{ "schema_version": 2, "generated_from_author_sources": true, "conforms_to": "https://aogavrilov.com/openapi.json#/components/schemas/FullTextCitationContextIndex", "canonical_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/citations.json", "source_full_text_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/", "source_fragment_sha256": "f549e42d8ec03aa3bc730da9cd5766818cb5ecb78f8ecf3218d67f6cf285411a", "section_index_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/sections.json", "section_index_sha256": "78da1d7d0dff1259a965853e5f04d80b0c5ad788cd8cf93af8657c9d4d1e3d80", "bibliographic_identity_source_url": "https://aogavrilov.com/publications/where-quality-breaks/metadata.jsonld", "provenance": { "bibliography_text": "author-provided-camera-ready-tex", "bibliographic_identity": "validated-publication-record-copy-only", "citation_context": "exact-source-block", "identifier_inference": false }, "title": "Where Quality Breaks in Compressed Short-Text Generation: Staged Bottleneck Localization", "language": "en", "granularity": [ "bibliographic-reference", "bibliographic-identity", "citation-occurrence", "source-block" ], "statistics": { "references": 15, "cited_references": 15, "citation_occurrences": 15, "citing_blocks": 3 }, "references": [ { "id": "bib.bib1", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib1", "label": "[1]", "citation_text": "S. S. Sahoo, M. Arriola, Y. Schiff, A. Gokaslan, E. Marroquin, J. T. Chiu, A. Rush, and V. Kuleshov. Simple and Effective Masked Diffusion Language Models. NeurIPS, 2024.", "citation_text_sha256": "631da46aa0c9b2528532a1d24dc5be65c148b04deb00521fcc592f9478a2c4fc", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p2.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p2.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "1", "citation_cluster_text": "[1, 4, 5, 6, 7, 8, 9, 10]", "text": "For sequence generation in the latent domain, our implementation follows masked refinement and discrete diffusion ideas rather than fully autoregressive decoding. MDLM and related categorical diffusion or masked-refinement methods provide a practical way to denoise short discrete sequences in parallel [1, 4, 5, 6, 7, 8, 9, 10]. Their engineering promise is straightforward: if the codec preserves enough information, then generation in a shorter code sequence can offer a better quality-efficiency trade-off than direct token-space diffusion.", "text_sha256": "aad8e1ca673865d8ac358720ec7511a2fcf88c28ee5a2e470a7d8cd657ccef37" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/2406.07524", "number": 1, "title": "Simple and Effective Masked Diffusion Language Models", "authors": [ "Subham Sekhar Sahoo", "Marianne Arriola", "Yair Schiff", "Aaron Gokaslan", "Edgar Marroquin", "Justin T Chiu", "Alexander Rush", "Volodymyr Kuleshov" ], "year": 2024, "venue": "Advances in Neural Information Processing Systems", "url": "https://arxiv.org/abs/2406.07524", "doi": null, "arxiv": "2406.07524", "same_as": [] } }, { "id": "bib.bib2", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib2", "label": "[2]", "citation_text": "A. van den Oord, O. Vinyals, and K. Kavukcuoglu. Neural Discrete Representation Learning. NeurIPS, 2017.", "citation_text_sha256": "da02ca69a535ddd1387dc5d42cb5f8c31ae1f1018f46e85dc74b2750ecdc4a28", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p1.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p1.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "2", "citation_cluster_text": "[2, 3]", "text": "Discrete latent modeling in text extends VQ-VAE and VQ-VAE-2 style representation learning, where a codec maps a long sequence into a shorter discrete representation [2, 3]. In image generation, such tokenizers are widely used to make downstream generation cheaper. In text, the same factorization is attractive for short-sequence workloads, but semantic fidelity is more fragile because even a small reconstruction error can change meaning, syntax, or local factual detail.", "text_sha256": "93c7b9be0020e24ff3b7788362b3dcbb80bd4ea26bb4a85f3e31783631698770" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/1711.00937", "number": 2, "title": "Neural Discrete Representation Learning", "authors": [ "Aaron van den Oord", "Oriol Vinyals", "Koray Kavukcuoglu" ], "year": 2017, "venue": "Advances in Neural Information Processing Systems", "url": "https://arxiv.org/abs/1711.00937", "doi": null, "arxiv": "1711.00937", "same_as": [] } }, { "id": "bib.bib3", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib3", "label": "[3]", "citation_text": "A. Razavi, A. van den Oord, and O. Vinyals. Generating Diverse High-Fidelity Images with VQ-VAE-2. NeurIPS, 2019.", "citation_text_sha256": "a2a5fa677076fe5edca3ee0e52da86b05576e20bef320c90555d9c7e4609f40e", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p1.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p1.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "3", "citation_cluster_text": "[2, 3]", "text": "Discrete latent modeling in text extends VQ-VAE and VQ-VAE-2 style representation learning, where a codec maps a long sequence into a shorter discrete representation [2, 3]. In image generation, such tokenizers are widely used to make downstream generation cheaper. In text, the same factorization is attractive for short-sequence workloads, but semantic fidelity is more fragile because even a small reconstruction error can change meaning, syntax, or local factual detail.", "text_sha256": "93c7b9be0020e24ff3b7788362b3dcbb80bd4ea26bb4a85f3e31783631698770" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/1906.00446", "number": 3, "title": "Generating Diverse High-Fidelity Images with VQ-VAE-2", "authors": [ "Ali Razavi", "Aaron van den Oord", "Oriol Vinyals" ], "year": 2019, "venue": "Advances in Neural Information Processing Systems", "url": "https://arxiv.org/abs/1906.00446", "doi": null, "arxiv": "1906.00446", "same_as": [] } }, { "id": "bib.bib4", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib4", "label": "[4]", "citation_text": "J. Austin, D. D. Johnson, J. Ho, D. Tarlow, and R. van den Berg. Structured Denoising Diffusion Models in Discrete State-Spaces. NeurIPS, 2021.", "citation_text_sha256": "2f4847c3c01ef7a5e27d3df3421715671d3311fe6863750670d3c5e740c07d40", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p2.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p2.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "4", "citation_cluster_text": "[1, 4, 5, 6, 7, 8, 9, 10]", "text": "For sequence generation in the latent domain, our implementation follows masked refinement and discrete diffusion ideas rather than fully autoregressive decoding. MDLM and related categorical diffusion or masked-refinement methods provide a practical way to denoise short discrete sequences in parallel [1, 4, 5, 6, 7, 8, 9, 10]. Their engineering promise is straightforward: if the codec preserves enough information, then generation in a shorter code sequence can offer a better quality-efficiency trade-off than direct token-space diffusion.", "text_sha256": "aad8e1ca673865d8ac358720ec7511a2fcf88c28ee5a2e470a7d8cd657ccef37" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/2107.03006", "number": 4, "title": "Structured Denoising Diffusion Models in Discrete State-Spaces", "authors": [ "Jacob Austin", "Daniel D. Johnson", "Jonathan Ho", "Daniel Tarlow", "Rianne van den Berg" ], "year": 2021, "venue": "Advances in Neural Information Processing Systems", "url": "https://arxiv.org/abs/2107.03006", "doi": null, "arxiv": "2107.03006", "same_as": [] } }, { "id": "bib.bib5", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib5", "label": "[5]", "citation_text": "A. Lou, C. Meng, and S. Ermon. Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution. ICML, 2024.", "citation_text_sha256": "b19e06ab4089f4c6fae6c6032955689ddf581346df85329b3400f318c4f5913a", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p2.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p2.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "5", "citation_cluster_text": "[1, 4, 5, 6, 7, 8, 9, 10]", "text": "For sequence generation in the latent domain, our implementation follows masked refinement and discrete diffusion ideas rather than fully autoregressive decoding. MDLM and related categorical diffusion or masked-refinement methods provide a practical way to denoise short discrete sequences in parallel [1, 4, 5, 6, 7, 8, 9, 10]. Their engineering promise is straightforward: if the codec preserves enough information, then generation in a shorter code sequence can offer a better quality-efficiency trade-off than direct token-space diffusion.", "text_sha256": "aad8e1ca673865d8ac358720ec7511a2fcf88c28ee5a2e470a7d8cd657ccef37" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/2310.16834", "number": 5, "title": "Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution", "authors": [ "Aaron Lou", "Chenlin Meng", "Stefano Ermon" ], "year": 2024, "venue": "Proceedings of the 41st International Conference on Machine Learning", "url": "https://arxiv.org/abs/2310.16834", "doi": null, "arxiv": "2310.16834", "same_as": [] } }, { "id": "bib.bib6", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib6", "label": "[6]", "citation_text": "X. Han, S. Kumar, and Y. Tsvetkov. SSD-LM: Semi-autoregressive Simplex-based Diffusion Language Model for Text Generation and Modular Control. ACL, 2023.", "citation_text_sha256": "0f1493105a708a0c0d6676d0b7d235e0f1297f090e4fe29f6221234f30732fe0", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p2.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p2.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "6", "citation_cluster_text": "[1, 4, 5, 6, 7, 8, 9, 10]", "text": "For sequence generation in the latent domain, our implementation follows masked refinement and discrete diffusion ideas rather than fully autoregressive decoding. MDLM and related categorical diffusion or masked-refinement methods provide a practical way to denoise short discrete sequences in parallel [1, 4, 5, 6, 7, 8, 9, 10]. Their engineering promise is straightforward: if the codec preserves enough information, then generation in a shorter code sequence can offer a better quality-efficiency trade-off than direct token-space diffusion.", "text_sha256": "aad8e1ca673865d8ac358720ec7511a2fcf88c28ee5a2e470a7d8cd657ccef37" } ], "bibliographic_record": { "id": "https://doi.org/10.18653/v1/2023.acl-long.647", "number": 6, "title": "SSD-LM: Semi-autoregressive Simplex-based Diffusion Language Model for Text Generation and Modular Control", "authors": [ "Xiaochuang Han", "Sachin Kumar", "Yulia Tsvetkov" ], "year": 2023, "venue": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", "url": "https://doi.org/10.18653/v1/2023.acl-long.647", "doi": "10.18653/v1/2023.acl-long.647", "arxiv": "2210.17432", "same_as": [ "https://arxiv.org/abs/2210.17432" ] } }, { "id": "bib.bib7", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib7", "label": "[7]", "citation_text": "X. Li, J. Thickstun, I. Gulrajani, P. S. Liang, and T. B. Hashimoto. Diffusion-LM Improves Controllable Text Generation. NeurIPS, 2022.", "citation_text_sha256": "2c4471633ba568adc072e03bb9c628cca46ed30d12f40947ef1eaa12ca29a2d1", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p2.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p2.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "7", "citation_cluster_text": "[1, 4, 5, 6, 7, 8, 9, 10]", "text": "For sequence generation in the latent domain, our implementation follows masked refinement and discrete diffusion ideas rather than fully autoregressive decoding. MDLM and related categorical diffusion or masked-refinement methods provide a practical way to denoise short discrete sequences in parallel [1, 4, 5, 6, 7, 8, 9, 10]. Their engineering promise is straightforward: if the codec preserves enough information, then generation in a shorter code sequence can offer a better quality-efficiency trade-off than direct token-space diffusion.", "text_sha256": "aad8e1ca673865d8ac358720ec7511a2fcf88c28ee5a2e470a7d8cd657ccef37" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/2205.14217", "number": 7, "title": "Diffusion-LM Improves Controllable Text Generation", "authors": [ "Xiang Lisa Li", "John Thickstun", "Ishaan Gulrajani", "Percy Liang", "Tatsunori B. Hashimoto" ], "year": 2022, "venue": "Advances in Neural Information Processing Systems", "url": "https://arxiv.org/abs/2205.14217", "doi": null, "arxiv": "2205.14217", "same_as": [] } }, { "id": "bib.bib8", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib8", "label": "[8]", "citation_text": "H. Chang, H. Zhang, L. Jiang, C. Liu, and W. T. Freeman. MaskGIT: Masked Generative Image Transformer. CVPR, 2022.", "citation_text_sha256": "7a65407fc31048e1ec5249046eeb2f25ae22e9add8ad85bf316c4014206a2640", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p2.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p2.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "8", "citation_cluster_text": "[1, 4, 5, 6, 7, 8, 9, 10]", "text": "For sequence generation in the latent domain, our implementation follows masked refinement and discrete diffusion ideas rather than fully autoregressive decoding. MDLM and related categorical diffusion or masked-refinement methods provide a practical way to denoise short discrete sequences in parallel [1, 4, 5, 6, 7, 8, 9, 10]. Their engineering promise is straightforward: if the codec preserves enough information, then generation in a shorter code sequence can offer a better quality-efficiency trade-off than direct token-space diffusion.", "text_sha256": "aad8e1ca673865d8ac358720ec7511a2fcf88c28ee5a2e470a7d8cd657ccef37" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/2202.04200", "number": 8, "title": "MaskGIT: Masked Generative Image Transformer", "authors": [ "Huiwen Chang", "Han Zhang", "Lu Jiang", "Ce Liu", "William T. Freeman" ], "year": 2022, "venue": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition", "url": "https://openaccess.thecvf.com/content/CVPR2022/html/Chang_MaskGIT_Masked_Generative_Image_Transformer_CVPR_2022_paper.html", "doi": null, "arxiv": "2202.04200", "same_as": [ "https://openaccess.thecvf.com/content/CVPR2022/html/Chang_MaskGIT_Masked_Generative_Image_Transformer_CVPR_2022_paper.html" ] } }, { "id": "bib.bib9", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib9", "label": "[9]", "citation_text": "M. Ghazvininejad, O. Levy, Y. Liu, and L. Zettlemoyer. Mask-Predict: Parallel Decoding of Conditional Masked Language Models. EMNLP-IJCNLP, 2019.", "citation_text_sha256": "cc198841a814409218100c326d9f8d66aaf0f88c5036b3a0c96e72c8e4f16ba3", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p2.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p2.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "9", "citation_cluster_text": "[1, 4, 5, 6, 7, 8, 9, 10]", "text": "For sequence generation in the latent domain, our implementation follows masked refinement and discrete diffusion ideas rather than fully autoregressive decoding. MDLM and related categorical diffusion or masked-refinement methods provide a practical way to denoise short discrete sequences in parallel [1, 4, 5, 6, 7, 8, 9, 10]. Their engineering promise is straightforward: if the codec preserves enough information, then generation in a shorter code sequence can offer a better quality-efficiency trade-off than direct token-space diffusion.", "text_sha256": "aad8e1ca673865d8ac358720ec7511a2fcf88c28ee5a2e470a7d8cd657ccef37" } ], "bibliographic_record": { "id": "https://doi.org/10.18653/v1/D19-1633", "number": 9, "title": "Mask-Predict: Parallel Decoding of Conditional Masked Language Models", "authors": [ "Marjan Ghazvininejad", "Omer Levy", "Yinhan Liu", "Luke Zettlemoyer" ], "year": 2019, "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing", "url": "https://doi.org/10.18653/v1/D19-1633", "doi": "10.18653/v1/D19-1633", "arxiv": "1904.09324", "same_as": [ "https://arxiv.org/abs/1904.09324" ] } }, { "id": "bib.bib10", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib10", "label": "[10]", "citation_text": "E. Hoogeboom, D. Nielsen, P. Jaini, P. Forré, and M. Welling. Argmax Flows and Multinomial Diffusion: Learning Categorical Distributions. NeurIPS, 2021.", "citation_text_sha256": "58c1892a913e365a889be4b9eb57ed4d0d14f6a7c9ffb607e685229231c63279", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p2.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p2.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "10", "citation_cluster_text": "[1, 4, 5, 6, 7, 8, 9, 10]", "text": "For sequence generation in the latent domain, our implementation follows masked refinement and discrete diffusion ideas rather than fully autoregressive decoding. MDLM and related categorical diffusion or masked-refinement methods provide a practical way to denoise short discrete sequences in parallel [1, 4, 5, 6, 7, 8, 9, 10]. Their engineering promise is straightforward: if the codec preserves enough information, then generation in a shorter code sequence can offer a better quality-efficiency trade-off than direct token-space diffusion.", "text_sha256": "aad8e1ca673865d8ac358720ec7511a2fcf88c28ee5a2e470a7d8cd657ccef37" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/2102.05379", "number": 10, "title": "Argmax Flows and Multinomial Diffusion: Learning Categorical Distributions", "authors": [ "Emiel Hoogeboom", "Didrik Nielsen", "Priyank Jaini", "Patrick Forré", "Max Welling" ], "year": 2021, "venue": "Advances in Neural Information Processing Systems", "url": "https://arxiv.org/abs/2102.05379", "doi": null, "arxiv": "2102.05379", "same_as": [] } }, { "id": "bib.bib11", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib11", "label": "[11]", "citation_text": "A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever. Language Models are Unsupervised Multitask Learners. OpenAI Technical Report, 2019.", "citation_text_sha256": "0e2326af7bba12ccf980c62bbeef1b6cafc8b170b855e9d45f8aef6bcd7ce038", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p3.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p3.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "11", "citation_cluster_text": "[11]", "text": "Our evaluation setup combines one shared anchor metric with complementary semantic metrics. We use an external GPT-2 scorer because it can be applied consistently to originals, reconstructions, and generated text under one protocol [11]. For the geometry study, we additionally report SBERT similarity, BERTScore, MAUVE, and an LLM-judge summary as complementary text-level views rather than as interchangeable summary scores [12, 13, 14, 15]. Our contribution is therefore methodological and diagnostic: staged validation is used to localize the dominant bottleneck in one fixed implementation, not merely to report another benchmark comparison.", "text_sha256": "3b291a5209bc4ff759fd278b7d6e7edf0d33d36a8415117720f5def6b5d2f3be" } ], "bibliographic_record": { "id": "https://cdn.openai.com/better-language-models/language-models.pdf", "number": 11, "title": "Language Models are Unsupervised Multitask Learners", "authors": [ "Alec Radford", "Jeffrey Wu", "Rewon Child", "David Luan", "Dario Amodei", "Ilya Sutskever" ], "year": 2019, "venue": "OpenAI Technical Report", "url": "https://cdn.openai.com/better-language-models/language-models.pdf", "doi": null, "arxiv": null, "same_as": [] } }, { "id": "bib.bib12", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib12", "label": "[12]", "citation_text": "N. Reimers and I. Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. EMNLP-IJCNLP, 2019.", "citation_text_sha256": "5e826842d209d965c50ae4b64ac8ca55a53a8e6fccf93e72245895e30b2d8921", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p3.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p3.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "12", "citation_cluster_text": "[12, 13, 14, 15]", "text": "Our evaluation setup combines one shared anchor metric with complementary semantic metrics. We use an external GPT-2 scorer because it can be applied consistently to originals, reconstructions, and generated text under one protocol [11]. For the geometry study, we additionally report SBERT similarity, BERTScore, MAUVE, and an LLM-judge summary as complementary text-level views rather than as interchangeable summary scores [12, 13, 14, 15]. Our contribution is therefore methodological and diagnostic: staged validation is used to localize the dominant bottleneck in one fixed implementation, not merely to report another benchmark comparison.", "text_sha256": "3b291a5209bc4ff759fd278b7d6e7edf0d33d36a8415117720f5def6b5d2f3be" } ], "bibliographic_record": { "id": "https://doi.org/10.18653/v1/D19-1410", "number": 12, "title": "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks", "authors": [ "Nils Reimers", "Iryna Gurevych" ], "year": 2019, "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing", "url": "https://doi.org/10.18653/v1/D19-1410", "doi": "10.18653/v1/D19-1410", "arxiv": "1908.10084", "same_as": [ "https://arxiv.org/abs/1908.10084" ] } }, { "id": "bib.bib13", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib13", "label": "[13]", "citation_text": "T. Zhang, V. Kishore, F. Wu, K. Q. Weinberger, and Y. Artzi. BERTScore: Evaluating Text Generation with BERT. ICLR, 2020.", "citation_text_sha256": "3826d5e42ced4456f8f58b3e82ae93952955994f7a56cb54dc76416901b5c492", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p3.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p3.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "13", "citation_cluster_text": "[12, 13, 14, 15]", "text": "Our evaluation setup combines one shared anchor metric with complementary semantic metrics. We use an external GPT-2 scorer because it can be applied consistently to originals, reconstructions, and generated text under one protocol [11]. For the geometry study, we additionally report SBERT similarity, BERTScore, MAUVE, and an LLM-judge summary as complementary text-level views rather than as interchangeable summary scores [12, 13, 14, 15]. Our contribution is therefore methodological and diagnostic: staged validation is used to localize the dominant bottleneck in one fixed implementation, not merely to report another benchmark comparison.", "text_sha256": "3b291a5209bc4ff759fd278b7d6e7edf0d33d36a8415117720f5def6b5d2f3be" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/1904.09675", "number": 13, "title": "BERTScore: Evaluating Text Generation with BERT", "authors": [ "Tianyi Zhang", "Varsha Kishore", "Felix Wu", "Kilian Q. Weinberger", "Yoav Artzi" ], "year": 2020, "venue": "International Conference on Learning Representations", "url": "https://arxiv.org/abs/1904.09675", "doi": null, "arxiv": "1904.09675", "same_as": [] } }, { "id": "bib.bib14", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib14", "label": "[14]", "citation_text": "K. Pillutla, S. Swayamdipta, R. Zellers, J. Thickstun, S. Welleck, Y. Choi, and Z. Harchaoui. MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers. NeurIPS, 2021.", "citation_text_sha256": "2a1e9a72a7d6bc835fc11fde4f4940b5f09f1ef207e355c50f0a64688296dfba", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p3.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p3.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "14", "citation_cluster_text": "[12, 13, 14, 15]", "text": "Our evaluation setup combines one shared anchor metric with complementary semantic metrics. We use an external GPT-2 scorer because it can be applied consistently to originals, reconstructions, and generated text under one protocol [11]. For the geometry study, we additionally report SBERT similarity, BERTScore, MAUVE, and an LLM-judge summary as complementary text-level views rather than as interchangeable summary scores [12, 13, 14, 15]. Our contribution is therefore methodological and diagnostic: staged validation is used to localize the dominant bottleneck in one fixed implementation, not merely to report another benchmark comparison.", "text_sha256": "3b291a5209bc4ff759fd278b7d6e7edf0d33d36a8415117720f5def6b5d2f3be" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/2102.01454", "number": 14, "title": "MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers", "authors": [ "Krishna Pillutla", "Swabha Swayamdipta", "Rowan Zellers", "John Thickstun", "Sean Welleck", "Yejin Choi", "Zaid Harchaoui" ], "year": 2021, "venue": "Advances in Neural Information Processing Systems", "url": "https://arxiv.org/abs/2102.01454", "doi": null, "arxiv": "2102.01454", "same_as": [] } }, { "id": "bib.bib15", "url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#bib.bib15", "label": "[15]", "citation_text": "L. Zheng, W.-L. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, and I. Stoica. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS 2023 Datasets and Benchmarks Track.", "citation_text_sha256": "a2ad01cc2fe120570115c335dff254bb2406103ce41cc45e961f230236d4ceb7", "occurrence_count": 1, "contexts": [ { "block_id": "S2.p3.text", "block_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2.p3.text", "section_id": "S2", "section_url": "https://aogavrilov.com/publications/where-quality-breaks/full-text/#S2", "section_heading": "II Related Work", "anchor_text": "15", "citation_cluster_text": "[12, 13, 14, 15]", "text": "Our evaluation setup combines one shared anchor metric with complementary semantic metrics. We use an external GPT-2 scorer because it can be applied consistently to originals, reconstructions, and generated text under one protocol [11]. For the geometry study, we additionally report SBERT similarity, BERTScore, MAUVE, and an LLM-judge summary as complementary text-level views rather than as interchangeable summary scores [12, 13, 14, 15]. Our contribution is therefore methodological and diagnostic: staged validation is used to localize the dominant bottleneck in one fixed implementation, not merely to report another benchmark comparison.", "text_sha256": "3b291a5209bc4ff759fd278b7d6e7edf0d33d36a8415117720f5def6b5d2f3be" } ], "bibliographic_record": { "id": "https://arxiv.org/abs/2306.05685", "number": 15, "title": "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena", "authors": [ "Lianmin Zheng", "Wei-Lin Chiang", "Ying Sheng", "Siyuan Zhuang", "Zhanghao Wu", "Yonghao Zhuang", "Zi Lin", "Zhuohan Li", "Dacheng Li", "Eric P. Xing", "Hao Zhang", "Joseph E. Gonzalez", "Ion Stoica" ], "year": 2023, "venue": "Advances in Neural Information Processing Systems, Datasets and Benchmarks Track", "url": "https://arxiv.org/abs/2306.05685", "doi": null, "arxiv": "2306.05685", "same_as": [] } } ] }