Pith. sign in

REVIEW 3 major objections 2 minor 79 references

ExploreGS: Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors

T0 review · 3 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read ExploreGS claims that adding diffusion-refined pseudo-views from information-gain-placed virtual cameras to a 3D Gaussian Splatting scene makes arbitrary-view rendering artifact-free and beats existing 3DGS methods on its Wild-Explore bench

desk verdict The submission artifact is unverifiable: the claimed paper's full text is actually a different LLM-probing paper, so the abstract is the only evidence available. read the letter →

arxiv 2508.06014 v1 pith:PO4COKXP submitted 2025-08-08 cs.CV

Minsu Kim , Subin Jeon , In Cho , Mijin Yoo , Seon Joo Kim This is my paper
classification cs.CV
keywords 3DGaussianSplattingnovelviewsynthesissceneexplorationvirtualcameraplacementinformationgainvideodiffusionpriorsarbitrary-viewrenderingWild-Explorebenchmark
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

3D Gaussian Splatting renders captured scenes in real time, but quality collapses when the requested viewpoint leaves the original camera path, producing artifacts and holes that block seamless exploration. ExploreGS argues that the fix is not a better renderer but a better training set: place virtual cameras where the reconstruction knows least, render what they would see, refine those renders with a video diffusion model that enforces temporal consistency, and fine-tune the Gaussians on the result. The paper reports that this loop significantly improves reconstruction quality, and it introduces Wild-Explore, a benchmark in which evaluation viewpoints deliberately depart from the training trajectory, to demonstrate the gain. If the claim holds, real-time 3DGS renderers become explorer-friendly without changing the splatting machinery underneath.

What carries the argument

The load-bearing combination is a virtual camera placement policy driven by information gain — the expected reduction in uncertainty about unobserved scene content — which concentrates new training views where the reconstruction is least confident — together with a video diffusion prior used as a refiner, which turns the splatted pseudo-views into temporally coherent images before they are fed back as training data. The third component is the fine-tuning loop on the 3D Gaussians themselves: the pseudo-views act as additional supervision, so the splat parameters absorb newly covered regions instead of only the original images. The information-gain term decides where to look; the diffusion pri

What would settle it

One decisive check: on a scene with a mirror or repeated texture, hold out all views of that region, run ExploreGS, and measure PSNR/SSIM on the held-out region against ground truth. If the diffusion-refined pseudo-views are wrong where geometry is ambiguous but metrics still rise, the prior is painting plausible noise; if metrics fall relative to the non-augmented baseline, the claim that video diffusion priors reliably support viewpoint extrapolation fails on exactly the hard cases the benchmark is meant to test.

Watch

Extended reading notes

Core claim

The central claim is that the failure of 3D Gaussian Splatting under viewpoint extrapolation is a coverage problem, not a rendering problem, and can be fixed by augmenting the Gaussian scene with training views it never had. The pipeline has three stages. First, an information-gain-driven virtual camera sampler proposes viewpoints that maximize new scene coverage, explicitly targeting under-reconstructed regions rather than adding redundant views. Second, a video diffusion prior, chosen for temporal consistency across frames, refines the coarse splat-rendered images from those cameras into clean pseudo-views. Third, the 3D Gaussians are fine-tuned on these enhanced views so that holes and ar

Load-bearing premise

The video diffusion prior must refine pseudo-views in a way that agrees with the actual scene geometry; if it fabricates plausible content that the fine-tuning then bakes into the Gaussians, real novel-view quality can degrade instead of improve.

Editorial extensions

If this is right

  • Arbitrary-view rendering from a 3DGS scene becomes artifact-free enough for seamless exploration, not just interpolation near the training views.
  • The video diffusion prior acts as a training-signal generator: its refined pseudo-views, not just real images, become supervision that fills in under-reconstructed regions.
  • The Wild-Explore benchmark gives the community a common yardstick for extrapolation quality — viewpoints that depart from the training trajectory rather than near duplicates of it.
  • According to the reported experiments, the augmented pipeline outperforms existing 3DGS-based methods on this benchmark while keeping the real-time rendering property of 3DGS.

Reading between the lines

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

  • Editorial extension: the information-gain camera placement is the same optimization as next-best-view selection in active perception; the criterion could transfer to robots or drones choosing where to look next, with the diffusion prior standing in for 'what might be behind the corner.'
  • Editorial extension: a stress test on scenes with mirrors, glass, or repeated texture would map the failure boundary — these are exactly the cases where a diffusion refiner is most likely to produce plausible content that disagrees with true geometry.
  • Editorial extension: the paper fixes viewpoint coverage at training time; a natural iterative variant is an explore-then-tune loop that alternates camera sampling and fine-tuning until the information gain saturates, which the paper does not analyze.
  • Editorial extension: because the pseudo-views come from a generative prior, the reported metric gains may partly reflect the prior's texture statistics rather than recovered structure; comparing against a depth-based pseudo-view baseline would separate the two effects.
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

3 major / 2 minor

Summary. The abstract describes ExploreGS, a 3D Gaussian Splatting pipeline for novel view synthesis that (i) selects virtual camera poses via an information-gain criterion to increase scene coverage, (ii) refines rendered images with video diffusion priors, and (iii) fine-tunes the 3D Gaussians on these enhanced views. The abstract further claims that this improves reconstruction quality over existing 3DGS-based methods and introduces a new benchmark, Wild-Explore, for challenging scene exploration. However, the full text supplied with the manuscript is an unrelated paper, 'Efficient Knowledge Probing of Large Language Models by Adapting Pre-trained Embeddings' (PEEK), which concerns estimating LLM factual knowledge with proxy embeddings. None of the ExploreGS components—3D Gaussian Splatting, virtual camera placement, information gain, video diffusion, or Wild-Explore—appear anywhere in the full text. The central claims of the abstract are therefore unsupported by the document under review.

Significance. If the ExploreGS method were actually described and evaluated as the abstract suggests, the work could be relevant to the 3DGS/NVS community: the idea of actively selecting virtual views and using diffusion priors to reduce artifacts caused by large viewpoint changes is plausible and worth testing. The proposed benchmark, Wild-Explore, could also be a useful resource if properly designed. However, because the submitted text contains no method description, no experimental protocol, and no results, the significance cannot be assessed from this manuscript. The present document provides no evidence for any of its stated contributions.

major comments (3)
  1. [Entire full text (Abstract vs. Full text)] The supplied full text does not match the abstract. It is a different paper about LLM knowledge probing (PEEK) and contains no description of 3D Gaussian Splatting, virtual camera sampling, information-gain-driven placement, video diffusion priors, or the Wild-Explore benchmark. Consequently, every claim in the abstract—including the central empirical claim that the method 'outperforms existing 3DGS-based methods'—is unsupported by the document under review. This is a verifiability failure, not a minor editing issue.
  2. [Experimental evaluation (absent)] The full text contains no experimental setup, datasets, baselines, metrics, quantitative results, ablations, or error bars. There is no way to check the asserted 'significant improvement' in reconstruction quality. The absence of ablations also makes it impossible to separate the contribution of information-gain camera placement from that of diffusion-based refinement, which is exactly the kind of attribution the abstract claims.
  3. [Wild-Explore benchmark (absent)] The abstract introduces Wild-Explore as a new evaluation benchmark, but the full text contains no description of its scenes, capture protocol, train/test viewpoint splits, or evaluation metrics. Without these details, the benchmark cannot be used, compared, or reproduced. This is a load-bearing omission because the paper's main empirical claim depends on this benchmark.
minor comments (2)
  1. [Metadata consistency] The full-text header and page footers refer to arXiv:2508.06030v2 and the PEEK paper, while the claimed manuscript is arXiv:2508.06014. The metadata should be reconciled in any revised version.
  2. [Abstract wording] Even at the abstract level, 'significantly improves reconstruction quality' is too vague to be falsifiable. Quantitative comparisons or at least a pointer to specific tables would be needed for a verifiable claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found in the supplied text; however, the full text is an unrelated paper, so the ExploreGS claims cannot be verified from the manuscript.

full rationale

The abstract claims a 3D Gaussian Splatting pipeline with information-gain-driven virtual camera placement and video diffusion priors, but the supplied full text is 'Efficient Knowledge Probing of Large Language Models by Adapting Pre-trained Embeddings' (PEEK), a paper about proxy embeddings for estimating LLM knowledge. No derivation chain, equations, or experimental setup for ExploreGS appears in the manuscript. There is therefore no circular step to exhibit: the central claim is unsupported by the provided document, which is a verifiability/correctness failure rather than a circularity. For the PEEK content that is present, the method fits a linear head on labeled LLM facts and evaluates on a held-out set; that is standard supervised evaluation, not a fitted-input-called-prediction circularity. No load-bearing self-citation chain or redefinition of terms is present. The reader's concern about evaluation viewpoints overlapping virtual training viewpoints cannot be assessed because the ExploreGS methodology is absent; it remains an empirical risk, not a demonstrated circular reduction.

Assumptions & free parameters 2 free parameters · 2 assumptions · 1 invented entities

Abstract-only review; the supplied full text is an unrelated paper, so the ledger is reconstructed from the abstract and marked unverifiable.

free parameters (2)
  • Information-gain camera placement hyperparameters (number of virtual views, gain threshold)
    Abstract describes an information-gain strategy but gives no values; these would be tuned and are unverifiable from the abstract alone.
  • Diffusion prior guidance/refinement strength
    Abstract does not specify; the degree of refinement would affect how much diffusion content is mixed into the training views.
assumptions (2)
  • domain assumption Video diffusion priors can refine rough renders into plausible, consistent frames suitable for training 3D Gaussians.
    Core to the method; if refinement is implausible, fine-tuning inherits errors. Invoked by the pipeline description.
  • domain assumption Information-gain-driven virtual camera placement maximizes scene coverage and reduces rendering artifacts.
    Design choice claimed in the abstract; not proven.
invented entities (1)
  • Wild-Explore benchmark
    purpose: Standardized evaluation for challenging scene exploration viewpoints.
    Abstract announces the benchmark, but no dataset documentation, statistics, or download link is given in the abstract; cannot verify independent existence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ExploreGS: Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors." pith.science (2026). https://pith.science/paper/PO4COKXP

@misc{pith2026250806014,
  author       = {Pith},
  title        = {Pith review of: ExploreGS: Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PO4COKXP}},
  note         = {Machine review of arXiv:2508.06014}
}
read the original abstract

Recent advances in novel view synthesis (NVS) have enabled real-time rendering with 3D Gaussian Splatting (3DGS). However, existing methods struggle with artifacts and missing regions when rendering from viewpoints that deviate from the training trajectory, limiting seamless scene exploration. To address this, we propose a 3DGS-based pipeline that generates additional training views to enhance reconstruction. We introduce an information-gain-driven virtual camera placement strategy to maximize scene coverage, followed by video diffusion priors to refine rendered results. Fine-tuning 3D Gaussians with these enhanced views significantly improves reconstruction quality. To evaluate our method, we present Wild-Explore, a benchmark designed for challenging scene exploration. Experiments demonstrate that our approach outperforms existing 3DGS-based methods, enabling high-quality, artifact-free rendering from arbitrary viewpoints. https://exploregs.github.io

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

79 extracted references · 27 canonical work pages

  1. [1]

    Dbpedia–a large-scale, multilingual knowledge base extracted from wikipedia.Semantic Web, 6(2):167–195, 2015

    Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, Dimitris Kontokostas, Pablo N Mendes, Sebastian Hellmann, Mohamed Morsey, Patrick van Kleef, Sören Auer, et al. Dbpedia–a large-scale, multilingual knowledge base extracted from wikipedia.Semantic Web, 6(2):167–195, 2015

    work page 2015

  2. [2]

    Yago: A core of semantic knowledge

    Fabian M Suchanek, Gjergji Kasneci, and Gerhard Weikum. Yago: A core of semantic knowledge. InProceedings of the 16th international conference on World Wide Web, pages 697–706, 2007

    work page 2007

  3. [3]

    Language models as knowledge bases?arXiv:1909.01066, 2019

    Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases?arXiv:1909.01066, 2019

    arXiv 1909

  4. [4]

    Large language models encode clinical knowledge.Nature, 620(7972):172–180, 2023

    Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al. Large language models encode clinical knowledge.Nature, 620(7972):172–180, 2023

    2023

  5. [5]

    Legalbench: A collaboratively built benchmark for measuring legal reasoning in large language models.Advances in Neural Information Processing Systems, 36:44123–44279, 2023

    Neel Guha, Julian Nyarko, Daniel Ho, Christopher Ré, Adam Chilton, Alex Chohlas-Wood, Austin Peters, Brandon Waldon, Daniel Rockmore, Diego Zambrano, et al. Legalbench: A collaboratively built benchmark for measuring legal reasoning in large language models.Advances in Neural Information Processing Systems, 36:44123–44279, 2023

    work page 2023

  6. [6]

    Llm-sr: Scientific equation discovery via programming with large language models

    Parshin Shojaee, Kazem Meidani, Shashank Gupta, Amir Barati Farimani, and Chandan K Reddy. Llm-sr: Scientific equation discovery via programming with large language models. arXiv:2404.18400, 2024

    arXiv 2024

  7. [7]

    Llm and simulation as bilevel optimizers: a new paradigm to advance physical scientific discovery

    Pingchuan Ma, Tsun-Hsuan Wang, Minghao Guo, Zhiqing Sun, Joshua B Tenenbaum, Daniela Rus, Chuang Gan, and Wojciech Matusik. Llm and simulation as bilevel optimizers: a new paradigm to advance physical scientific discovery. InProceedings of the 41st International Conference on Machine Learning, pages 33940–33962, 2024

    work page 2024

  8. [8]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv:2311.05232, 2023

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv:2311.05232, 2023

    arXiv 2023

Show all 79 references
  1. [9]

    Model evaluation for extreme risks.arXiv:2305.15324, 2023

    Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, et al. Model evaluation for extreme risks.arXiv:2305.15324, 2023

    2023 arXiv

  2. [10]

    A hazard analysis framework for code synthesis large language models.arXiv:2207.14157, 2022

    Heidy Khlaaf, Pamela Mishkin, Joshua Achiam, Gretchen Krueger, and Miles Brundage. A hazard analysis framework for code synthesis large language models.arXiv:2207.14157, 2022

    2022 arXiv

  3. [11]

    Kglens: Towards efficient and effective knowledge probing of large language models with knowledge graphs

    Shangshang Zheng, He Bai, Yizhe Zhang, Yi Su, Xiaochuan Niu, and Navdeep Jaitly. Kglens: Towards efficient and effective knowledge probing of large language models with knowledge graphs. arXiv:2312.11539, 2023

    2023 arXiv

  4. [12]

    Large language models struggle to learn long-tail knowledge

    Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. Large language models struggle to learn long-tail knowledge. InInternational Conference on Machine Learning, pages 15696–15707. PMLR, 2023

    2023

  5. [13]

    Head-to-tail: How knowledge- able are large language models (llm)? aka will llms replace knowledge graphs?arXiv:2308.10168, 2023

    Kai Sun, Yifan Ethan Xu, Hanwen Zha, Yue Liu, and Xin Luna Dong. Head-to-tail: How knowledge- able are large language models (llm)? aka will llms replace knowledge graphs?arXiv:2308.10168, 2023

    2023 arXiv

  6. [14]

    Systematic assessment of factual knowledge in large language models.arXiv:2310.11638, 2023

    Linhao Luo, Thuy-Trang Vu, Dinh Phung, and Gholamreza Haffari. Systematic assessment of factual knowledge in large language models.arXiv:2310.11638, 2023. 11

    2023 arXiv

  7. [15]

    How can we know what language models know?Transactions of the Association for Computational Linguistics, 8:423–438, 2020

    Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know?Transactions of the Association for Computational Linguistics, 8:423–438, 2020

    2020

  8. [16]

    The internal state of an llm knows when it’s lying

    Amos Azaria and Tom Mitchell. The internal state of an llm knows when it’s lying. arXiv:2304.13734, 2023

    2023 arXiv

  9. [17]

    Estimating knowledge in large language models without generating a single token.arXiv:2406.12673, 2024

    Daniela Gottesman and Mor Geva. Estimating knowledge in large language models without generating a single token.arXiv:2406.12673, 2024

    2024 arXiv

  10. [18]

    Faclens: Transfer- able probe for foreseeing non-factuality in large language models.arXiv:2406.05328, 2024

    Yanling Wang, Haoyang Li, Hao Zou, Jing Zhang, Xinlei He, Qi Li, and Ke Xu. Faclens: Transfer- able probe for foreseeing non-factuality in large language models.arXiv:2406.05328, 2024

    2024 arXiv

  11. [19]

    Distinguishing the knowable from the unknowable with language models.arXiv:2402.03563, 2024

    Gustaf Ahdritz, Tian Qin, Nikhil Vyas, Boaz Barak, and Benjamin L Edelman. Distinguishing the knowable from the unknowable with language models.arXiv:2402.03563, 2024

    2024 arXiv

  12. [20]

    Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models.arXiv:2303.08896, 2023

    Potsawee Manakul, Adian Liusie, and Mark JF Gales. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models.arXiv:2303.08896, 2023

    2023 arXiv

  13. [21]

    Estimating the hallucination rate of generative ai.Advances in Neural Information Processing Systems, 37:31154–31201, 2025

    Andrew Jesson, Nicolas Beltran Velez, Quentin Chu, Sweta Karlekar, Jannik Kossen, Yarin Gal, John P Cunningham, and David Blei. Estimating the hallucination rate of generative ai.Advances in Neural Information Processing Systems, 37:31154–31201, 2025

    2025

  14. [22]

    Mteb: Massive text embed- ding benchmark.arXiv:2210.07316, 2022

    Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. Mteb: Massive text embed- ding benchmark.arXiv:2210.07316, 2022

    2022 arXiv

  15. [23]

    Towards foundation models for knowledge graph reasoning.2310.04562, 2023

    Mikhail Galkin, Xinyu Yuan, Hesham Mostafa, Jian Tang, and Zhaocheng Zhu. Towards foundation models for knowledge graph reasoning.2310.04562, 2023

    2023 arXiv

  16. [24]

    Llm2vec: Large language models are secretly powerful text encoders

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Cha- pados, and Siva Reddy. Llm2vec: Large language models are secretly powerful text encoders. arXiv:2404.05961, 2024

    2024 arXiv

  17. [25]

    Llms as zero-shot graph learners: Alignment of gnn representations with llm token embeddings.Advances in Neural Information Processing Systems, 37:5950–5973, 2024

    Duo Wang, Yuan Zuo, Fengzhi Li, and Junjie Wu. Llms as zero-shot graph learners: Alignment of gnn representations with llm token embeddings.Advances in Neural Information Processing Systems, 37:5950–5973, 2024

    2024

  18. [26]

    Give me the facts! a survey on factual knowledge probing in pre-trained language models.arXiv:2310.16570, 2023

    Paul Youssef, Osman Alperen Kora¸ s, Meijie Li, Jörg Schlötterer, and Christin Seifert. Give me the facts! a survey on factual knowledge probing in pre-trained language models.arXiv:2310.16570, 2023

    2023 arXiv

  19. [27]

    Factscore: Fine-grained atomic evaluation of factual precision in long form text generation.arXiv:2305.14251, 2023

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. Factscore: Fine-grained atomic evaluation of factual precision in long form text generation.arXiv:2305.14251, 2023

    2023 arXiv

  20. [28]

    Kgquiz: Evaluating the generalization of encoded knowledge in large language models

    Yuyang Bai, Shangbin Feng, Vidhisha Balachandran, Zhaoxuan Tan, Shiqi Lou, Tianxing He, and Yulia Tsvetkov. Kgquiz: Evaluating the generalization of encoded knowledge in large language models. InProceedings of the ACM Web Conference 2024, pages 2226–2237, 2024

    2024

  21. [29]

    Inference-time intervention: Eliciting truthful answers from a language model.Advances in Neural Information Processing Systems, 36:41451–41530, 2023

    Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model.Advances in Neural Information Processing Systems, 36:41451–41530, 2023

    2023

  22. [30]

    Cognitive dissonance: Why do language model outputs disagree with internal representations of truthfulness?arXiv:2312.03729, 2023

    Kevin Liu, Stephen Casper, Dylan Hadfield-Menell, and Jacob Andreas. Cognitive dissonance: Why do language model outputs disagree with internal representations of truthfulness?arXiv:2312.03729, 2023

    2023 arXiv

  23. [31]

    Hallucination detection in llms: Fast and memory-efficient finetuned models.arXiv:2409.02976, 2024

    Gabriel Y Arteaga, Thomas B Schön, and Nicolas Pielawski. Hallucination detection in llms: Fast and memory-efficient finetuned models.arXiv:2409.02976, 2024. 12

    2024 arXiv

  24. [32]

    Hallucination detection: Robustly discerning reliable answers in large language models

    Yuyan Chen, Qiang Fu, Yichen Yuan, Zhihao Wen, Ge Fan, Dayiheng Liu, Dongmei Zhang, Zhixu Li, and Yanghua Xiao. Hallucination detection: Robustly discerning reliable answers in large language models. InProceedings of the 32nd ACM International Conference on Information and Kno...

    2023

  25. [33]

    Teaching models to express their uncertainty in words.arXiv:2205.14334, 2022

    Stephanie Lin, Jacob Hilton, and Owain Evans. Teaching models to express their uncertainty in words.arXiv:2205.14334, 2022

    2022 arXiv

  26. [34]

    Boolq: Exploring the surprising difficulty of natural yes/no questions.arXiv:1905.10044, 2019

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions.arXiv:1905.10044, 2019

    1905 arXiv

  27. [35]

    Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models.arXiv:2104.08663, 2021

    Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models.arXiv:2104.08663, 2021

    2021 arXiv

  28. [36]

    V-measure: A conditional entropy-based external cluster evaluation measure

    Andrew Rosenberg and Julia Hirschberg. V-measure: A conditional entropy-based external cluster evaluation measure. InProceedings of the 2007 joint conference on empirical methods in natural language processing and computational natural language learning (EMNLP-CoNLL), pages 41...

    2007

  29. [37]

    Understanding the reasoning ability of language models from the perspective of reasoning paths aggregation.arXiv:2402.03268, 2024

    Xinyi Wang, Alfonso Amayuelas, Kexun Zhang, Liangming Pan, Wenhu Chen, and William Yang Wang. Understanding the reasoning ability of language models from the perspective of reasoning paths aggregation.arXiv:2402.03268, 2024

    2024 arXiv

  30. [38]

    On the opportunities and risks of foundation models.arXiv:2108.07258, 2021

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models.arXiv:2108.07258, 2021

    2021 arXiv

  31. [39]

    Lora: Low-rank adaptation of large language models.arXiv:2106.09685, 2021

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.arXiv:2106.09685, 2021

    2021 arXiv

  32. [40]

    Creating training corpora for NLG micro-planning

    Claire Gardent, Anastasia Shimorina, Shashi Narayan, and Laura Perez-Beltrachini. Creating training corpora for NLG micro-planning. InProceedings of the 10th International Conference on Natural Language Generation, pages 111–120. Association for Computational Linguistics, 2017

    2017

  33. [41]

    The webnlg challenge: Generating text from rdf data

    Claire Gardent, Anastasia Shimorina, Shashi Narayan, and Laura Perez-Beltrachini. The webnlg challenge: Generating text from rdf data. InProceedings of the 10th international conference on natural language generation, pages 124–133, 2017

    2017

  34. [42]

    Mpnet: Masked and permuted pre-training for language understanding.NeurIPS, 33:16857–16867, 2020

    Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. Mpnet: Masked and permuted pre-training for language understanding.NeurIPS, 33:16857–16867, 2020

    2020

  35. [43]

    Nv-embed: Improved techniques for training llms as generalist embedding models

    Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. InICLR, 2025

    2025

  36. [44]

    Linq-embed-mistral: Elevating text retrieval with improved gpt data through task-specific control and quality refinement

    Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, Jy yong Sohn, and Chanyeol Choi. Linq-embed-mistral: Elevating text retrieval with improved gpt data through task-specific control and quality refinement. Linq AI Research Blog, 2024

    2024

  37. [45]

    Towards general text embeddings with multi-stage contrastive learning.arXiv:2308.03281, 2023

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning.arXiv:2308.03281, 2023

    2023 arXiv

  38. [46]

    Solatorio

    Aivin V . Solatorio. Gistembed: Guided in-sample selection of training negatives for text embedding fine-tuning.arXiv:2402.16829, 2024. 13

    2024 arXiv

  39. [47]

    Angle-optimized text embeddings.arXiv preprint arXiv:2309.12871, 2023

    Xianming Li and Jing Li. Angle-optimized text embeddings.arXiv preprint arXiv:2309.12871, 2023

    2023 arXiv

  40. [48]

    Mteb: Massive text embed- ding benchmark

    Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. Mteb: Massive text embed- ding benchmark. InEACL, pages 2014–2037, 2023

    2014

  41. [49]

    The llama 3 herd of models

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv:2407.21783, 2024

    2024 arXiv

  42. [50]

    Gpt-4 technical report.Arxiv Preprint, arXiv:2303.08774, 2023

    OpenAI. Gpt-4 technical report.Arxiv Preprint, arXiv:2303.08774, 2023

    2023 arXiv

  43. [51]

    Large language models are zero-shot reasoners.Advances in neural information processing systems, 35:22199–22213, 2022

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners.Advances in neural information processing systems, 35:22199–22213, 2022

    2022

  44. [52]

    Stanford alpaca: An instruction-following llama model, 2023

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model, 2023

    2023

  45. [53]

    Stable lm 2 1.6b technical report.arXiv:2402.17834, 2024

    Marco Bellagente, Jonathan Tow, Dakota Mahan, Duy Phung, Maksym Zhuravinskyi, Reshinth Adithyan, James Baicoianu, Ben Brooks, Nathan Cooper, Ashish Datta, et al. Stable lm 2 1.6b technical report.arXiv:2402.17834, 2024

    2024 arXiv

  46. [54]

    ChatGPT: Conversational ai platform

    OpenAI. ChatGPT: Conversational ai platform. Online, 2022

    2022

  47. [55]

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality.See https://vicuna

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality.See https://vicuna. lmsys. org (accessed 14 April 2023), 2(...

    2023

  48. [56]

    Training language models to follow instructions with human feedback.Advances in Neural Information Processing Systems, 35:27730– 27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in Neural Information Processing Systems, 35:27730...

    2022

  49. [57]

    Fever: a large-scale dataset for fact extraction and verification.arXiv:1803.05355, 2018

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. Fever: a large-scale dataset for fact extraction and verification.arXiv:1803.05355, 2018

    2018 arXiv

  50. [58]

    Hotpotqa: A dataset for diverse, explainable multi-hop question answering.arXiv:1809.09600, 2018

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering.arXiv:1809.09600, 2018

    2018 arXiv

  51. [59]

    Sparse autoencoder.CS294A Lecture notes, 72(2011):1–19, 2011

    Andrew Ng et al. Sparse autoencoder.CS294A Lecture notes, 72(2011):1–19, 2011

    2011

  52. [60]

    Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2.arXiv:2408.05147, 2024

    Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, János Kramár, Anca Dragan, Rohin Shah, and Neel Nanda. Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2.arXiv:2408.05147, 2024

    2024 arXiv

  53. [61]

    Explainability for large language models: A survey.ACM Transac- tions on Intelligent Systems and Technology, 15(2):1–38, 2024

    Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey.ACM Transac- tions on Intelligent Systems and Technology, 15(2):1–38, 2024

    2024

  54. [62]

    Investigating the limitations of transformers with simple arithmetic tasks.arXiv:2102.13019, 2021

    Rodrigo Nogueira, Zhiying Jiang, and Jimmy Lin. Investigating the limitations of transformers with simple arithmetic tasks.arXiv:2102.13019, 2021

    2021 arXiv

  55. [63]

    Understanding addition in transformers.arXiv:2310.13121, 2023

    Philip Quirke and Fazl Barez. Understanding addition in transformers.arXiv:2310.13121, 2023. 14

    2023 arXiv

  56. [64]

    Generalization through memorization: Nearest neighbor language models.arXiv:1911.00172, 2019

    Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models.arXiv:1911.00172, 2019

    1911 arXiv

  57. [65]

    Retrieval augmented language model pre-training

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. Retrieval augmented language model pre-training. InInternational conference on machine learning, pages 3929–3938. PMLR, 2020

    2020

  58. [66]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.NeurIPS, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.NeurIPS, 33:9459–9474, 2020

    2020

  59. [67]

    Compact: Compressing retrieved documents actively for question answering

    Chanwoong Yoon, Taewhoo Lee, Hyeon Hwang, Minbyul Jeong, and Jaewoo Kang. Compact: Compressing retrieved documents actively for question answering. InEMNLP, 2024

    2024

  60. [68]

    Lift yourself up: Retrieval-augmented text generation with self-memory.NeurIPS, 36, 2024

    Xin Cheng, Di Luo, Xiuying Chen, Lemao Liu, Dongyan Zhao, and Rui Yan. Lift yourself up: Retrieval-augmented text generation with self-memory.NeurIPS, 36, 2024

    2024

  61. [69]

    Recomp: Improving retrieval-augmented lms with compression and selective augmentation

    Fangyuan Xu, Weijia Shi, and Eunsol Choi. Recomp: Improving retrieval-augmented lms with compression and selective augmentation. InICLR, 2024

    2024

  62. [70]

    Halueval: A large-scale hallucination evaluation benchmark for large language models.arXiv:2305.11747, 2023

    Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. Halueval: A large-scale hallucination evaluation benchmark for large language models.arXiv:2305.11747, 2023

    2023 arXiv

  63. [71]

    Do llms know about hallucination? an empirical investigation of llm’s hidden states.arXiv:2402.09733, 2024

    Hanyu Duan, Yi Yang, and Kar Yan Tam. Do llms know about hallucination? an empirical investigation of llm’s hidden states.arXiv:2402.09733, 2024

    2024 arXiv

  64. [72]

    Benchmarking uncertainty quantification methods for large language models with lm-polygraph

    Roman Vashurin, Ekaterina Fadeeva, Artem Vazhentsev, Lyudmila Rvanova, Akim Tsvigun, Daniil Vasilev, Rui Xing, Abdelrahman Boda Sadallah, Kirill Grishchenkov, Sergey Petrakov, et al. Benchmarking uncertainty quantification methods for large language models with lm-polygraph. a...

    2024 arXiv

  65. [73]

    Bertnesia: Investigating the capture and forgetting of knowledge in bert.arXiv:2106.02902, 2021

    Jonas Wallat, Jaspreet Singh, and Avishek Anand. Bertnesia: Investigating the capture and forgetting of knowledge in bert.arXiv:2106.02902, 2021

    2021 arXiv

  66. [74]

    How much knowledge can you pack into the parameters of a language model?arXiv:2002.08910, 2020

    Adam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model?arXiv:2002.08910, 2020

    2002 arXiv

  67. [75]

    Locating and editing factual associations in gpt.Advances in Neural Information Processing Systems, 35:17359–17372, 2022

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt.Advances in Neural Information Processing Systems, 35:17359–17372, 2022

    2022

  68. [76]

    Challenges with unsupervised llm knowledge discovery.arXiv:2312.10029, 2023

    Sebastian Farquhar, Vikrant Varma, Zachary Kenton, Johannes Gasteiger, Vladimir Mikulik, and Rohin Shah. Challenges with unsupervised llm knowledge discovery.arXiv:2312.10029, 2023

    2023 arXiv

  69. [77]

    Fast model editing at scale.arXiv:2110.11309, 2021

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale.arXiv:2110.11309, 2021

    2021 arXiv

  70. [78]

    Propa- gating knowledge updates to lms through distillation.Advances in Neural Information Processing Systems, 36, 2024

    Shankar Padmanabhan, Yasumasa Onoe, Michael Zhang, Greg Durrett, and Eunsol Choi. Propa- gating knowledge updates to lms through distillation.Advances in Neural Information Processing Systems, 36, 2024

    2024

  71. [79]

    Rlaif: Scaling reinforcement learning from human feedback with ai feedback.arXiv:2309.00267, 2023

    Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. Rlaif: Scaling reinforcement learning from human feedback with ai feedback.arXiv:2309.00267, 2023. 15 Appendix A Extended Related Work Interpretable LLM...

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.