pith. sign in

arxiv: 2606.25519 · v1 · pith:4KJIW2UUnew · submitted 2026-06-24 · 💻 cs.AI · cs.LG

Quantization Inflates Reasoning: Token Inflation as a Hidden Cost of Low-Bit Reasoning Models

Xinyu Lian , Walid Krichene , Beichen Huang , Masahiro Tanaka , Olatunji Ruwase , Li Zhang , Minjia Zhang This is my paper

Pith reviewed 2026-06-25 20:54 UTC · model grok-4.3

classification 💻 cs.AI cs.LG
keywords quantizationchain of thoughtreasoning modelstoken inflationlarge language modelsinference efficiencypost-training quantization
0
0 comments X

The pith

Low-bit post-training quantization makes reasoning models generate longer chains of thought even when accuracy stays the same.

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

The paper establishes that quantizing reasoning large language models to low bits such as INT4 or INT3 preserves final-answer accuracy across math, code, science, and agent benchmarks but causes the models to produce longer reasoning traces. This token inflation creates an extra test-time compute cost that can offset the expected per-token speed gains from quantization. The authors quantify the effect with a new metric called the CoT Token Inflation Ratio, which averages the increase in reasoning tokens between quantized and full-precision versions. They also document accompanying changes such as more intermediate steps and greater semantic repetition in the traces, which lead to measurable end-to-end serving penalties. Prompting and sampling mitigations give inconsistent results, while quantization-aware training reduces both accuracy loss and token inflation.

Core claim

Low-bit post-training quantization preserves accuracy on reasoning tasks but systematically increases the number of tokens used in the chain-of-thought, introducing a hidden inference-time cost that is captured by the CoT Token Inflation Ratio and visible in longer traces with more steps and repetition.

What carries the argument

The CoT Token Inflation Ratio, defined as the average ratio of reasoning-token counts between quantized and full-precision models across all evaluation benchmarks.

If this is right

  • Accuracy alone is not sufficient to judge the efficiency of quantized reasoning models.
  • End-to-end latency and energy use can rise even when per-token latency falls.
  • Quantization-aware training reduces both accuracy degradation and token inflation more reliably than post-hoc prompting fixes.

Where Pith is reading between the lines

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

  • Total energy budgets for large-scale inference deployments may need re-evaluation when quantized reasoning models are used.
  • The same inflation pattern could appear under other compression methods such as pruning or distillation.

Load-bearing premise

The measured increase in reasoning length is produced by the quantization step itself rather than by differences in decoding settings, prompts, or benchmark construction.

What would settle it

An experiment that applies identical prompts, decoding parameters, and temperature to the same model before and after quantization and finds no statistically significant rise in average reasoning-token count would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.25519 by Beichen Huang, Li Zhang, Masahiro Tanaka, Minjia Zhang, Olatunji Ruwase, Walid Krichene, Xinyu Lian.

Figure 2
Figure 2. Figure 2: Per-instance CoT step repetitiveness (higher [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: Per-instance CoT step counts under BF16 vs. INT3 GPTQ for Qwen3-4B (left) and Qwen3-30B (right) on MBPP. Each point plots BF16 steps (x-axis) against GPTQ steps (y-axis); the dashed line indicates parity. Points above the diagonal indicate more reason￾ing steps under quantization. Inflation increases step-level repetition. Longer traces are not necessarily worse if the extra tokens correspond to useful add… view at source ↗
Figure 4
Figure 4. Figure 4: GPTQ quantized Qwen3-4B evaluation on AIME25, showing a trade-off between CoT token and accuracy under different repetition penalty settings. appeared in the prompt or previously generated out￾put. The default repetition_penalty is 1, and we test values r ∈ {1.1, 1.2, 1.3, 1.4} on Qwen3- 4B and evaluate the performance on AIME25 as reported in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy–efficiency trade-off across calibra [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Correctness-conditioned CTIR for Qwen3-30B. We split questions by whether BF16 answers them [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
read the original abstract

Quantization is widely used to reduce the inference cost of large language models, but its effect on reasoning models is not fully captured by final-answer accuracy or per-token latency. We show that low-bit post-training quantization can introduce a hidden test-time compute cost: quantized reasoning models often generate longer chains of thought even when they still answer correctly. Across mathematical reasoning, code generation, scientific question answering, and agentic tool-use benchmarks, we find that INT4/INT3 quantization can preserve accuracy but increase reasoning-token usage, offsetting the expected per-token speedup. To measure this effect, we introduce the CoT Token Inflation Ratio, which compares reasoning length between quantized and full-precision models averaged across all evaluation benchmarks. We further show that token inflation is accompanied by behavioral changes in the reasoning trace, including more intermediate steps and greater semantic repetition. These changes translate into measurable end-to-end real-world serving penalties. Finally, we evaluate mitigation strategies and find that prompting and decoding-time sampling offer inconsistent accuracy-length trade-offs, while quantization-aware training shows more promise in reducing both accuracy degradation and token inflation. Our results suggest that reasoning-token usage should be reported alongside accuracy when evaluating quantized reasoning models.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 2 minor

Summary. The paper claims that low-bit post-training quantization of reasoning models can preserve final-answer accuracy while increasing chain-of-thought token counts, introducing a hidden test-time compute cost. It introduces the CoT Token Inflation Ratio to quantify this effect across mathematical reasoning, code generation, scientific question answering, and agentic tool-use benchmarks, reports accompanying changes such as more intermediate steps and semantic repetition, demonstrates end-to-end serving penalties, and evaluates mitigation strategies including prompting, sampling, and quantization-aware training.

Significance. If the central empirical claim holds after addressing controls, the work is significant for showing that accuracy alone is an incomplete metric for quantized reasoning models and that token usage must be reported alongside it. The multi-benchmark empirical measurement and introduction of a concrete ratio are strengths; the observation of behavioral changes in traces adds value beyond latency considerations.

major comments (3)
  1. [Abstract and Experiments] Abstract and Experiments section: the claim that quantization causes longer CoT on correctly answered questions is not supported by evidence that token lengths are conditioned on the intersection of questions solved correctly by both models or by paired per-question comparisons; without this, observed inflation could reflect differing success sets rather than per-instance behavioral change.
  2. [Abstract] Abstract: no information is supplied on statistical tests, variance across runs, or controls for confounding variables such as decoding hyperparameters, which prevents verification of the data-to-claim link for the reported inflation ratios.
  3. [Mitigation evaluation] Mitigation evaluation: the statement that quantization-aware training reduces both accuracy degradation and token inflation lacks reported effect sizes, confidence intervals, or ablation on the exact training configurations used, making the comparative claim difficult to assess.
minor comments (2)
  1. [Method] The exact mathematical definition of the CoT Token Inflation Ratio should be stated explicitly with an equation number rather than described only in prose.
  2. [Tables] Table captions should include the precise model sizes, bit widths, and benchmark subsets used for each reported ratio.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We have revised the manuscript to address the concerns regarding evidence for per-instance behavioral changes, statistical reporting, and mitigation evaluation details. Point-by-point responses follow.

read point-by-point responses

  1. Referee: [Abstract and Experiments] Abstract and Experiments section: the claim that quantization causes longer CoT on correctly answered questions is not supported by evidence that token lengths are conditioned on the intersection of questions solved correctly by both models or by paired per-question comparisons; without this, observed inflation could reflect differing success sets rather than per-instance behavioral change.

    Authors: We agree that aggregate comparisons alone leave open the possibility that inflation arises from differing success sets. In the revised manuscript we add (i) per-question paired token-length differences restricted to questions answered correctly by both models and (ii) explicit conditioning on the intersection set. These new analyses confirm that the inflation ratio remains positive and statistically detectable on the common correct subset, supporting the per-instance interpretation. revision: yes

  2. Referee: [Abstract] Abstract: no information is supplied on statistical tests, variance across runs, or controls for confounding variables such as decoding hyperparameters, which prevents verification of the data-to-claim link for the reported inflation ratios.

    Authors: We have expanded the Experiments section with (a) results from five independent runs using different random seeds, reporting mean inflation ratios together with standard deviations, (b) paired t-tests and Wilcoxon signed-rank tests on per-question token counts, and (c) an explicit statement that all compared models used identical decoding settings (greedy decoding, temperature = 0, top-p = 1). These additions directly link the reported ratios to controlled, reproducible measurements. revision: yes

  3. Referee: [Mitigation evaluation] Mitigation evaluation: the statement that quantization-aware training reduces both accuracy degradation and token inflation lacks reported effect sizes, confidence intervals, or ablation on the exact training configurations used, making the comparative claim difficult to assess.

    Authors: The revised Mitigation section now reports (i) absolute and relative effect sizes for both accuracy recovery and token-inflation reduction, (ii) 95 % confidence intervals obtained from the same multi-run protocol, and (iii) a full ablation table varying learning rate, number of epochs, and calibration-set size for the QAT procedure. These additions allow quantitative assessment of the mitigation claim. revision: yes

Circularity Check

0 steps flagged

No circularity: direct empirical measurements with no derivations or self-referential definitions

full rationale

The paper presents an empirical study measuring token lengths in quantized vs. full-precision models across benchmarks. It defines the CoT Token Inflation Ratio explicitly as a comparison of observed reasoning lengths, with no equations, fitted parameters, or predictions that reduce to inputs by construction. No self-citations are invoked as load-bearing premises, uniqueness theorems, or ansatzes. The central claim rests on reported measurements rather than any derivation chain that collapses to tautology. This is self-contained empirical reporting.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper reports empirical observations from benchmarks without introducing new free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5765 in / 1048 out tokens · 25466 ms · 2026-06-25T20:54:56.000240+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

300 extracted references · 45 canonical work pages

  1. [1]

    CVPR '12 , year =

    Lempitsky, Victor , title =. CVPR '12 , year =

  2. [2]

    Proceedings of the 8th USENIX Conference on Hot Topics in Storage and File Systems , series =

    Kim, Hyeong-Jun and Lee, Young-Sik and Kim, Jin-Soo , title =. Proceedings of the 8th USENIX Conference on Hot Topics in Storage and File Systems , series =. 2016 , location =

    2016

  3. [3]

    TPAMI 2011 , year =

    Jegou, Herve and Douze, Matthijs and Schmid, Cordelia , booktitle =. TPAMI 2011 , year =

    2011

  4. [4]

    Conference of the European Chapter of the Association for Computational Linguistics , year=

    What Makes Medical Claims (Un)Verifiable? Analyzing Entity and Relation Properties for Fact Verification , author=. Conference of the European Chapter of the Association for Computational Linguistics , year=

  5. [5]

    2023 , eprint=

    Mistral 7B , author=. 2023 , eprint=

    2023

  6. [6]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    2024

  7. [7]

    In: Proceedings of the 44th International ACM SIGIR Con- ference on Research and Development in Information Retrieval

    Lin, Jimmy and Ma, Xueguang and Lin, Sheng-Chieh and Yang, Jheng-Hong and Pradeep, Ronak and Nogueira, Rodrigo , title =. Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2021 , isbn =. doi:10.1145/3404835.3463238 , abstract =

    work page doi:10.1145/3404835.3463238 2021

  8. [8]

    Fleet , title =

    Mohammad Norouzi and David J. Fleet , title =. CVPR 2013 , Xpages =

    2013

  9. [9]

    CVPR 2013 , Xpages =

    Tiezheng Ge and Kaiming He and Qifa Ke and Jian Sun , title =. CVPR 2013 , Xpages =

    2013

  10. [10]

    Avrithis , title =

    Yannis Kalantidis and Yannis S. Avrithis , title =

  11. [11]

    Information Systems , pages =

    Yury Malkov and Alexander Ponomarenko and Andrey Logvinov and Vladimir Krylov , title =. Information Systems , pages =

  12. [12]

    Malkov and D

    Yury A. Malkov and D. A. Yashunin , title =. CoRR , volume =

  13. [13]

    Frank Hsu and Xiaojie Lan and Gabriel Miller and David Baird , title =

    D. Frank Hsu and Xiaojie Lan and Gabriel Miller and David Baird , title =. DASC'17 , Xpages =

  14. [14]

    2016 , location =

    Mueller, Jonas and Thyagarajan, Aditya , title =. 2016 , location =

    2016

  15. [15]

    Proceedings of the Seventh International Conference on World Wide Web 7 , series =

    Brin, Sergey and Page, Lawrence , title =. Proceedings of the Seventh International Conference on World Wide Web 7 , series =. 1998 , location =

    1998

  16. [16]

    Proceedings of the Sixth ACM International Conference on Web Search and Data Mining , series =

    Risvik, Knut Magne and Chilimbi, Trishul and Tan, Henry and Kalyanaraman, Karthik and Anderson, Chris , title =. Proceedings of the Sixth ACM International Conference on Web Search and Data Mining , series =. 2013 , isbn =

    2013

  17. [17]

    Barla and Baeza-Yates, Ricardo , title =

    Cambazoglu, B. Barla and Baeza-Yates, Ricardo , title =. Proceedings of the 37th International ACM SIGIR Conference on Research &\#38; Development in Information Retrieval , series =. 2014 , isbn =. doi:10.1145/2600428.2602291 , acmid =

    work page doi:10.1145/2600428.2602291 2014

  18. [18]

    CIKM '13 , year =

    Huang, Po-Sen and He, Xiaodong and Gao, Jianfeng and Deng, Li and Acero, Alex and Heck, Larry , title =. CIKM '13 , year =

  19. [19]

    Andoni, Alexandr and Indyk, Piotr and Razenshteyn, Ilya , journal=

  20. [20]

    2016 , url =

    Junaid Shuja and Kashif Bilal and Sajjad Ahmad Madani and Mazliza Othman and Rajiv Ranjan and Pavan Balaji and Samee Ullah Khan , title =. 2016 , url =. doi:10.1109/JSYST.2014.2315823 , timestamp =

    work page doi:10.1109/jsyst.2014.2315823 2016

  21. [21]

    Andersen and Michael Kaminsky , title =

    Hyeontaek Lim and Bin Fan and David G. Andersen and Michael Kaminsky , title =

  22. [22]

    Proceedings of the 42nd Annual International Symposium on Computer Architecture, Portland, OR, USA, June 13-17, 2015 , pages =

    David Lo and Liqun Cheng and Rama Govindaraju and Parthasarathy Ranganathan and Christos Kozyrakis , title =. Proceedings of the 42nd Annual International Symposium on Computer Architecture, Portland, OR, USA, June 13-17, 2015 , pages =. 2015 , crossref =. doi:10.1145/2749469.2749475 , timestamp =

    work page doi:10.1145/2749469.2749475 2015

  23. [23]

    Communications of the ACM , issue_date =

    Bentley, Jon Louis , title =. Communications of the ACM , issue_date =. 1975 , issn =

    1975

  24. [24]

    , title =

    Yianilos, Peter N. , title =. SODA '93 , year =

  25. [25]

    Lee, D. T. and Wong, C. K. , title =. Acta Informatica , issue_date =. 1977 , pages =

    1977

  26. [26]

    Proceedings of the ACM Conference of the Special Interest Group on Data Communication , series =

    Flach, Tobias and Dukkipati, Nandita and Terzis, Andreas and Raghavan, Barath and Cardwell, Neal and Cheng, Yuchung and Jain, Ankur and Hao, Shuai and Katz-Bassett, Ethan and Govindan, Ramesh , title =. Proceedings of the ACM Conference of the Special Interest Group on Data Communication , series =. 2013 , isbn =. doi:10.1145/2486001.2486014 , acmid =

    work page doi:10.1145/2486001.2486014 2013

  27. [27]

    Proceedings of the 32 Annual ACM Symposium on Theory of Computing , series =

    Kleinberg, Jon , title =. Proceedings of the 32 Annual ACM Symposium on Theory of Computing , series =. 2000 , isbn =. doi:10.1145/335305.335325 , acmid =

    work page doi:10.1145/335305.335325 2000

  28. [28]

    Gray, R. M. and Neuhoff, D. L. , title =. IEEE Trans. Inf. Theor. , issue_date =. 2006 , issn =. doi:10.1109/18.720541 , acmid =

    work page doi:10.1109/18.720541 2006

  29. [29]

    Mirrokni

    Mayur Datar and Nicole Immorlica and Piotr Indyk and Vahab S. Mirrokni , title =. Proceedings of the 20th. 2004 , crossref =. doi:10.1145/997817.997857 , timestamp =

    work page doi:10.1145/997817.997857 2004

  30. [30]

    VLDB'99 , pages =

    Aristides Gionis and Piotr Indyk and Rajeev Motwani , title =. VLDB'99 , pages =

  31. [31]

    Lloyd , title =

    Stuart P. Lloyd , title =

  32. [32]

    Similarity Search and Applications - 6th International Conference , series =

    Leonid Boytsov and Bilegsaikhan Naidan , title =. Similarity Search and Applications - 6th International Conference , series =

  33. [33]

    and Shaeffer, Ian and Gopalakrishnan, Liji and Lo, David and Lee, Benjamin C

    Malladi, Krishna T. and Shaeffer, Ian and Gopalakrishnan, Liji and Lo, David and Lee, Benjamin C. and Horowitz, Mark , title =. Proceedings of the 2012 45th Annual IEEE/ACM International Symposium on Microarchitecture , series =. 2012 , isbn =

    2012

  34. [34]

    Ding, Yufei and Zhao, Yue and Shen, Xipeng and Musuvathi, Madanlal and Mytkowicz, Todd , booktitle=

  35. [35]

    Proceedings of the 1995 ACM SIGMOD International Conference on Management of Data , series =

    Roussopoulos, Nick and Kelley, Stephen and Vincent, Fr. Proceedings of the 1995 ACM SIGMOD International Conference on Management of Data , series =. 1995 , isbn =

    1995

  36. [36]

    , title =

    Watts, Duncan J. , title =. 1999 , Xisbn =

    1999

  37. [37]

    IEEE Transactions on Knowledge and Data Engineering , year =

    Wen Li and Ying Zhang and Yifang Sun and Wei Wang and Wenjie Zhang and Xuemin Lin , title =. IEEE Transactions on Knowledge and Data Engineering , year =

  38. [38]

    [p1, 3] T

    Golovin, Daniel and Solnik, Benjamin and Moitra, Subhodeep and Kochanski, Greg and Karro, John and Sculley, D. , title =. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , series =. 2017 , isbn =. doi:10.1145/3097983.3098043 , acmid =

    work page doi:10.1145/3097983.3098043 2017

  39. [39]

    Proceedings of the 18th ACM/IFIP/USENIX Middleware Conference , series =

    Rasley, Jeff and He, Yuxiong and Yan, Feng and Ruwase, Olatunji and Fonseca, Rodrigo , title =. Proceedings of the 18th ACM/IFIP/USENIX Middleware Conference , series =. 2017 , isbn =. doi:10.1145/3135974.3135994 , acmid =

    work page doi:10.1145/3135974.3135994 2017

  40. [40]

    , title =

    Lowe, David G. , title =. Int. J. Comput. Vision , issue_date =

  41. [41]

    Lempitsky , title =

    Artem Babenko and Victor S. Lempitsky , title =

  42. [42]

    Lempitsky , title =

    Artem Babenko and Victor S. Lempitsky , title =. CoRR , volume =. 2014 , url =

    2014

  43. [43]

    and Swanson, Steven , title =

    Andersen, David G. and Swanson, Steven , title =. IEEE Micro , issue_date =. 2010 , issn =. doi:10.1109/MM.2010.71 , acmid =

    work page doi:10.1109/mm.2010.71 2010

  44. [44]

    Leventhal, Adam , title =. Commun. ACM , issue_date =. 2008 , issn =. doi:10.1145/1364782.1364796 , acmid =

    work page doi:10.1145/1364782.1364796 2008

  45. [45]

    Matthews, Jeanna and Trika, Sanjeev and Hensgen, Debra and Coulson, Rick and Grimsrud, Knut , title =. Trans. Storage , issue_date =. 2008 , issn =. doi:10.1145/1367829.1367830 , acmid =

    work page doi:10.1145/1367829.1367830 2008

  46. [46]

    Proceedings of the 36th International ACM SIGIR Conference on Research and Development in Information Retrieval , series =

    Wang, Jianguo and Lo, Eric and Yiu, Man Lung and Tong, Jiancong and Wang, Gang and Liu, Xiaoguang , title =. Proceedings of the 36th International ACM SIGIR Conference on Research and Development in Information Retrieval , series =. 2013 , isbn =

    2013

  47. [47]

    ACM Trans

    Wang, Jianguo and Lo, Eric and Yiu, Man Lung and Tong, Jiancong and Wang, Gang and Liu, Xiaoguang , title =. ACM Trans. Inf. Syst. , issue_date =. 2014 , issn =. doi:10.1145/2661629 , acmid =

    work page doi:10.1145/2661629 2014

  48. [48]

    Thomas Claburn , title =

  49. [49]

    2008 , howpublished =

    Miranda Miller , title =. 2008 , howpublished =

    2008

  50. [50]

    Petersen, Casper and Simonsen, Jakob Grue and Lioma, Christina , title =

  51. [51]

    SIGIR '17 , year =

    Goodwin, Bob and Hopcroft, Michael and Luu, Dan and Clemmer, Alex and Curmei, Mihaela and Elnikety, Sameh and He, Yuxiong , title =. SIGIR '17 , year =

  52. [52]

    CoRR , volume =

    Jeff Johnson and Matthijs Douze and Herv. CoRR , volume =. 2017 , url =

    2017

  53. [53]

    Ying Shan and Jian Jiao and Jie Zhu and J. C. Mao , title =. CoRR , volume =. 2018 , url =

    2018

  54. [54]

    CoRR , volume =

    Patrick Wieschollek and Oliver Wang and Alexander Sorkine. CoRR , volume =. 2017 , url =

    2017

  55. [55]

    Zhang, Jialiang and Khoram, Soroosh and Li, Jing , booktitle=

  56. [56]

    Matthijs Douze and Herv

  57. [57]

    1982 , crossref =

    Biing. 1982 , crossref =. doi:10.1109/ICASSP.1982.1171604 , timestamp =

    work page doi:10.1109/icassp.1982.1171604 1982

  58. [58]

    Proceedings of the 20th International Conference on World Wide Web,

    Wei Dong and Moses Charikar and Kai Li , title =. Proceedings of the 20th International Conference on World Wide Web,. 2011 , crossref =. doi:10.1145/1963405.1963487 , timestamp =

    work page doi:10.1145/1963405.1963487 2011

  59. [59]

    2018 , howpublished =

    Google AI Blog , title =. 2018 , howpublished =

    2018

  60. [60]

    2018 , howpublished =

    Bing blogs , title =. 2018 , howpublished =

    2018

  61. [61]

    2018 , howpublished =

    Adaline Lau , title =. 2018 , howpublished =

    2018

  62. [62]

    2018 , howpublished =

    Hamel Husain , title =. 2018 , howpublished =

    2018

  63. [63]

    2018 , Xpages =

    Matthijs Douze and Alexandre Sablayrolles and Herv. 2018 , Xpages =

    2018

  64. [64]

    Lowe , title =

    Marius Muja and David G. Lowe , title =. TPAMI 2014 , volume =

    2014

  65. [65]

    Razenshteyn and Ludwig Schmidt , title =

    Alexandr Andoni and Piotr Indyk and Thijs Laarhoven and Ilya P. Razenshteyn and Ludwig Schmidt , title =. Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015 , pages =

    2015

  66. [66]

    Jure Leskovec and Rok Sosic , title =

  67. [67]

    Herve Jegou and Romain Tavenard and Matthijs Douze and Laurent Amsaleg , title =

  68. [68]

    ICMR 2015 , pages =

    Shoou. ICMR 2015 , pages =

    2015

  69. [69]

    Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015 , pages =

    Xiang Zhang and Junbo Jake Zhao and Yann LeCun , title =. Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015 , pages =

    2015

  70. [70]

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

    Tomas Mikolov and Wen. Human Language Technologies: Conference of the North American Chapter of the Association of Computational Linguistics , pages =

  71. [71]

    Accessed: 05-20-2019 , title =

    2019

  72. [72]

    2018 , Xnote =

    Microsoft AI & Research , month =. 2018 , Xnote =

    2018

  73. [73]

    VLDB'19 , year =

    Fu, Cong and Xiang, Chao and Wang, Changxu and Cai, Deng , title =. VLDB'19 , year =

  74. [74]

    FAQ: All about the Google RankBrain algorithm , author=

  75. [75]

    CoRR , volume =

    Lei Yu and Karl Moritz Hermann and Phil Blunsom and Stephen Pulman , title =. CoRR , volume =. 2014 , Xurl =

    2014

  76. [76]

    CIKM '16 , year =

    Van Gysel, Christophe and de Rijke, Maarten and Kanoulas, Evangelos , title =. CIKM '16 , year =

  77. [77]

    Bruce Croft , title =

    Mostafa Dehghani and Hamed Zamani and Aliaksei Severyn and Jaap Kamps and W. Bruce Croft , title =. SIGIR 2017 , pages =

    2017

  78. [78]

    Bruce Croft , title =

    Jiafeng Guo and Yixing Fan and Qingyao Ai and W. Bruce Croft , title =

  79. [79]

    Bhaskar Mitra and Fernando Diaz and Nick Craswell , title =

  80. [80]

    SIGIR 2017 , pages =

    Chenyan Xiong and Zhuyun Dai and Jamie Callan and Zhiyuan Liu and Russell Power , title =. SIGIR 2017 , pages =

    2017

Showing first 80 references.