pith. sign in

arxiv: 2606.12303 · v1 · pith:CAO5SB4Anew · submitted 2026-06-10 · 💻 cs.CV

From 2D Grids to 1D Tokens: Reforming Shared Representations for Multimodal Image Fusion

Yuchen Xian , Yunqiu Xu , Yang He , Yi Yang This is my paper

Pith reviewed 2026-06-27 09:56 UTC · model grok-4.3

classification 💻 cs.CV
keywords multimodal image fusion1D tokensselective token editingglobal coherencelocal fidelityfrozen pretrained tokenizershared representations
0
0 comments X

The pith

Multimodal image fusion improves by using 1D tokens for global appearance alongside 2D grids for local details.

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

This paper establishes that shared representations for multimodal image fusion can be reformed by adding a compact 1D token interface from a frozen pretrained image tokenizer. The 1D space models non-local appearance and base factors while the 2D pathway handles local structure restoration. Selective Token Editing sparsely updates a small set of critical tokens to steer global coherence without fine-tuning the tokenizer or adding extra losses. Experiments across four benchmarks demonstrate the best overall performance with gains in both global coherence and local fidelity metrics. This hybrid approach addresses the limitation of pure 2D grids in capturing image-level global factors.

Core claim

By introducing a 1D token interface based on a frozen pretrained image tokenizer for modeling non-local appearance factors and Selective Token Editing to sparsely update critical tokens, the approach achieves superior balance between global consistency and local fidelity in multimodal image fusion.

What carries the argument

Selective Token Editing (STE), which provides a lightweight mechanism to steer global appearance coherence by sparsely updating or replacing critical tokens in the 1D token space.

Load-bearing premise

The frozen pretrained image tokenizer supplies a 1D token space that can serve as an effective global carrier for non-local appearance factors without requiring fine-tuning or serving as a reconstruction backbone.

What would settle it

A direct comparison on the four benchmarks where the proposed method does not outperform existing 2D grid approaches in overall performance or multi-metric scores would falsify the advantage of the 1D token interface.

Figures

Figures reproduced from arXiv: 2606.12303 by Yang He, Yi Yang, Yuchen Xian, Yunqiu Xu.

Figure 1
Figure 1. Figure 1: 2D grids vs. 1D tokens for base/detail decoupling in multimodal image fusion. (a) A 2D shared feature map entangles global base appearance with local details. (b) We represent base in a compact 1D token set Z, map it to a base map via π(·), and combine it with a spatial detail map D for decoding. serving structural backgrounds, target saliency, fine-grained textures, and coherent global appearance. For ins… view at source ↗
Figure 2
Figure 2. Figure 2: Two-stage training scheme of our multimodal image fusion framework. The key design is to introduce a compact 1D token interface for global appearance/base modeling, while preserving the 2D pathway for local detail reconstruction. (A) Stage I (Reconstruction Warm-Up): the model learns stable modality-specific base/detail representations through per-modality reconstruction. (B) Stage II (Fusion Training): th… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparisons on M3 FD, RoadScene, TNO and Harvard datasets. Benefiting from the concentrated semantic information encoded in our 1D token representation, our method produces fused images with enhanced global coherence and sharper local structures compared to existing methods. The reconstruction loss is defined as L (m) rec = αssim LSSIM I (m) , ˆI (m)  +αmse [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 4
Figure 4. Figure 4: (a)Qualitative comparisons of object detection performance in smoke scene. (b)Qualitative comparisons of semantic segmentation performance in nighttime scene. Quantitative Comparison [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Hard selection frequency of the learned token-position selector under the 2-slot setting, selecting position 12, while Slot 1 consistently selects position 18, for both visible (VI) and infrared (IR) inputs. verify that the STE positions are not manually assigned, we use a lightweight Gumbel-Softmax (Jang et al., 2017; Mad￾dison et al., 2017) selector to learn token-position choices under different slot bu… view at source ↗
Figure 6
Figure 6. Figure 6: shows the full Gumbel-Softmax selector distributions under different slot budgets. With one slot, the selector concentrates on a single dominant position. With two slots, which is our final setting, the selector identifies two complemen￾tary positions. When the number of slots is further increased, additional slots are assigned to weaker residual positions, suggesting diminishing returns beyond the 2-slot … view at source ↗
Figure 7
Figure 7. Figure 7: Slot-wise qualitative comparison of sparse token manipulation. Slot 0 mainly produces a sharpening-oriented effect, while Slot 1 mainly produces a background-smoothing-oriented effect. Jointly editing Slot 0 and Slot 1 yields a better balance between detail enhancement and global appearance consistency. C.3. Interpretation of the Selected Slots The selected slots should not be interpreted as universal sema… view at source ↗
Figure 8
Figure 8. Figure 8: provides representative qualitative comparisons across several challenging fusion scenarios. In urban night scenes, where visible textures are weak and thermal targets dominate, our method enhances salient infrared responses while preserving the surrounding visible structures. In road and low-light surveillance scenes, existing methods tend to suffer from over-smoothed textures, unstable brightness, or dis… view at source ↗
Figure 9
Figure 9. Figure 9: Additional qualitative comparisons on the M3 FD dataset. Our method better preserves salient infrared targets while maintaining visible structural details under nighttime and low-illumination conditions. Ir CDDFuse Vis DCEvo Text-DiFuse Ours Ir CDDFuse Vis DCEvo Text-DiFuse Ours Ir CDDFuse Vis DCEvo Text-DiFuse Ours Ir CDDFuse Vis DCEvo Text-DiFuse Ours Ir Vis DCEvo CDDFuse Text-DiFuse Ours Ir Vis DCEvo CD… view at source ↗
Figure 10
Figure 10. Figure 10: Additional qualitative comparisons on the M3 FD dataset under challenging road scenes. Compared with existing methods, our method produces more coherent global brightness and clearer object boundaries. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional qualitative comparisons on the M3 FD dataset. Our method reduces visual artifacts and improves the balance between thermal saliency and visible-scene structure. MRI CT DCEvo CDDFuse Text-DiFuse Ours MRI CT DCEvo CDDFuse Text-DiFuse Ours MRI PET DCEvo CDDFuse Text-DiFuse Ours MRI SPECT DCEvo CDDFuse Text-DiFuse Ours [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional qualitative comparisons on the Harvard medical image fusion dataset. Our method preserves anatomical structures while maintaining complementary modality-specific information from CT, PET, and SPECT images. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
read the original abstract

Multimodal image fusion aims to integrate complementary information from different modalities into a fused image that preserves rich local details while maintaining globally consistent appearance. Existing approaches build shared representations on 2D feature grids, which excel at modeling local structures but offer limited leverage over image-level global appearance factors. To balance these objectives, we introduce a compact 1D token interface based on a frozen pretrained image tokenizer for modeling non-local appearance/base factors. Rather than using the tokenizer as a reconstruction backbone, our design uses the 1D token space as a global carrier while retaining the 2D spatial pathway for local structure restoration. Specifically, we introduce Selective Token Editing (STE), which sparsely updates/replaces a small set of critical tokens, providing a lightweight mechanism to steer global appearance coherence while keeping the fusion backbone unchanged and avoiding extra losses. Experiments on four commonly used benchmarks show that our method achieves the best overall performance, with consistent, multi-metric improvements in both global coherence and local fidelity. Project page: https://zju-xyc.github.io/1D-Fusion-Project-Page/

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

0 major / 2 minor

Summary. The paper claims that multimodal image fusion can be improved by introducing a compact 1D token interface from a frozen pretrained image tokenizer to model non-local/global appearance factors, while retaining the conventional 2D spatial pathway for local structure. The core mechanism is Selective Token Editing (STE), which sparsely updates or replaces a small number of critical tokens to steer global coherence without fine-tuning the tokenizer, without using it as a reconstruction backbone, and without extra losses. The authors report that this yields the best overall performance across four standard benchmarks, with consistent gains in both global coherence and local fidelity metrics.

Significance. If the reported results and ablations hold, the work demonstrates a lightweight architectural separation of global and local pathways that leverages existing frozen tokenizers as global carriers. This avoids the computational cost of full fine-tuning or reconstruction objectives and provides a sparse editing interface that could generalize to other multimodal vision tasks. The approach is notable for being parameter-light on the global side and for supplying controls that isolate the 1D interface contribution.

minor comments (2)
  1. [Abstract] Abstract: the performance claim ('best overall performance... consistent, multi-metric improvements') would be more informative if the abstract briefly cited one or two representative metric gains or the number of baselines beaten, even if full tables appear later.
  2. [Method] The description of STE would benefit from an explicit statement of how the 'critical tokens' are selected (e.g., attention-based, gradient-based, or fixed heuristic) and whether this selection is deterministic across runs.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of our work on the 1D token interface and Selective Token Editing for multimodal image fusion, as well as for recommending acceptance.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The manuscript presents an architectural choice: a frozen pretrained tokenizer supplies a 1D token space used as a global carrier, paired with a retained 2D pathway and a new Selective Token Editing (STE) operator. No equations, predictions, or uniqueness claims are shown that reduce by construction to fitted parameters defined inside the method itself. The central claims rest on explicit design decisions plus multi-metric experimental results on external benchmarks, with ablations that isolate the 1D interface contribution. No self-citation chain, ansatz smuggling, or renaming of known results appears in the provided derivation steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that a frozen tokenizer's 1D space can model global appearance without adaptation, and introduces STE as a new mechanism whose effectiveness is asserted via benchmark results.

axioms (1)
  • domain assumption The 1D token space from a frozen pretrained image tokenizer can effectively model non-local appearance and base factors when used as a global carrier.
    Invoked to justify retaining the 2D pathway for local restoration while delegating global coherence to the 1D tokens.
invented entities (1)
  • Selective Token Editing (STE) no independent evidence
    purpose: Sparsely update or replace a small set of critical tokens to steer global appearance coherence.
    New lightweight editing operation introduced to avoid extra losses and backbone changes.

pith-pipeline@v0.9.1-grok · 5730 in / 1304 out tokens · 24361 ms · 2026-06-27T09:56:21.115924+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Multi-modality Image Fusion under Adverse Weather: Mask-Guided Feature Restoration and Interaction

    cs.CV 2026-06 unverdicted novelty 5.0

    Mask-guided MMIF method with pseudo ground truth and cross-attention for feature restoration and interaction under adverse weather, claiming SOTA results on synthetic and real data.

Reference graph

Works this paper leans on

142 extracted references · 8 linked inside Pith · cited by 1 Pith paper

  1. [1]

    ICML , pages=

    Crafting papers on machine learning , author=. ICML , pages=

  2. [2]

    IEEE TIP , year=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE TIP , year=

  3. [3]

    IEEE TCOM , year=

    Image quality measures and their performance , author=. IEEE TCOM , year=

  4. [4]

    International Journal of Engineering Science Invention , year=

    Hybrid multimodality medical image fusion technique for feature enhancement in medical diagnosis , author=. International Journal of Engineering Science Invention , year=

  5. [5]

    Optics Communications , year=

    Detail preserved fusion of visible and infrared images using regional saliency extraction and multi-scale image decomposition , author=. Optics Communications , year=

  6. [6]

    arXiv preprint arXiv:2401.00421 , year=

    From text to pixels: A context-aware semantic synergy solution for infrared and visible image fusion , author=. arXiv preprint arXiv:2401.00421 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2603.03871 , year=

    Bridging Human Evaluation to Infrared and Visible Image Fusion , author=. arXiv preprint arXiv:2603.03871 , year=

    arXiv

  8. [8]

    ACM MM , year=

    Toward a Training-Free Plug-and-Play Refinement Framework for Infrared and Visible Image Registration and Fusion , author=. ACM MM , year=

  9. [9]

    NeurIPS , year=

    Efficient Rectified Flow for Image Fusion , author=. NeurIPS , year=

  10. [10]

    arXiv preprint arXiv:2603.14214 , year=

    UniFusion: A Unified Image Fusion Framework with Robust Representation and Source-Aware Preservation , author=. arXiv preprint arXiv:2603.14214 , year=

    arXiv

  11. [11]

    AAAI , year=

    Domain Adaptation Guided Infrared and Visible Image Fusion , author=. AAAI , year=

  12. [12]

    IEEE/CAA Journal of Automatica Sinica , year=

    PromptFusion: Harmonized semantic prompt learning for infrared and visible image fusion , author=. IEEE/CAA Journal of Automatica Sinica , year=

  13. [13]

    arXiv preprint arXiv:2601.03955 , year=

    ResTok: Learning Hierarchical Residuals in 1D Visual Tokenizers for Autoregressive Image Generation , author=. arXiv preprint arXiv:2601.03955 , year=

    arXiv

  14. [14]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  15. [15]

    M. J. Kearns , title =

  16. [16]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  17. [17]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  18. [18]

    Suppressed for Anonymity , author=

  19. [19]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

  20. [20]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959

  21. [21]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  22. [22]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  23. [23]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  24. [24]

    Proceedings of the 40th International Conference on Machine Learning (ICML) , pages =

    Fast inference from transformers via speculative decoding , author =. Proceedings of the 40th International Conference on Machine Learning (ICML) , pages =. 2023 , volume =

    2023

  25. [25]

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

    SpecTr: Fast Speculative Decoding via Optimal Transport , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  26. [26]

    IEEE Transactions on Information Theory , volume =

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

  27. [27]

    arXiv preprint arXiv:2307.02764 , year =

    When Does Confidence-Based Cascade Deferral Suffice? , author =. arXiv preprint arXiv:2307.02764 , year =

    arXiv

  28. [28]

    International Conference on Learning Representations (ICLR) , year =

    Language Model Cascades: Token-level Uncertainty and Beyond , author =. International Conference on Learning Representations (ICLR) , year =

  29. [29]

    Proceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS) , year =

    Speculative Decoding with Big Little Decoder , author =. Proceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS) , year =

  30. [30]

    Advances in Neural Information Processing Systems , volume =

    Blockwise parallel decoding for deep autoregressive models , author =. Advances in Neural Information Processing Systems , volume =

  31. [31]

    Patterson , title =

    David A. Patterson , title =. Communications of the ACM , volume =. 2004 , publisher =

    2004

  32. [32]

    Hennessy and David A

    John L. Hennessy and David A. Patterson , title =. 2012 , publisher =

    2012

  33. [33]

    CoRR , volume =

    Xiaoxuan Liu and Lanxiang Hu and Peter Bailis and Ion Stoica and Zhijie Deng and Alvin Cheung and Hao Zhang , title =. CoRR , volume =. 2023 , url =

    2023

  34. [34]

    Svirschevski and A

    R. Svirschevski and A. May and Z. Chen and B. Chen and Z. Jia and M. Ryabinin , title =. arXiv preprint arXiv:2406.02532 , year =

    arXiv

  35. [35]

    Hooper and S

    C. Hooper and S. Kim and H. Mohammadzadeh and H. Genc and K. Keutzer and A. Gholami and S. Shao , title =. arXiv preprint arXiv:2310.12072 , year =

    arXiv

  36. [36]

    Shazeer , title =

    Noam M. Shazeer , title =. arXiv preprint arXiv:1911.02150 , year =

    Pith/arXiv arXiv 1911

  37. [37]

    F. W. Burton , title =. IEEE Transactions on Computers , volume =. 1985 , doi =

    1985

  38. [38]

    The Thirteenth International Conference on Learning Representations , year=

    Faster Cascades via Speculative Decoding , author=. The Thirteenth International Conference on Learning Representations , year=

  39. [39]

    Information Geometry and Its Applications , author =

  40. [40]

    Elements of Information Theory (2nd ed.) , author =

  41. [41]

    2016 , publisher=

    Information Geometry and Its Applications , author=. 2016 , publisher=

    2016

  42. [42]

    arXiv preprint arXiv:2406.17276 , year =

    Opt-Tree: Speculative Decoding with Adaptive Draft Tree Structure , author =. arXiv preprint arXiv:2406.17276 , year =

    arXiv

  43. [43]

    arXiv preprint arXiv:2305.09781 , year =

    SpecInfer: Accelerating Generative Large Language Model Serving with Tree-based Speculative Inference and Verification , author =. arXiv preprint arXiv:2305.09781 , year =

    arXiv

  44. [44]

    arXiv preprint arXiv:2401.10774 , year =

    Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads , author =. arXiv preprint arXiv:2401.10774 , year =

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2401.15077 , year =

    Eagle: Speculative Sampling Requires Rethinking Feature Uncertainty , author =. arXiv preprint arXiv:2401.15077 , year =

    Pith/arXiv arXiv

  46. [46]

    CoRR , volume =

    Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding , author =. CoRR , volume =. 2023 , url =

    2023

  47. [47]

    CoRR , volume =

    Predictive Pipelined Decoding: A Compute-Latency Trade-off for Exact LLM Decoding , author =. CoRR , volume =. 2023 , url =

    2023

  48. [48]

    arXiv preprint arXiv:2402.12374 , year =

    Sequoia: Scalable, Robust, and Hardware-Aware Speculative Decoding , author =. arXiv preprint arXiv:2402.12374 , year =

    arXiv

  49. [49]

    arXiv preprint arXiv:2409.16560 , year =

    Dynamic-Width Speculative Beam Decoding for Efficient LLM Inference , author =. arXiv preprint arXiv:2409.16560 , year =

    arXiv

  50. [50]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages =

    Speculative decoding: Exploiting speculative execution for accelerating seq2seq generation , author =. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages =. 2023 , address =

    2023

  51. [51]

    2006 , publisher=

    Elements of Information Theory , author=. 2006 , publisher=

    2006

  52. [52]

    Transactions of the Association for Computational Linguistics , year=

    Speculative decoding with token-wise acceptance prediction , author=. Transactions of the Association for Computational Linguistics , year=

  53. [53]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Training deeper neural networks by skip-layer connections , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  54. [54]

    CVPR , year=

    Densely connected convolutional networks , author=. CVPR , year=

  55. [55]

    NeurIPS , year=

    LLM-ZIP: Efficient LLM Inference via Layer Skipping , author=. NeurIPS , year=

  56. [56]

    2023 , eprint=

    DistillSpec: Improving Speculative Decoding via Knowledge Distillation , author=. 2023 , eprint=

    2023

  57. [57]

    Leibler , title =

    Solomon Kullback and Richard A. Leibler , title =. Annals of Mathematical Statistics , volume =. 1951 , publisher =

    1951

  58. [58]

    2025 , eprint=

    Cascade Speculative Drafting for Even Faster LLM Inference , author=. 2025 , eprint=

    2025

  59. [59]

    arXiv preprint arXiv:2412.18934 , year=

    Dovetail: A CPU/GPU heterogeneous speculative decoding for LLM inference , author=. arXiv preprint arXiv:2412.18934 , year=

    arXiv

  60. [60]

    and Zhou, Z

    Li, C. and Zhou, Z. and Zheng, S. and Zhang, J. and Liang, Y. and Sun, G. , booktitle=. 2024 , publisher=

    2024

  61. [61]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), Long Papers , pages =

    CLaSp: In-Context Layer Skip for Self-Speculative Decoding , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), Long Papers , pages =. 2025 , publisher =

    2025

  62. [62]

    Advances in Neural Information Processing Systems , volume =

    Speculative Decoding with Big Little Decoder , author =. Advances in Neural Information Processing Systems , volume =. 2023 , publisher =

    2023

  63. [63]

    arXiv preprint arXiv:2302.01318 , year=

    Accelerating large language model decoding with speculative sampling , author=. arXiv preprint arXiv:2302.01318 , year=

    Pith/arXiv arXiv

  64. [64]

    2025 , url=

    Heming Xia and Yongqi Li and Jun Zhang and Cunxiao Du and Wenjie Li , booktitle=. 2025 , url=

    2025

  65. [65]

    Unsupervised Thoughts (blog) , author=

    An optimal lossy variant of speculative decoding , url=. Unsupervised Thoughts (blog) , author=

  66. [66]

    arXiv preprint arXiv:2403.06075 , year=

    Multisize dataset condensation , author=. arXiv preprint arXiv:2403.06075 , year=

    arXiv

  67. [67]

    Journal of Machine Learning Research , volume=

    Exploring the limits of transfer learning with a unified text-to-text transformer , author=. Journal of Machine Learning Research , volume=. 2020 , url=

    2020

  68. [68]

    arXiv preprint arXiv:2408.00118 , year=

    Gemma 2: Improving open language models at a practical size , author=. arXiv preprint arXiv:2408.00118 , year=

    Pith/arXiv arXiv

  69. [69]

    2017 , publisher=

    Markov chains and mixing times , author=. 2017 , publisher=

    2017

  70. [70]

    Proceedings of the Ninth Workshop on Statistical Machine Translation , pages=

    Findings of the 2014 workshop on statistical machine translation , author=. Proceedings of the Ninth Workshop on Statistical Machine Translation , pages=

    2014

  71. [71]

    arXiv preprint arXiv:2408.11850 , year=

    Pearl: Parallel speculative decoding with adaptive draft length , author=. arXiv preprint arXiv:2408.11850 , year=

    arXiv

  72. [72]

    arXiv preprint arXiv:2406.16858 , year=

    Eagle-2: Faster inference of language models with dynamic draft trees , author=. arXiv preprint arXiv:2406.16858 , year=

    arXiv

  73. [73]

    arXiv preprint arXiv:2503.01840 , year=

    Eagle-3: Scaling up inference acceleration of large language models via training-time test , author=. arXiv preprint arXiv:2503.01840 , year=

    Pith/arXiv arXiv

  74. [74]

    Advances in Neural Information Processing Systems , volume=

    Teaching machines to read and comprehend , author=. Advances in Neural Information Processing Systems , volume=

  75. [75]

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages=

    Don't give me the details, just the summary! Topic-aware convolutional neural networks for extreme summarization , author=. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages=

    2018

  76. [76]

    arXiv preprint arXiv:2108.07732 , year=

    Program synthesis with large language models , author=. arXiv preprint arXiv:2108.07732 , year=

    Pith/arXiv arXiv

  77. [77]

    Transactions of the Association for Computational Linguistics , volume=

    Natural questions: A benchmark for question answering research , author=. Transactions of the Association for Computational Linguistics , volume=. 2019 , doi=

    2019

  78. [78]

    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 , doi=

    2017

  79. [79]

    Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing , pages=

    Semantic parsing on freebase from question-answer pairs , author=. Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing , pages=

    2013

  80. [80]

    Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing , pages=

    SQuAD: 100,000+ questions for machine comprehension of text , author=. Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing , pages=. 2016 , doi=

    2016

Showing first 80 references.