC heatmap: 22 methods × 6 corpora on Llama-3.3-70B. Left block (teacher-forced) shows uniformly high val￾ues; right block (live-generation) shows the artifact-free detection ceiling. Model TruthfulQA HaluBench DRIFT SAPLMA DRIFT-concat SC† DRIFT SAPLMA DRIFT-concat SC† Llama-3.3-70B 0.729 0.751 0.737 0.558 0.915 0.911 0.911 0.535 Llama-3.1-8B 0.628 0.656 0.664 0.535 0.839 0.824 0.823 0.579 Qwen3-8B 0.6… view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy–efficiency Pareto frontier for internal-state probes on TruthfulQA with Llama-3.3- [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: ROC curves for the six strongest methods on TruthfulQA (left) and RAGTruth (right) with [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Stacker (A+C+E+F) vs. best individual method AUROC as a function of training set size [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: AUROC vs. training set size N for DRIFT, SAPLMA, DRIFT-concat, and LT-SVD on TruthfulQA (Llama-3.3-70B). Results are 10-seed averages with ±1 std shaded. DRIFT and SAPLMA both reach near-peak performance at N=250; LT-SVD does not reliably exceed chance below N=500. The dotted vertical line marks N=250. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
read the original abstract

Large language models (LLMs) hallucinate with confidence: their outputs can be fluent, authoritative, and simply wrong. In medical, legal, and scientific applications this failure causes direct harm, and detecting it from internal model states offers a path to safer deployment. A growing body of work reports that this problem is increasingly tractable, with recent methods achieving high detection performance on widely used benchmarks. We show, however, that much of this apparent progress does not survive scrutiny. Four of the six corpora embed the ground-truth answer directly in the input prompt. A na\"{i}ve text-similarity baseline we call \textsc{TxTemb} exploits this to achieve near-perfect detection scores without any access to model internals. To measure what genuine detection capability remains once these artifacts are controlled, we conduct a large-scale evaluation spanning twenty-two detection methods, twelve open-source models spanning six architectural families, and six corpora. We further introduce \textbf{DRIFT}, a supervised probe over inter-layer hidden-state transitions, as a point of comparison for live-generation detection. Our findings suggest that the field's reported progress on hallucination detection is substantially explained by benchmark construction artifacts in widely used corpora, and that the majority of established baselines perform near chance under controlled conditions; the consistent exceptions are SAPLMA and DRIFT, both supervised probes on upper-layer hidden states.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper claims that four of six widely used hallucination detection corpora embed ground-truth answers directly in the input prompts, allowing a naive text-similarity baseline (TxTemb) to achieve near-perfect scores without model internals. After excising these embedded answers to create controlled conditions, a large-scale evaluation across 22 detection methods, 12 open-source models from six families, and the six corpora shows that most established baselines perform near chance; the consistent exceptions are SAPLMA and the newly introduced DRIFT (a supervised probe on inter-layer hidden-state transitions). The central conclusion is that reported progress on hallucination detection is substantially explained by benchmark construction artifacts.

Significance. If the controlled evaluation is robust, the work is significant for exposing how prompt artifacts can inflate detection performance and for showing that genuine internal-state detection remains difficult for most methods. The scale of the evaluation and the introduction of DRIFT as a live-generation comparison point are strengths that could help redirect the field toward artifact-free benchmarks and methods that truly rely on model internals rather than surface cues.

major comments (1)
  1. [Evaluation on cleaned corpora] The central claim that 'the majority of established baselines perform near chance under controlled conditions' and that progress is 'substantially explained by benchmark construction artifacts' depends on the cleaned prompts (with ground-truth answers removed from the four affected corpora) constituting a fair test of genuine hallucination detection. Removing embedded answers necessarily alters prompt length, structure, and information content, which may shift generation behavior, output distributions, or residual statistical patterns in ways that upper-layer probes like DRIFT or SAPLMA could still exploit. Without explicit controls or analysis demonstrating that these shifts do not introduce new confounds, the conclusion that only supervised hidden-state methods succeed for non-artifactual reasons is not fully supported.
minor comments (2)
  1. [Abstract] The abstract states that the evaluation spans 'twenty-two detection methods' and 'six corpora' but provides no table or section reference listing the exact methods, models, or corpora used; adding such a summary table would improve reproducibility and allow readers to verify coverage.
  2. [DRIFT introduction] The description of DRIFT as 'a supervised probe over inter-layer hidden-state transitions' is introduced without a precise definition of the transition features or training procedure in the provided text; a short formal definition or pseudocode would clarify how it differs from SAPLMA.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for this constructive comment on our evaluation methodology. We address the concern directly below and have revised the manuscript to incorporate additional controls.

read point-by-point responses

  1. Referee: [Evaluation on cleaned corpora] The central claim that 'the majority of established baselines perform near chance under controlled conditions' and that progress is 'substantially explained by benchmark construction artifacts' depends on the cleaned prompts (with ground-truth answers removed from the four affected corpora) constituting a fair test of genuine hallucination detection. Removing embedded answers necessarily alters prompt length, structure, and information content, which may shift generation behavior, output distributions, or residual statistical patterns in ways that upper-layer probes like DRIFT or SAPLMA could still exploit. Without explicit controls or analysis demonstrating that these shifts do not introduce new confounds, the conclusion that only supervised hidden-state methods succeed for non-artifactual reasons is not fully supported.

    Authors: We agree that excising the embedded ground-truth answers alters prompt length, structure, and information content, and that this could in principle introduce new confounds. The original manuscript already reports performance on both original and cleaned versions of the four affected corpora to isolate the contribution of the artifact. To strengthen the claim, the revised manuscript adds explicit controls: we normalize prompt lengths across conditions via padding/truncation, compute lexical diversity and n-gram overlap statistics, and verify that output token distributions do not exhibit new patterns that correlate with DRIFT or SAPLMA scores. Under these controls, the majority of baselines remain near chance while SAPLMA and DRIFT retain their advantage, consistent with their reliance on internal hidden-state transitions rather than surface cues. We have expanded the relevant subsection of the experimental analysis and added a limitations paragraph acknowledging residual prompt effects. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical artifact audit with external baselines

full rationale

The paper identifies embedded ground-truth answers in four corpora by direct inspection, constructs the naive TxTemb similarity baseline to exploit that artifact, removes the answers to produce controlled prompts, and then reports performance of 22 methods (including prior SAPLMA and the newly introduced DRIFT probe) on the modified data. No equation or result is obtained by fitting a parameter to the target detection scores and relabeling it a prediction; no uniqueness theorem or ansatz is imported via self-citation to force the conclusion; and the central claim rests on comparative empirical numbers rather than any self-definitional reduction. The evaluation is therefore self-contained against the external benchmarks and baselines.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claims rest primarily on empirical re-evaluation of existing corpora and methods with minimal additional postulates beyond standard assumptions in LLM evaluation.

axioms (1)
  • domain assumption Hallucinations can be detected from internal model states in LLMs
    Invoked to motivate the focus on hidden-state probes such as SAPLMA and DRIFT.
invented entities (1)
  • DRIFT no independent evidence
    purpose: Supervised probe over inter-layer hidden-state transitions for live-generation hallucination detection
    New probe introduced as a point of comparison; no external falsifiable evidence provided in the abstract.

pith-pipeline@v0.9.0 · 5773 in / 1404 out tokens · 42978 ms · 2026-05-19T20:06:20.249089+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

91 extracted references · 91 canonical work pages · 6 internal anchors

  1. [1]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

    Red Teaming Language Models with Language Models , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

    work page 2022

  2. [2]

    Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned

    Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned , author=. arXiv preprint arXiv:2209.07858 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Ness, Robert and Zhang, Junaid and Rao, Anchit and Gryz, Jarek , journal=

    work page

  4. [4]

    Communications Medicine (Nature) , year=

    Multi-Model Assurance Analysis Showing Large Language Models Are Highly Vulnerable to Adversarial Hallucination Attacks During Clinical Decision Support , author=. Communications Medicine (Nature) , year=

    work page

  5. [5]

    Red Teaming

    Chang, Haw-Shiuan and others , journal=. Red Teaming

    work page

  6. [6]

    arXiv preprint arXiv:2602.22775 , year=

    work page arXiv

  7. [7]

    DIAL: Direct Iterative Adversarial Learning for Realistic Multi-Turn Dialogue Simulation

    Adversarial Training for Failure-Sensitive User Simulation in Mental Health Support , author=. arXiv preprint arXiv:2512.20773 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Wang, Peng and others , journal=

    work page

  9. [9]

    arXiv preprint arXiv:2505.08775 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Representation Engineering: A Top-Down Approach to

    Zou, Andy and Phan, Long and Chen, Sarah and Campbell, James and Guo, Phillip and Ren, Richard and Pan, Alexander and Yin, Xuwang and Mazeika, Mantas and Dombrowski, Ann-Kathrin and others , journal=. Representation Engineering: A Top-Down Approach to

    work page

  11. [11]

    Steering

    Panickssery, Nina and Gabrieli, Nick and Schulz, Julian and Tong, Meg and Hubinger, Evan and Turner, Alexander Matt , journal=. Steering

    work page

  12. [12]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Activation Addition: Steering Language Models Without Optimization , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

    work page

  13. [13]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Inference-Time Intervention: Eliciting Truthful Answers from a Language Model , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

    work page

  14. [14]

    The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets

    The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets , author=. arXiv preprint arXiv:2310.06824 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    International Conference on Learning Representations (ICLR) , year=

    Discovering Latent Knowledge in Language Models Without Supervision , author=. International Conference on Learning Representations (ICLR) , year=

    work page

  16. [16]

    The Internal State of an

    Azaria, Amos and Mitchell, Tom , booktitle=. The Internal State of an. 2023 , note=

    work page 2023

  17. [17]

    Adaptive Activation Steering: A Tuning-Free

    Wang, Anqi and others , booktitle=. Adaptive Activation Steering: A Tuning-Free. 2025 , note=

    work page 2025

  18. [18]

    2025 , note=

    Orgad, Hadas and Belinkov, Yonatan and others , booktitle=. 2025 , note=

    work page 2025

  19. [19]

    Proceedings of the International Conference on Machine Learning (ICML) , year=

    How Language Model Hallucinations Can Snowball , author=. Proceedings of the International Conference on Machine Learning (ICML) , year=

    work page

  20. [20]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

    work page

  21. [21]

    Findings of the Association for Computational Linguistics: ACL 2022 , year=

    Why Exposure Bias Matters: An Imitation Learning Perspective of Error Accumulation in Language Generation , author=. Findings of the Association for Computational Linguistics: ACL 2022 , year=

    work page 2022

  22. [22]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , year=

    On Exposure Bias, Hallucination and Domain Shift in Neural Machine Translation , author=. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , year=

    work page

  23. [23]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Faith and Fate: Limits of Transformers on Compositionality , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

    work page

  24. [24]

    The Illusion of Diminishing Returns: Measuring Long Horizon Execution in

    Sinha, Aaditya and others , journal=. The Illusion of Diminishing Returns: Measuring Long Horizon Execution in

    work page

  25. [25]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Factuality Enhanced Language Models for Open-Ended Text Generation , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

    work page

  26. [26]

    Nature , year=

    Detecting Hallucinations in Large Language Models Using Semantic Entropy , author=. Nature , year=

    work page

  27. [27]

    Beyond Exponential Decay: Rethinking Error Accumulation in Large Language Models

    Beyond Exponential Decay: Rethinking Error Accumulation in Autoregressive Models , author=. arXiv preprint arXiv:2505.24187 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Steering When Necessary: Flexible Steering Large Language Models with Backtracking , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

    work page

  29. [29]

    2024 , note=

    Su, Weili and others , booktitle=. 2024 , note=

    work page 2024

  30. [30]

    2024 , note=

    Chen, Chao and others , booktitle=. 2024 , note=

    work page 2024

  31. [31]

    2024 , note=

    Chuang, Yung-Sung and Xie, Yujia and Luo, Hongyin and Kim, Yoon and Glass, James and He, Pengcheng , booktitle=. 2024 , note=

    work page 2024

  32. [32]

    arXiv preprint arXiv:2602.21704 , year=

    Dynamic Multimodal Activation Steering , author=. arXiv preprint arXiv:2602.21704 , year=

    work page arXiv

  33. [33]

    arXiv preprint arXiv:2405.19648 , year=

    Token-Level Hallucination Detection via Simple Features , author=. arXiv preprint arXiv:2405.19648 , year=

    work page arXiv

  34. [34]

    2025 , doi=

    Asgari, Elham and others , journal=. 2025 , doi=

    work page 2025

  35. [35]

    Being Kind Isn't Always Being Safe: Diagnosing Affective Hallucination in

    Kim, Hyeonseok and others , journal=. Being Kind Isn't Always Being Safe: Diagnosing Affective Hallucination in

    work page

  36. [36]

    arXiv preprint arXiv:2412.18947 , year=

    work page arXiv

  37. [37]

    Available: https://arxiv.org/abs/2503.05777

    Medical Hallucination in Foundation Models and Their Impact on Healthcare , author=. arXiv preprint arXiv:2503.05777 , year=

    work page arXiv

  38. [38]

    arXiv preprint arXiv:2510.19032 , year=

    When Can We Trust. arXiv preprint arXiv:2510.19032 , year=

    work page arXiv

  39. [39]

    PMC , year=

    Charting the Evolution of Artificial Intelligence Mental Health Chatbots from Rule-Based Systems to Large Language Models: A Systematic Review , author=. PMC , year=

    work page

  40. [40]

    Ferrando, Javier and de Heredia, Oscar and Betancourt, Aitor Ormazabal , booktitle=. Do. 2025 , note=

    work page 2025

  41. [41]

    Sun, Zhiwei and others , journal=

    work page

  42. [42]

    Dassen, Ian and others , journal=

    work page

  43. [43]

    Steering

    Hua, Mingxuan and others , journal=. Steering

    work page

  44. [44]

    2026 , eprint=

    Toward Faithful Retrieval-Augmented Generation with Sparse Autoencoders , author=. arXiv preprint arXiv:2512.08892 , year=

    work page arXiv

  45. [45]

    arXiv preprint arXiv:2505.12886 , year=

    Detection and Mitigation of Hallucination in Large Reasoning Models: A Mechanistic Perspective , author=. arXiv preprint arXiv:2505.12886 , year=

    work page arXiv

  46. [46]

    arXiv preprint arXiv:2502.17601 , year=

    Representation Engineering for Large-Language Models: Survey and Research Challenges , author=. arXiv preprint arXiv:2502.17601 , year=

    work page arXiv

  47. [47]

    Siren's Song in the

    Zhang, Yue and Li, Yafu and Cui, Leyang and Cai, Deng and Liu, Lemao and Fu, Tingchen and Huang, Xinting and Zhao, Enbo and Zhang, Yu and Chen, Yulong and others , journal=. Siren's Song in the. 2023 , note=

    work page 2023

  48. [48]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL) , pages=

    A Token-level Reference-free Hallucination Detection Benchmark for Free-form Text Generation , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL) , pages=

    work page

  49. [49]

    Niu, Cheng and Wu, Yuanhao and Zhu, Juno and Xu, Siliang and Shum, KaShun and Zhong, Randy and Song, Juntong and Zhang, Tong , booktitle=

    work page

  50. [50]

    2023 , note=

    Li, Junyi and Cheng, Xiaoxue and Zhao, Wayne Xin and Nie, Jian-Yun and Wen, Ji-Rong , journal=. 2023 , note=

    work page 2023

  51. [51]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L

    A Single Direction of Truth: An Observer Model's Linear Residual Probe Exposes and Steers Contextual Hallucinations , author=. arXiv preprint arXiv:2507.23221 , year=

    work page arXiv

  52. [52]

    arXiv preprint arXiv:2509.03531 , year=

    Real-Time Detection of Hallucinated Entities in Long-Form Generation , author=. arXiv preprint arXiv:2509.03531 , year=

    work page arXiv

  53. [53]

    Beyond Token Probes: Hallucination Detection via Activation Tensors with

    Bar-Shalom, Yaniv and others , journal=. Beyond Token Probes: Hallucination Detection via Activation Tensors with

    work page

  54. [54]

    Robust Hallucination Detection in

    Niu, Feng and others , journal=. Robust Hallucination Detection in

    work page

  55. [55]

    arXiv preprint arXiv:2505.11741 , year=

    work page arXiv

  56. [56]

    arXiv preprint arXiv:2512.20949 , year=

    Neural Probe-Based Hallucination Detection for Large Language Models , author=. arXiv preprint arXiv:2512.20949 , year=

    work page arXiv

  57. [57]

    arXiv preprint arXiv:2507.02990 , year=

    Adversarial Jailbreaking in Mental Health Contexts , author=. arXiv preprint arXiv:2507.02990 , year=

    work page arXiv

  58. [58]

    arXiv preprint arXiv:2509.14254 , year=

    Hallucination Detection with the Internal Layers of. arXiv preprint arXiv:2509.14254 , year=

    work page arXiv

  59. [59]

    , booktitle=

    Manakul, Potsawee and Liusie, Adian and Gales, Mark J.F. , booktitle=. 2023 , note=

    work page 2023

  60. [60]

    Lin, Stephanie and Hilton, Jacob and Evans, Owain , booktitle=

    work page

  61. [61]

    Du, Xuefeng and Gao, Chaowei and Li, Shanghang and Li, Yixuan , booktitle=

    work page

  62. [62]

    Dasgupta, Sharanya and Garg, Daksh and Gupta, Manish , journal=

    work page

  63. [63]

    Advances in Neural Information Processing Systems , volume=

    Attention is All You Need , author=. Advances in Neural Information Processing Systems , volume=

    work page

  64. [64]

    Touvron, Hugo and Martin, Louis and Stone, Kevin and Albert, Peter and Almahairi, Amjad and Babaei, Yasmine and Bashlykov, Nikolay and Batra, Soumya and Bhargava, Prajjwal and Bhosale, Shruti and others , journal=

    work page

  65. [65]

    Dubey, Abhimanyu and Jauhri, Abhinav and Pandey, Abhinav and Kadian, Abhishek and Al-Dahle, Ahmad and others , journal=. The

    work page

  66. [66]

    Achiam, Josh and Adler, Steven and Agarwal, Sandhini and Ahmad, Lama and Akkaya, Ilge and others , journal=

    work page

  67. [67]

    Advances in Neural Information Processing Systems , volume=

    Training Language Models to Follow Instructions with Human Feedback , author=. Advances in Neural Information Processing Systems , volume=

    work page

  68. [68]

    Advances in Neural Information Processing Systems , volume=

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. Advances in Neural Information Processing Systems , volume=

    work page

  69. [69]

    ACM Computing Surveys , volume=

    Survey of Hallucination in Natural Language Generation , author=. ACM Computing Surveys , volume=

    work page

  70. [70]

    A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions

    A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions , author=. arXiv preprint arXiv:2311.05232 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  71. [71]

    Proceedings of the 34th International Conference on Machine Learning (ICML) , year=

    On Calibration of Modern Neural Networks , author=. Proceedings of the 34th International Conference on Machine Learning (ICML) , year=

    work page

  72. [72]

    NeurIPS Workshop on Trustworthy and Socially Responsible Machine Learning , year=

    Language Models (Mostly) Know What They Know , author=. NeurIPS Workshop on Trustworthy and Socially Responsible Machine Learning , year=

    work page

  73. [73]

    International Conference on Learning Representations (ICLR) , year=

    Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation , author=. International Conference on Learning Representations (ICLR) , year=

    work page

  74. [74]

    Min, Sewon and Krishna, Kalpesh and Lyu, Xinxi and Lewis, Mike and Yih, Wen-tau and Koh, Pang Wei and Iyyer, Mohit and Zettlemoyer, Luke and Hajishirzi, Hannaneh , booktitle=

    work page

  75. [75]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , year=

    On Faithfulness and Factuality in Abstractive Summarization , author=. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , year=

    work page

  76. [76]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL) , year=

    When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL) , year=

    work page

  77. [77]

    Laban, Philippe and Schnabel, Tobias and Bennett, Paul N and Hearst, Marti A , booktitle=

    work page

  78. [78]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , volume=

    work page

  79. [79]

    Transformer Circuits Thread , year=

    A Mathematical Framework for Transformer Circuits , author=. Transformer Circuits Thread , year=

    work page

  80. [80]

    Locating and Editing Factual Associations in

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , journal=. Locating and Editing Factual Associations in

    work page

Showing first 80 references.