pith. sign in

arxiv: 2605.21604 · v1 · pith:SSIVSNTHnew · submitted 2026-05-20 · 💻 cs.MA

Argo: Efficient Importance Labeling for Enterprise Email Systems

Siddhant Ray , Ganesh Ananthanarayanan , Kevin Chian , Yan Guo , Cristina St Hill , Jack W. Stokes , Victor Wang , Junchen Jiang This is my paper

Pith reviewed 2026-05-22 08:46 UTC · model grok-4.3

classification 💻 cs.MA
keywords email importance labelinglarge language modelscost reductionprofilingenterprise emailinference optimizationon-demand scaling
0
0 comments X

The pith

Argo's profiler identifies labeling schemes that cut inference costs by 148-167X while preserving near-GPT quality for enterprise email.

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

The paper introduces Argo to address the high cost of using large language models for labeling email importance at enterprise scale. It creates a profiler that explores cheaper alternative labeling methods and selects those offering similar quality at much lower expense. An additional on-demand system adjusts resources to real-time demand and avoids cost spikes. Experiments across three public email datasets confirm large reductions in both inference and profiling expenses with minimal impact on label accuracy. This makes intelligent, context-sensitive email prioritization viable for businesses handling high volumes of messages.

Core claim

Argo constructs a profiler to search the space of possible labeling schemes for cost-efficient options that approach the quality of GPT-4.1, then applies these at scale using on-demand provisioning to manage variable loads. The result is practical large-scale context-aware labeling without the prohibitive costs of full-scale LLM inference.

What carries the argument

The profiler, which searches the cost-quality trade-off space to identify efficient labeling alternatives to full GPT-4.1 models.

Load-bearing premise

Cheaper labeling schemes found by the profiler will deliver nearly the same quality as GPT-4.1 when applied to actual enterprise email data and distributions.

What would settle it

Testing the Argo-chosen schemes on a large, private enterprise email corpus and finding that labeling accuracy falls significantly below GPT-4.1 levels.

Figures

Figures reproduced from arXiv: 2605.21604 by Cristina St Hill, Ganesh Ananthanarayanan, Jack W. Stokes, Junchen Jiang, Kevin Chian, Siddhant Ray, Victor Wang, Yan Guo.

Figure 1
Figure 1. Figure 1: At real-scale and increasing email exchange vol￾umes, cost of labeling with LLMs becomes prohibitive exchanged per day across the world and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Label overlaps vary significantly across different subsets which should be used to guide the labeling system study on the distribution of several importance labels , which can be assigned to emails in a user’s inbox. • Priority - This label provides an overall importance ranking on how important the email is measured on a discrete scale from 1 (lowest) to 5 (highest). • Needs Reply - This binary label dete… view at source ↗
Figure 3
Figure 3. Figure 3: Lower-cost alternatives to LLM labeling Knob Description L SLM cascade or embedding classifier M Different SLMs and embedding models T Confidence values used in cascade O Order of SLMs in cascade S Size of calibration email set [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Argo performs offline profiling to determine knob values and decisions to be used for email labeling [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Embedding classifier suffices for binary labeling Norm. Quality Norm. Cost Reduction Cost-Focus Quality-Focus Balanced The dashed line is the Pareto-Front [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Argo’s default policy chooses the Balanced point generated embeddings are reused across all binary labels. Ad￾ditionally, the classifier is completely offloaded to the CPU for training and inference , which saves on GPU compute costs. Further micro-optimizations in hyperparameter tuning such as weight decay [31], onecycle learning rate scheduling [53] and dropout [55] improves the classifier performance. T… view at source ↗
Figure 10
Figure 10. Figure 10: Efficiently choosing confidence thresholds on the Pareto Front for use in the SLM cascade Below these thresholds, individual SLMs achieve poor quality-cost tradeoffs [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Log-prob confidence distribution for two open￾source SLMs on a calibration set from the Enron Dataset labeling of any enterprise-scale messaging workloads, even beyond emails (e.g., instant messaging), as our system does not make any specific assumptions constrained to emails. 4.3 How do we manage the cost of profiling? Incrementally building the calibration email set: A nat￾ural challenge in Argo’s desig… view at source ↗
Figure 12
Figure 12. Figure 12: Argo finds multiple Pareto Efficient cost-quality tradeoff points for the SLM cascade and chooses the balance point (blue square) which achieves 148-167× cost reduction with comparable quality Better Better [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Argo resource provisioning is 2.2-3.8× lower cost than baselines across multiple penalty factors and capacity bottlenecks (results on Enron Dataset) Better Better [PITH_FULL_IMAGE:figures/full_fig_p011_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Argo achieves 20-640000× lower profiling costs compared to all baselines maintaining same or higher quality • 10% Sample - This profiler sweeps over all profiler config￾uration knobs from a random 10% email calibration set. • Reduced Cascade - This profiler only profiles with chosen 3 SLMs for the cascade with best individual quality-cost tradeoffs on the full calibration set. • Reduced Thresholds - This … view at source ↗
Figure 16
Figure 16. Figure 16: Only using the SLM cascade or embedding clas￾sifier achieves 2× higher cost or 13-15% lower quality Better Better [PITH_FULL_IMAGE:figures/full_fig_p012_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Argo enables specifying operator constraints (e.g., SLM subsets to be used) and provides a tradeoff solution which enforces the constraints labels. Argo achieves the same quality as the exhaustive profiler but is 640000× cheaper. The 1% and 10% sample baselines, while closer to Argo in cost (31000× and 4850× cheaper than exhaustive respectively), achieve 33% and 21% lower quality respectively. The Reduced… view at source ↗
read the original abstract

Email importance labeling has long been a critical yet challenging problem for businesses and individuals. Traditional approaches; such as keyword matching, user-defined rules, and sender-based heuristics; demand extensive manual feature engineering and fail to scale effectively or generalize. Recent advances in large language models (LLMs) demonstrate strong potential and a natural fit for this task, offering deep contextual understanding and superior labeling quality. However, using LLM models like GPT-4.1 at enterprise email volumes incurs prohibitive computational costs and hinders real-world deployment. We explore the trade-off space of using alternative labeling schemes as opposed to GPT4.1 scale LLMs, with the goal of achieving near GPT level labeling quality with significantly lower cost. We develop Argo, an enterprise email labeling framework, where we construct a profiler to efficiently search the cost quality trade-off space of labeling and identify cost-efficient alternatives to labeling emails. Additionally, we design an on-demand provisioning scheme to intelligently scale Argo with real time load, to minimize cost increases during peak load inference. Over 3 open-source email datasets, Argo achieves 148-167X inference cost reduction with negligible quality degradation and 20-640000X lower profiling costs, making large-scale, context-aware email labeling practical for enterprises.

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

1 major / 2 minor

Summary. The manuscript introduces Argo, an enterprise email labeling framework that constructs a profiler to search the cost-quality trade-off space among labeling schemes and designs an on-demand provisioning scheme to scale inference under real-time load. The central empirical claim is that, over three open-source email datasets, Argo delivers 148-167X inference cost reduction relative to GPT-4.1 with negligible quality degradation and 20-640000X lower profiling costs, thereby making large-scale context-aware email importance labeling practical.

Significance. If the reported cost-quality trade-offs hold under enterprise conditions, the work would remove a major computational barrier to deploying LLM-based contextual labeling at business scale. The profiler-plus-provisioning design directly targets the inference-cost bottleneck that currently prevents adoption of high-quality models for high-volume email streams.

major comments (1)
  1. [Abstract] Abstract: All quantitative results (148-167X inference reduction, negligible quality loss, 20-640000X profiling-cost savings) are reported exclusively on three open-source email datasets. Enterprise email differs systematically in volume, thread length, sender diversity, privacy constraints, and content distribution. No domain-shift experiments, ablation on enterprise-like characteristics, or representativeness argument is supplied to show that the profiler-selected cheaper schemes retain near-GPT fidelity when these distributional differences are present. This extrapolation is load-bearing for the central deployment claim.
minor comments (2)
  1. [Abstract] Abstract: punctuation is inconsistent ('Traditional approaches; such as keyword matching, user-defined rules, and sender-based heuristics; demand'). Replace the semicolons with commas or restructure the sentence for readability.
  2. [Abstract] Abstract: model name alternates between 'GPT-4.1' and 'GPT4.1'. Standardize throughout the manuscript.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's significance. We address the major comment on generalizability to enterprise settings below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: All quantitative results (148-167X inference reduction, negligible quality loss, 20-640000X profiling-cost savings) are reported exclusively on three open-source email datasets. Enterprise email differs systematically in volume, thread length, sender diversity, privacy constraints, and content distribution. No domain-shift experiments, ablation on enterprise-like characteristics, or representativeness argument is supplied to show that the profiler-selected cheaper schemes retain near-GPT fidelity when these distributional differences are present. This extrapolation is load-bearing for the central deployment claim.

    Authors: We acknowledge that all reported numbers come from the three open-source datasets and that no direct domain-shift experiments on enterprise data are included. Enterprise email does differ in the ways noted, and we cannot perform experiments on proprietary enterprise corpora due to privacy constraints. However, the profiler is explicitly designed to be run on whatever target distribution is available, empirically locating the cost-quality frontier for that specific data rather than assuming a fixed scheme. The Enron, Avocado, and third dataset already contain substantial variation in thread structure, sender diversity, and topical content. In the revision we will add an explicit representativeness discussion comparing key statistics of these datasets to published enterprise email characterizations, plus a limitations paragraph on the extrapolation. We believe this addresses the concern without overclaiming while preserving the central methodological contribution. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical profiler results on open-source datasets

full rationale

The paper describes an empirical systems framework that constructs a profiler to search labeling cost-quality trade-offs and reports measured speedups on three open-source email datasets. No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear in the abstract or described approach. The central claims rest on direct experimental measurements rather than any reduction to inputs by construction, satisfying the self-contained benchmark criterion.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so no free parameters, axioms, or invented entities can be identified from the text.

pith-pipeline@v0.9.0 · 5773 in / 1080 out tokens · 42151 ms · 2026-05-22T08:46:50.580794+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

71 extracted references · 71 canonical work pages

  1. [1]

    Qwen2.5 32b instruct

    Alibaba/Qwen Team. Qwen2.5 32b instruct. https://huggingface.co/Q wen/Qwen2.5-32B-Instruct, 2025. SLM, 20×cheaper

    work page 2025

  2. [2]

    Work hard, play hard: Email classification on the avocado and enron corpora

    Sakhar Alkhereyf and Owen Rambow. Work hard, play hard: Email classification on the avocado and enron corpora. InProceedings of TextGraphs-11: the Workshop on Graph-based Methods for Natural Lan- guage Processing, pages 57–65. Association for Computational Linguis- tics, 2017

    work page 2017

  3. [3]

    Show emails from vip senders in mail on mac

    Apple. Show emails from vip senders in mail on mac. https://support.ap ple.com/guide/mail/show-emails-from-vip-senders-mail40589/mac. Accessed: 2025-12-08

    work page 2025

  4. [4]

    Summarize notifications and reduce interruptions with apple intelligence on iphone

    Apple Support. Summarize notifications and reduce interruptions with apple intelligence on iphone. https://support.apple.com/is- is/guide/iphone/iph1fbe7d2b9/ios, 2025. Accessed December 8, 2025

    work page 2025

  5. [5]

    Normalized wasser- stein distance for mixture distributions with applications in adversarial learning and domain adaptation, 2019

    Yogesh Balaji, Rama Chellappa, and Soheil Feizi. Normalized wasser- stein distance for mixture distributions with applications in adversarial learning and domain adaptation, 2019

    work page 2019

  6. [6]

    Automatic categorization of email into folders: Bench- mark experiments on enron and sri corpora

    Ron Bekkerman. Automatic categorization of email into folders: Bench- mark experiments on enron and sri corpora

    work page

  7. [7]

    Supervised learning of universal sentence representa- tions from natural language inference data

    Alexis Conneau, Douwe Kiela, Holger Schwenk, Loïc Barrault, and Antoine Bordes. Supervised learning of universal sentence representa- tions from natural language inference data. In Martha Palmer, Rebecca Hwa, and Sebastian Riedel, editors,Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 670–680, Copenhagen, Den...

    work page 2017

  8. [8]

    Deepseek distilled models

    DeepSeek AI. Deepseek distilled models. https://huggingface.co /deepseek-ai, 2024. Distilled variants including DeepSeek-LLM and R1-Distill

    work page 2024

  9. [9]

    Hierarchical attention networks for email classification

    Sheng Deng, Wei Wang, and Jian Sun. Hierarchical attention networks for email classification. InAAAI Conference on Artificial Intelligence, 2018

    work page 2018

  10. [10]

    Bert: Pre-training of deep bidirectional transformers for language understanding.Proceedings of NAACL, 2019

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding.Proceedings of NAACL, 2019

    work page 2019

  11. [11]

    Dujian Ding, Ankur Mallick, Chi Wang, Robert Sim, Subhabrata Mukherjee, Victor Ruhle, Laks V. S. Lakshmanan, and Ahmed Hassan Awadallah. Hybrid llm: Cost-efficient and quality-aware query routing, 2024

    work page 2024

  12. [12]

    E-mail as habitat: an explo- ration of embedded personal information management.Interactions, 8(5):30–38, September 2001

    Nicolas Ducheneaut and Victoria Bellotti. E-mail as habitat: an explo- ration of embedded personal information management.Interactions, 8(5):30–38, September 2001

    work page 2001

  13. [13]

    Gpt-4 technical report, 2024

    OpenAI et al. Gpt-4 technical report, 2024

    work page 2024

  14. [14]

    Pytorch lightning

    William Falcon and the PyTorch Lightning Team. Pytorch lightning. https://github.com/Lightning-AI/pytorch-lightning, 2025. Version 2.5.4 (accessed 2025-12-09)

    work page 2025

  15. [15]

    Enron email dataset

    Federal Energy Regulatory Commission. Enron email dataset. https: //www.cs.cmu.edu/~enron/, 2004. Accessed 2025-12-09

    work page 2004

  16. [16]

    Gemma 3 27b it

    Google DeepMind. Gemma 3 27b it. https://huggingface.co/google/ge mma-3-27b-it, 2025. SLM, 90×cheaper

    work page 2025

  17. [17]

    Recipient recommendation in enterprises using communication graphs and email content

    David Graus, David van Dijk, Manos Tsagkias, Wouter Weerkamp, and Maarten de Rijke. Recipient recommendation in enterprises using communication graphs and email content. InProceedings of the 37th International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR ’14, page 1079–1082, New York, NY, USA,

    work page

  18. [19]

    Language model cascades: Token-level uncertainty and beyond

    Neha Gupta, Harikrishna Narasimhan, Wittawat Jitkrittum, Ankit Singh Rawat, Aditya Krishna Menon, and Sanjiv Kumar. Language model cascades: Token-level uncertainty and beyond. In B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun, editors,International Conference on Representation Learning, volume 2024, pages 4147–4180, 2024

    work page 2024

  19. [20]

    Haibo He and Edwardo A. Garcia. Learning from imbalanced data. IEEE Trans. on Knowl. and Data Eng., 21(9):1263–1284, September 2009

    work page 2009

  20. [21]

    Gradual domain adap- tation: Theory and algorithms.Journal of Machine Learning Research, 25(361):1–40, 2024

    Yifei He, Haoxiang Wang, Bo Li, and Han Zhao. Gradual domain adap- tation: Theory and algorithms.Journal of Machine Learning Research, 25(361):1–40, 2024

    work page 2024

  21. [22]

    Gibbons, and Onur Mutlu

    Kevin Hsieh, Ganesh Ananthanarayanan, Peter Bodik, Shivaram Venkataraman, Paramvir Bahl, Matthai Philipose, Phillip B. Gibbons, and Onur Mutlu. Focus: Querying large video datasets with low la- tency and low cost. In13th USENIX Symposium on Operating Systems Design and Implementation (OSDI 18), pages 269–286, Carlsbad, CA, October 2018. USENIX Association

    work page 2018

  22. [23]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adapta- tion of large language models. InProceedings of the 10th International Conference on Learning Representations (ICLR), 2022

    work page 2022

  23. [24]

    Routerbench: A benchmark for multi-llm routing system, 2024

    Qitian Jason Hu, Jacob Bieker, Xiuyu Li, Nan Jiang, Benjamin Keigwin, Gaurav Ranganath, Kurt Keutzer, and Shriyash Kaustubh Upadhyay. Routerbench: A benchmark for multi-llm routing system, 2024

    work page 2024

  24. [25]

    Chameleon: scalable adaptation of video analytics

    Junchen Jiang, Ganesh Ananthanarayanan, Peter Bodik, Siddhartha Sen, and Ion Stoica. Chameleon: scalable adaptation of video analytics. InProceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication, SIGCOMM ’18, page 253–266, New York, NY, USA, 2018. Association for Computing Machinery

    work page 2018

  25. [26]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.International Conference on Learning Representations (ICLR), 2015

    work page 2015

  26. [27]

    Email classification with co-training

    Svetlana Kiritchenko and Stan Matwin. Email classification with co-training. InProceedings of the 2011 Conference of the Center for Ad- vanced Studies on Collaborative Research, CASCON ’11, page 301–312, USA, 2011. IBM Corp

    work page 2011

  27. [28]

    The enron corpus: A new dataset for email classification research

    Bryan Klimt and Yiming Yang. The enron corpus: A new dataset for email classification research. InEuropean Conference on Machine Learning (ECML), 2004

    work page 2004

  28. [29]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. InProceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles (SOSP), 2023

    work page 2023

  29. [30]

    Detecting emails containing requests for action

    Andrew Lampert, Robert Dale, and Cecile Paris. Detecting emails containing requests for action. InHuman Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association 13 for Computational Linguistics, pages 984–992, 2010

    work page 2010

  30. [31]

    Zhang, Shuyi Wang, Ziang Tang, Fang Han, Zohaib Hassan, Jianqiao Zheng, and Avinash Changrani

    Xunzhuo Liu, Huamin Chen, Samzong Lu, Yossi Ovadia, Guohong Wen, Hao Wu, Zhengda Tan, Jintao Zhang, Senan Zedan, Yehudit Kerido, Liav Weiss, Haichen Zhang, Bishen Yu, Asaad Balum, Noa Limoy, Abdallah Samara, Baofa Fan, Brent Salisbury, Ryan Cook, Zhijie Wang, Qiping Pan, Rehan Khan, Avishek Goswami, Houston H. Zhang, Shuyi Wang, Ziang Tang, Fang Han, Zoha...

    work page 2026

  31. [32]

    Decoupled weight decay regulariza- tion

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regulariza- tion. InInternational Conference on Learning Representations, 2019

    work page 2019

  32. [33]

    Lane, and Mengwei Xu

    Zhenyan Lu, Xiang Li, Dongqi Cai, Rongjie Yi, Fangming Liu, Wei Liu, Jian Luan, Xiwen Zhang, Nicholas D. Lane, and Mengwei Xu. Demys- tifying small language models for edge deployment. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors,Proceedings of the 63rd Annual Meeting of the Association for Computational Linguist...

    work page 2025

  33. [34]

    Wendy E. Mackay. More than just a communication system: Diversity in the use of electronic mail. InProceedings of the ACM Conference on Computer-Supported Cooperative Work (CSCW), pages 344–353. ACM, 1988

    work page 1988

  34. [35]

    Topic and role discovery in social networks with experiments on enron and academic email.J

    Andrew McCallum, Xuerui Wang, and Andrés Corrada-Emmanuel. Topic and role discovery in social networks with experiments on enron and academic email.J. Artif. Int. Res., 30(1):249–272, October 2007

    work page 2007

  35. [36]

    Llama 3.1 8b instruct

    Meta AI. Llama 3.1 8b instruct. https://huggingf ace.co/meta- llama/Llama-3.1-8B-Instruct, 2025. SLM, 100×cheaper

    work page 2025

  36. [37]

    Llama 3.3 70b instruct

    Meta AI. Llama 3.3 70b instruct. https://huggingf ace.co/meta- llama/Llama-3.3-70B-Instruct, 2025. SLM, 10×cheaper

    work page 2025

  37. [38]

    What is focused inbox? https://support.microsoft.com/en- us/office/what-is-focused-inbox-16b24373-dfa9-4139-ab19-08aa75 3a6055

    Microsoft. What is focused inbox? https://support.microsoft.com/en- us/office/what-is-focused-inbox-16b24373-dfa9-4139-ab19-08aa75 3a6055. Accessed: 2025-12-08

    work page 2025

  38. [39]

    Phi-4-mini-instruct

    Microsoft. Phi-4-mini-instruct. https://huggingface.co/microsoft/phi- 4-mini-instruct, 2025. SLM, 105×cheaper

    work page 2025

  39. [40]

    What is azure ai foundry? https://learn.microsoft.co m/en-us/azure/ai-foundry/what-is-azure-ai-foundry, 2025

    Microsoft Learn. What is azure ai foundry? https://learn.microsoft.co m/en-us/azure/ai-foundry/what-is-azure-ai-foundry, 2025. Accessed December 9, 2025

    work page 2025

  40. [41]

    Mistral 7b instruct

    Mistral AI. Mistral 7b instruct. https://huggingface.co/mistralai/Mistr al-7B-Instruct-v0.2, 2024. Open-weight 7B instruction-tuned model

    work page 2024

  41. [42]

    Towards mod- eling legitimate and unsolicited email traffic using social network properties

    Farnaz Moradi, Tomas Olovsson, and Philippas Tsigas. Towards mod- eling legitimate and unsolicited email traffic using social network properties. InProceedings of the Fifth Workshop on Social Network Sys- tems, SNS ’12, New York, NY, USA, 2012. Association for Computing Machinery

    work page 2012

  42. [43]

    Efficient large-scale language model training on gpu clusters

    Deepak Narayanan et al. Efficient large-scale language model training on gpu clusters. InUSENIX OSDI, 2021

    work page 2021

  43. [44]

    Gonzalez, M

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tian- hao Wu, Joseph E. Gonzalez, M. Waleed Kadous, and Ion Stoica. Routellm: Learning to route llms with preference data.arXiv preprint arXiv:2406.11635, 2024

    work page arXiv 2024

  44. [45]

    Onnx: Open neural network exchange

    ONNX Community. Onnx: Open neural network exchange. https: //onnx.ai, 2025. Accessed 2025-12-09

    work page 2025

  45. [46]

    Openai api

    OpenAI. Openai api. https://openai.com/blog/openai-api/, 2020. Accessed December 9, 2025

    work page 2020

  46. [47]

    Openai api pricing — text token costs

    OpenAI. Openai api pricing — text token costs. https://platform.ope nai.com/docs/pricing, 2025. Accessed December 9, 2025

    work page 2025

  47. [48]

    text-embedding-3-large

    OpenAI. text-embedding-3-large. https://platform.openai.com/docs /models, 2025. Embedding model, 100×cheaper

    work page 2025

  48. [49]

    Fauci emails dataset, 2021

    Hani Osman. Fauci emails dataset, 2021

    work page 2021

  49. [50]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Pytorch: An imperative style, high-performance deep learning library. InAdvances in Neural Information Processing Systems 32 (NeurIPS 2019), 2019

    work page 2019

  50. [51]

    Sentence-BERT: Sentence embed- dings using Siamese BERT-networks

    Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embed- dings using Siamese BERT-networks. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors,Proceedings of the 2019 Confer- ence on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP- IJCNLP), pages 3982–3...

    work page 2019

  51. [52]

    The enron email dataset: Database schema and brief statistical report

    Jigar Shetty and Jafar Adibi. The enron email dataset: Database schema and brief statistical report. InInformation Retrieval Research, 2004

    work page 2004

  52. [53]

    Learning with weak supervision for email intent detection

    Kai Shu, Subhabrata Mukherjee, Guoqing Zheng, Ahmed Hassan Awadallah, Milad Shokouhi, and Susan Dumais. Learning with weak supervision for email intent detection. InProceedings of the 43rd In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’20, page 1051–1060, New York, NY, USA,

    work page

  53. [54]

    Association for Computing Machinery

    work page

  54. [55]

    Cyclical learning rates for training neural net- works.2017 IEEE Winter Conference on Applications of Computer Vision (W ACV), pages 464–472, 2017

    Leslie N Smith. Cyclical learning rates for training neural net- works.2017 IEEE Winter Conference on Applications of Computer Vision (W ACV), pages 464–472, 2017

    work page 2017

  55. [56]

    Semantic agree- ment enables efficient open-ended LLM cascades

    Duncan Soiffer, Steven Kolawole, and Virginia Smith. Semantic agree- ment enables efficient open-ended LLM cascades. In Saloni Potdar, Lina Rojas-Barahona, and Sebastien Montella, editors,Proceedings of the 2025 Conference on Empirical Methods in Natural Language Process- ing: Industry Track, pages 2499–2537, Suzhou (China), November 2025. Association for...

    work page 2025

  56. [57]

    Dropout: A simple way to prevent neural networks from overfitting.Journal of Machine Learning Research, 15(1):1929–1958, 2014

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting.Journal of Machine Learning Research, 15(1):1929–1958, 2014

    work page 1929

  57. [58]

    Is your gmail inbox setup slowing you down? https://www.computerworld.com/article/3511582/is-your-gmail- inbox-setup-slowing-you-down.html

    Computerworld Staff. Is your gmail inbox setup slowing you down? https://www.computerworld.com/article/3511582/is-your-gmail- inbox-setup-slowing-you-down.html. Accessed: 2025-12-08

    work page arXiv 2025

  58. [59]

    Gmail categories and inbox tabs

    Jesicca Stockett. Gmail categories and inbox tabs. https://swatkb.atlas sian.net/wiki/spaces/GA/pages/19661188/Gmail+Categories+and+In box+Tabs. Accessed: 2025-12-08

    work page arXiv 2025

  59. [60]

    Energy and policy considerations for deep learning in nlp

    Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep learning in nlp. InACL, 2019

    work page 2019

  60. [61]

    Line: Large-scale information network embedding

    Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei. Line: Large-scale information network embedding. In Proceedings of the 24th International Conference on World Wide Web, WWW ’15, page 1067–1077, Republic and Canton of Geneva, CHE,

    work page

  61. [62]

    International World Wide Web Conferences Steering Committee

    work page

  62. [63]

    Department of State

    U.S. Department of State. Hillary clinton email archive. https://wikile aks.org/clinton-emails/, 2016. FOIA Release; Accessed 2025-12-09

    work page 2016

  63. [64]

    Improving text embeddings with large language models

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Ma- jumder, and Furu Wei. Improving text embeddings with large language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computa- tional Linguistics (Volume 1: Long Papers), pages 11897–11916, Bangkok, Thailand, August ...

    work page 2024

  64. [65]

    Context-aware intent identification in email conversa- tions

    Wei Wang, Saghar Hosseini, Ahmed Awadallah, Paul Bennett, and Chris Quirk. Context-aware intent identification in email conversa- tions. InProceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1209–1218. ACM, 2019

    work page 2019

  65. [66]

    Email overload: exploring per- sonal information management of email

    Steve Whittaker and Candace Sidner. Email overload: exploring per- sonal information management of email. InProceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI ’96, page 276–283, New York, NY, USA, 1996. Association for Computing Ma- chinery

    work page 1996

  66. [67]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Syl- vain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 14 Transformers: State-of-th...

    work page 2020

  67. [68]

    MadEye: Boosting live video analytics accuracy with adaptive camera configurations

    Mike Wong, Murali Ramanujam, Guha Balakrishnan, and Ravi Ne- travali. MadEye: Boosting live video analytics accuracy with adaptive camera configurations. In21st USENIX Symposium on Networked Sys- tems Design and Implementation (NSDI 24), pages 549–568, Santa Clara, CA, April 2024. USENIX Association

    work page 2024

  68. [69]

    Charac- terizing and predicting enterprise email reply behavior

    Liu Yang, Susan Dumais, Paul Bennett, and Ahmed Awadallah. Charac- terizing and predicting enterprise email reply behavior. InProceedings of the 40th International ACM SIGIR Conference on Research and Devel- opment in Information Retrieval, pages 505–514. ACM, 2017

    work page 2017

  69. [70]

    Hierarchical attention networks for document classifi- cation

    Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard Hovy. Hierarchical attention networks for document classifi- cation. InNAACL, 2016

    work page 2016

  70. [71]

    Freedman

    Haoyu Zhang, Ganesh Ananthanarayanan, Peter Bodik, Matthai Phili- pose, Paramvir Bahl, and Michael J. Freedman. Live video analytics at scale with approximation and Delay-Tolerance. In14th USENIX Symposium on Networked Systems Design and Implementation (NSDI 17), pages 377–392, Boston, MA, March 2017. USENIX Association

    work page 2017

  71. [72]

    Email reply and importance prediction using pre-trained language models.Information Processing and Management, 2021

    Rui Zhang and Chen Li. Email reply and importance prediction using pre-trained language models.Information Processing and Management, 2021. 15

    work page 2021