pith. sign in

arxiv: 2606.24790 · v1 · pith:EL5MI7URnew · submitted 2026-06-23 · 💻 cs.LG · cs.AI

Grad Detect: Gradient-Based Hallucination Detection in LLMs

Anand Kamat , Daniel Blake , Brent M. Werness This is my paper

Pith reviewed 2026-06-25 23:57 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords hallucination detectiongradient analysislarge language modelsmodel reliabilityinference-time detectionlayer-wise gradientsabstention prediction
0
0 comments X

The pith

Gradient patterns from a single forward-backward pass reveal whether an LLM output is a hallucination.

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

The paper aims to show that layer-wise gradient patterns computed during one forward-backward pass contain information about output correctness that output probabilities and sampling methods do not provide. This information allows better prediction of hallucinations and abstention decisions on Q&A benchmarks across multiple model families. If correct, the approach would enable low-overhead reliability checks focused on the final layers. A sympathetic reader would care because it offers a way to inspect model internals for reliability without extra model runs or changes.

Core claim

Grad Detect analyzes layer-wise gradient patterns from a single forward-backward pass to predict hallucinations. On Q&A benchmarks it outperforms confidence-based and sampling-based baselines for both hallucination detection and abstention prediction. The final five layers concentrate over 97 percent of the discriminative signal across eleven models from four families.

What carries the argument

Layer-wise gradient patterns obtained from one forward-backward pass on the input and generated output pair.

If this is right

  • Hallucination checks can run with only one extra backward pass at inference time.
  • Deployment can restrict gradient computation to the final five layers with little loss in performance.
  • The same gradient signals support unified prediction of both hallucinations and abstention.
  • Layer ablation reveals where failure signals originate inside the model.

Where Pith is reading between the lines

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

  • The gradient signal concentration in final layers may indicate that output reliability judgments depend mainly on late-stage processing.
  • Gradient patterns could be tested for detecting other output errors such as logical inconsistencies.
  • Hybrid detectors that combine gradients with existing output signals might yield further gains.

Load-bearing premise

Gradient patterns observed on Q&A benchmarks generalize as reliable indicators of hallucination across models, tasks, and real-world use without additional fitting or selection effects.

What would settle it

Testing the gradient patterns on a new task domain or model family and finding no correlation with independently labeled hallucination rates would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.24790 by Anand Kamat, Brent M. Werness, Daniel Blake.

Figure 1
Figure 1. Figure 1: Overview of the Grad Detect pipeline. At inference time (top), per-layer gradients extracted from probe responses are compared against precomputed reference gradients via cosine similarity, producing a compact feature matrix that a lightweight transformer encoder classifies as Correct, Incorrect, or Did Not Answer. Reference gradients (bottom) are constructed offline by averaging per-category gradients ove… view at source ↗
Figure 2
Figure 2. Figure 2: LDA projections of averaged layer-wise cosine similarity vectors for Qwen2.5-3B, Falcon3-3B, Gemma3-4B, and SmolLM3-3B. Clear separation between behavioral categories confirms that gradient directions carry discriminative information. Generation and labeling. Each question is presented with the instruction “Answer the following question” us￾ing greedy decoding at temperature = 0. Responses are labeled Corr… view at source ↗
Figure 3
Figure 3. Figure 3: Full task — Correct reference gradient. LDA projections of layer-wise cosine similarities for all four combinations of reference and training response types. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Full task — Incorrect reference gradient. LDA projections of layer-wise cosine similarities for all four combinations of reference and training response types. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Full task — Did Not Answer reference gradient. LDA projections of layer-wise cosine similarities for all four combinations of reference and training response types. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Correctness task — Correct reference gradient. LDA projections of layer-wise cosine similarities. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Correctness task — Incorrect reference gradient. LDA projections of layer-wise cosine similarities. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Response task — Answered reference gradient. LDA projections of layer-wise cosine similarities. Wide cluster separation reflects the near-perfect accuracy of abstention detection. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Response task — Did Not Answer reference gradient. LDA projections of layer-wise cosine similarities. The DNA reference produces the widest separation, confirming that abstention is the most distinctive behavioral mode in gradient space. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
read the original abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, yet they remain prone to generating hallucinations. Detecting these hallucinations is critical for deploying LLMs reliably in high-stakes applications. We present Grad Detect, a gradient-based approach for predicting hallucinations by analyzing layer-wise gradient patterns from a single forward-backward pass during inference. Our method shows that the internal gradient structure of a model carries rich information about the correctness of its output. This information is not accessible through output-level signals alone. We evaluate Grad Detect on several Q&A benchmarks across both hallucination detection and model abstention prediction, where it consistently outperforms confidence-based and sampling-based baselines. Through comprehensive layer ablation studies across all eleven models from four architectural families, we find that the final five layers concentrate over 97% of the discriminative gradient signal, enabling efficient deployment with minimal performance loss. Grad Detect provides a unified framework for predicting multiple dimensions of LLM reliability, offering strong predictive performance alongside interpretable insights into where and how model failures originate.

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

3 major / 1 minor

Summary. The manuscript introduces Grad Detect, a gradient-based method for hallucination detection and abstention prediction in LLMs that extracts layer-wise gradient patterns from a single forward-backward pass at inference time. It claims this internal-gradient approach consistently outperforms confidence-based and sampling-based baselines on multiple Q&A benchmarks, and that layer-ablation experiments across eleven models from four families show the final five layers concentrate over 97% of the discriminative signal, enabling efficient deployment.

Significance. If the gradient signals prove robust and generalizable without benchmark-specific tuning, the work would supply a practical, output-independent reliability signal together with an interpretable localization of where failures originate; the reported layer concentration would further support low-overhead deployment.

major comments (3)
  1. [Abstract] Abstract: the central claim of 'consistent outperformance' over baselines is asserted without any quantitative metrics, error bars, dataset sizes, or ablation tables, so the strength of the empirical result cannot be assessed from the provided text.
  2. [Layer ablation studies] Layer ablation studies (Abstract): the statement that the final five layers concentrate over 97% of the discriminative gradient signal is obtained via post-hoc ablation performed on the identical Q&A benchmarks and eleven models used for the main evaluation; this leaves open the possibility that both the concentration and the overall discriminative power are benchmark- or model-selection artifacts rather than intrinsic properties.
  3. [Evaluation] Evaluation sections: the claim that gradient patterns supply information 'not accessible through output-level signals alone' and that the method works 'without additional fitting' requires explicit controls showing that any downstream predictor or feature selection was not tuned on the test benchmarks; absent such controls the generalization argument does not follow from the reported single-pass Q&A results.
minor comments (1)
  1. [Abstract] The abstract would be strengthened by inclusion of at least one key performance number (e.g., AUROC or accuracy delta) to allow readers to gauge the magnitude of the reported gains.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment point-by-point below, indicating planned revisions where appropriate.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of 'consistent outperformance' over baselines is asserted without any quantitative metrics, error bars, dataset sizes, or ablation tables, so the strength of the empirical result cannot be assessed from the provided text.

    Authors: We agree the abstract would benefit from quantitative context. In revision we will add key metrics (e.g., average AUC gains over baselines), dataset sizes, model count, and a brief reference to the layer-ablation result already stated in the text. revision: yes

  2. Referee: [Layer ablation studies] Layer ablation studies (Abstract): the statement that the final five layers concentrate over 97% of the discriminative gradient signal is obtained via post-hoc ablation performed on the identical Q&A benchmarks and eleven models used for the main evaluation; this leaves open the possibility that both the concentration and the overall discriminative power are benchmark- or model-selection artifacts rather than intrinsic properties.

    Authors: The ablation was run across eleven models spanning four architectural families and multiple benchmarks, with the final-five-layer concentration holding uniformly. This cross-family consistency reduces the likelihood of pure artifact, but we acknowledge the post-hoc character on the same data. We will add an explicit limitations paragraph noting this and will consider reporting results on an additional held-out benchmark in revision. revision: partial

  3. Referee: [Evaluation] Evaluation sections: the claim that gradient patterns supply information 'not accessible through output-level signals alone' and that the method works 'without additional fitting' requires explicit controls showing that any downstream predictor or feature selection was not tuned on the test benchmarks; absent such controls the generalization argument does not follow from the reported single-pass Q&A results.

    Authors: Grad Detect computes layer-wise gradient statistics directly from one forward-backward pass and applies them without training, fine-tuning, or feature selection on any evaluation benchmark. No downstream predictor is learned. We will strengthen the manuscript by adding a dedicated paragraph clarifying the absence of fitting and by including an explicit statement that all reported numbers use the identical, untuned procedure across benchmarks. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical method evaluated on external benchmarks

full rationale

The paper presents Grad Detect as an empirical method that extracts layer-wise gradient patterns from a single forward-backward pass and evaluates them directly on Q&A benchmarks for hallucination detection and abstention prediction. No equations, derivations, fitted parameters renamed as predictions, or self-citations appear in the provided text. Ablation results (e.g., final-five-layer concentration) are reported as post-hoc observations on the same benchmarks rather than as load-bearing reductions that define the central claim by construction. The approach is therefore self-contained against external data and does not reduce to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no free parameters, axioms, or invented entities are described or can be identified.

pith-pipeline@v0.9.1-grok · 5704 in / 992 out tokens · 16053 ms · 2026-06-25T23:57:42.368881+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

81 extracted references · 23 canonical work pages · 13 internal anchors

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    Language Models are Few-Shot Learners , author=. Advances in Neural Information Processing Systems , volume=. 2020 , url=

    2020

  2. [2]

    Journal of Machine Learning Research , volume=

    PaLM: Scaling Language Modeling with Pathways , author=. Journal of Machine Learning Research , volume=. 2023 , url=

    2023

  3. [3]

    arXiv preprint arXiv:2302.13971 , year=

    LLaMA: Open and Efficient Foundation Language Models , author=. arXiv preprint arXiv:2302.13971 , year=

    Pith/arXiv arXiv

  4. [4]

    ACM Computing Surveys , volume=

    Survey of Hallucination in Natural Language Generation , author=. ACM Computing Surveys , volume=. 2023 , url=

    2023

  5. [5]

    Computational Linguistics , volume=

    Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models , author=. Computational Linguistics , volume=. 2025 , url=

    2025

  6. [6]

    arXiv preprint arXiv:2207.05221 , year=

    Language Models (Mostly) Know What They Know , author=. arXiv preprint arXiv:2207.05221 , year=

    Pith/arXiv arXiv

  7. [7]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=. 2023 , url=

    2023

  8. [8]

    International Conference on Learning Representations , year=

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

  9. [9]

    International Conference on Learning Representations , year=

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author=. International Conference on Learning Representations , year=

  10. [10]

    Transactions on Machine Learning Research , year=

    Generating with Confidence: Uncertainty Quantification for Black-Box Large Language Models , author=. Transactions on Machine Learning Research , year=

  11. [11]

    arXiv preprint arXiv:2302.12813 , year=

    Check Your Facts and Try Again: Improving Large Language Models with External Knowledge and Automated Feedback , author=. arXiv preprint arXiv:2302.12813 , year=

    Pith/arXiv arXiv

  12. [12]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    RARR: Researching and Revising What Language Models Say, Using Language Models , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2023 , url=

    2023

  13. [13]

    Advances in Neural Information Processing Systems , volume=

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks , author=. Advances in Neural Information Processing Systems , volume=. 2020 , url=

    2020

  14. [14]

    arXiv preprint arXiv:1312.6034 , year=

    Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps , author=. arXiv preprint arXiv:1312.6034 , year=

    Pith/arXiv arXiv

  15. [15]

    Proceedings of the 34th International Conference on Machine Learning , series=

    Axiomatic Attribution for Deep Networks , author=. Proceedings of the 34th International Conference on Machine Learning , series=. 2017 , url=

    2017

  16. [16]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Explaining How Transformers Use Context to Build Predictions , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2023 , doi=

    2023

  17. [17]

    International Conference on Learning Representations , year=

    Explaining and Harnessing Adversarial Examples , author=. International Conference on Learning Representations , year=

  18. [18]

    International Conference on Learning Representations , year=

    Towards Deep Learning Models Resistant to Adversarial Attacks , author=. International Conference on Learning Representations , year=

  19. [19]

    Proceedings of the 34th International Conference on Machine Learning , series=

    The Shattered Gradients Problem: If ResNets are the Answer, Then What is the Question? , author=. Proceedings of the 34th International Conference on Machine Learning , series=. 2017 , url=

    2017

  20. [20]

    Advances in Neural Information Processing Systems , volume=

    How Does Batch Normalization Help Optimization? , author=. Advances in Neural Information Processing Systems , volume=. 2018 , url=

    2018

  21. [21]

    Advances in Neural Information Processing Systems , volume=

    SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability , author=. Advances in Neural Information Processing Systems , volume=. 2017 , url=

    2017

  22. [22]

    IEEE Transactions on Information Theory , volume=

    On Optimum Recognition Error and Reject Tradeoff , author=. IEEE Transactions on Information Theory , volume=. 1970 , url=

    1970

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Selective Classification for Deep Neural Networks , author=. Advances in Neural Information Processing Systems , volume=. 2017 , url=

    2017

  24. [24]

    Transactions on Machine Learning Research , year=

    Teaching Models to Express Their Uncertainty in Words , author=. Transactions on Machine Learning Research , year=

  25. [25]

    Transformer Circuits Thread , year=

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

  26. [26]

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

    Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , year=

    2022

  27. [27]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

    BERT Rediscovers the Classical NLP Pipeline , author=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=. 2019 , doi=

    2019

  28. [28]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

    What Does BERT Learn about the Structure of Language? , author=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=. 2019 , url=

    2019

  29. [29]

    Advances in Neural Information Processing Systems , volume=

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena , author=. Advances in Neural Information Processing Systems , volume=. 2023 , url=

    2023

  30. [30]

    Advances in Neural Information Processing Systems , volume=

    AlpacaFarm: A Simulation Framework for Methods that Learn from Human Feedback , author=. Advances in Neural Information Processing Systems , volume=. 2023 , url=

    2023

  31. [31]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=. 2023 , url=

    2023

  32. [32]

    Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension , author=. Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2017 , url=

    2017

  33. [33]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    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 (Volume 1: Long Papers) , pages=. 2023 , url=

    2023

  34. [34]

    Proceedings of the 3rd Workshop on Noisy User-generated Text , pages=

    Crowdsourcing Multiple Choice Science Questions , author=. Proceedings of the 3rd Workshop on Noisy User-generated Text , pages=. 2017 , url=

    2017

  35. [35]

    Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=. 2019 , url=

    2019

  36. [36]

    Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , pages=

    WikiQA: A Challenge Dataset for Open-Domain Question Answering , author=. Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , pages=. 2015 , url=

    2015

  37. [37]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    TruthfulQA: Measuring How Models Mimic Human Falsehoods , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2022 , url=

    2022

  38. [38]

    Proceedings of the IEEE International Conference on Computer Vision , pages=

    Focal Loss for Dense Object Detection , author=. Proceedings of the IEEE International Conference on Computer Vision , pages=. 2017 , url=

    2017

  39. [39]

    International Conference on Learning Representations , year=

    Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=

  40. [40]

    Proceedings of the 34th International Conference on Machine Learning , series=

    On Calibration of Modern Neural Networks , author=. Proceedings of the 34th International Conference on Machine Learning , series=. 2017 , url=

    2017

  41. [41]

    Advances in Neural Information Processing Systems , volume=

    FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness , author=. Advances in Neural Information Processing Systems , volume=. 2022 , url=

    2022

  42. [42]

    ArXiv , year =

    A Survey of Large Language Models , author =. ArXiv , year =

  43. [43]

    Why Language Models Hallucinate

    Why Language Models Hallucinate , author =. arXiv preprint , year =. doi:10.48550/arXiv.2509.04664 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2509.04664

  44. [44]

    Ethical and social risks of harm from Language Models

    Ethical and social risks of harm from Language Models , author =. arXiv preprint , year =. doi:10.48550/arXiv.2112.04359 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2112.04359

  45. [45]

    ACM Trans

    A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions , author =. ACM Transactions on Information Systems , year =. doi:10.1145/3703155 , url =

    work page doi:10.1145/3703155

  46. [46]

    arXiv preprint , year =

    Transferable and Efficient Non-Factual Content Detection via Probe Training with Offline Consistency Checking , author =. arXiv preprint , year =. doi:10.48550/arXiv.2404.06742 , url =

    work page doi:10.48550/arxiv.2404.06742

  47. [47]

    NeurIPS 2024 , year =

    HaloScope: Harnessing Unlabeled LLM Generations for Hallucination Detection , author =. NeurIPS 2024 , year =. doi:10.48550/arXiv.2409.17504 , url =

    work page doi:10.48550/arxiv.2409.17504 2024

  48. [48]

    arXiv preprint , year =

    Unsupervised Real-Time Hallucination Detection based on the Internal States of Large Language Models , author =. arXiv preprint , year =. doi:10.48550/arXiv.2403.06448 , url =

    work page doi:10.48550/arxiv.2403.06448

  49. [49]

    ICAI'24 - The 26th Int'l Conf on Artificial Intelligence , year =

    Detecting Hallucinations in Large Language Model Generation: A Token Probability Approach , author =. ICAI'24 - The 26th Int'l Conf on Artificial Intelligence , year =. doi:10.48550/arXiv.2405.19648 , url =

    work page doi:10.48550/arxiv.2405.19648

  50. [50]

    Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs

    Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs , author =. arXiv preprint , year =. doi:10.48550/arXiv.2406.15927 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.15927

  51. [51]

    EMNLP 2023 Findings , year =

    Large Language Models are Better Reasoners with Self-Verification , author =. EMNLP 2023 Findings , year =. doi:10.48550/arXiv.2212.09561 , url =

    work page doi:10.48550/arxiv.2212.09561 2023

  52. [52]

    arXiv preprint , year =

    RCOT: Detecting and Rectifying Factual Inconsistency in Reasoning by Reversing Chain-of-Thought , author =. arXiv preprint , year =. doi:10.48550/arXiv.2305.11499 , url =

    work page doi:10.48550/arxiv.2305.11499

  53. [53]

    NAACL 2024 Main Conference , year =

    A Closer Look at the Self-Verification Abilities of Large Language Models in Logical Reasoning , author =. NAACL 2024 Main Conference , year =. doi:10.48550/arXiv.2311.07954 , url =

    work page doi:10.48550/arxiv.2311.07954 2024

  54. [54]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. arXiv preprint , year =. doi:10.48550/arXiv.2201.11903 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2201.11903

  55. [55]

    Attention Is All You Need

    Attention Is All You Need , author =. arXiv preprint , year =. doi:10.48550/arXiv.1706.03762 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1706.03762

  56. [56]

    ICCV 2021 , year =

    CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classification , author =. ICCV 2021 , year =. doi:10.48550/arXiv.2103.14899 , url =

    work page doi:10.48550/arxiv.2103.14899 2021

  57. [57]

    Focal Loss for Dense Object Detection

    Focal Loss for Dense Object Detection , author =. arXiv preprint , year =. doi:10.48550/arXiv.1708.02002 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1708.02002

  58. [58]

    The Internal State of an LLM Knows When It's Lying

    The Internal State of an LLM Knows When It's Lying , author =. arXiv preprint , year =. doi:10.48550/arXiv.2304.13734 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2304.13734

  59. [59]

    SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models

    SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models , author =. EMNLP 2023 , year =. doi:10.48550/arXiv.2303.08896 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2303.08896 2023

  60. [60]

    arXiv preprint , year =

    V-STaR: Training Verifiers for Self-Taught Reasoners , author =. arXiv preprint , year =. doi:10.48550/arXiv.2402.06457 , url =

    work page doi:10.48550/arxiv.2402.06457

  61. [61]

    Llama 2: Open Foundation and Fine-Tuned Chat Models

    Llama 2: Open Foundation and Fine-Tuned Chat Models , author =. arXiv preprint , year =. doi:10.48550/arXiv.2307.09288 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2307.09288

  62. [62]

    Qwen2.5 Technical Report

    Qwen2.5 Technical Report , author =. arXiv preprint , year =. doi:10.48550/arXiv.2412.15115 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115

  63. [63]

    TruthfulQA: Measuring How Models Mimic Human Falsehoods

    TruthfulQA: Measuring How Models Mimic Human Falsehoods , author =. ACL 2022 , year =. doi:10.48550/arXiv.2109.07958 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2109.07958 2022

  64. [64]

    Training Verifiers to Solve Math Word Problems

    Training Verifiers to Solve Math Word Problems , author =. arXiv preprint , year =. doi:10.48550/arXiv.2110.14168 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2110.14168

  65. [65]

    TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension

    TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension , author =. arXiv preprint , year =. doi:10.48550/arXiv.1705.03551 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1705.03551

  66. [66]

    arXiv preprint arXiv:2407.03282 , year=

    LLM Internal States Reveal Hallucination Risk Faced With a Query , author=. arXiv preprint arXiv:2407.03282 , year=

    arXiv

  67. [67]

    2021 , eprint=

    Training Verifiers to Solve Math Word Problems , author=. 2021 , eprint=

    2021

  68. [68]

    2023 , eprint=

    When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories , author=. 2023 , eprint=

    2023

  69. [69]

    2025 , eprint=

    Mixture of Experts in Large Language Models , author=. 2025 , eprint=

    2025

  70. [70]

    Advances in Neural Information Processing Systems , volume =

    Teaching Machines to Read and Comprehend , author =. Advances in Neural Information Processing Systems , volume =

  71. [71]

    and Lapata, Mirella , booktitle =

    Narayan, Shashi and Cohen, Shay B. and Lapata, Mirella , booktitle =. Don't Give Me the Details, Just the Summary!. 2018 , publisher =

    2018

  72. [72]

    2019 , publisher =

    Gliwa, Bogdan and Mochol, Iwona and Biesek, Maciej and Wawer, Aleksander , booktitle =. 2019 , publisher =

    2019

  73. [73]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages =

    Multi-News: A Large-Scale Multi-Document Summarization Dataset and Abstractive Hierarchical Model , author =. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages =. 2019 , publisher =

    2019

  74. [74]

    Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =

    A Discourse-Aware Attention Model for Abstractive Summarization of Long Documents , author =. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =. 2018 , publisher =

    2018

  75. [75]

    2308.03688 , archivePrefix =

    Liu, Xiao and Yu, Hao and Zhang, Hanchen and Xu, Yifan and Lei, Xuanyu and Lai, Hanyu and Gu, Yu and Ding, Hangliang and Men, Kaiwen and Yang, Kejuan and Zhang, Shudan and Deng, Xiang and Zeng, Aohan and Du, Zhengxiao and Zhang, Chenhui and Shen, Sheng and Zhang, Tianjun and Su, Yu and Tang, Jie and Yao, Juntao and Liu, Yankai and Li, Ruobing and Sun, Zhi...

    Pith/arXiv arXiv

  76. [76]

    2024 , url=

    Carlos E Jimenez and John Yang and Alexander Wettig and Shunyu Yao and Kexin Pei and Ofir Press and Karthik R Narasimhan , booktitle=. 2024 , url=

    2024

  77. [77]

    2023 , eprint =

    GAIA: a benchmark for General AI Assistants , author =. 2023 , eprint =

    2023

  78. [78]

    2024 , eprint=

    -bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains , author=. 2024 , eprint=

    2024

  79. [79]

    The Falcon 3 Family of Open Models , author =

  80. [80]

    arXiv preprint arXiv:2503.19786 , year =

    Gemma 3 Technical Report , author =. arXiv preprint arXiv:2503.19786 , year =

    Pith/arXiv arXiv

Showing first 80 references.