pith. sign in

arxiv: 2508.18167 · v2 · pith:346YKRKBnew · submitted 2025-08-25 · 💻 cs.CL · cs.HC

DiscussLLM: Teaching Large Language Models When to Speak

Deep Anil Patel , Iain Melvin , Christopher Malon , Martin Renqiang Min This is my paper

Pith reviewed 2026-05-21 22:04 UTC · model grok-4.3

classification 💻 cs.CL cs.HC
keywords large language modelsconversational AIintervention timingdata synthesissilent tokenmulti-turn discussionsproactive agentswhen to speak
0
0 comments X

The pith

LLMs can learn to decide when to speak by training on a dataset of synthesized multi-turn discussions annotated with intervention triggers.

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

The paper establishes that large language models can move beyond passive responses by learning to predict when an intervention would add value in ongoing human conversations. It achieves this with a two-stage pipeline that generates realistic multi-turn discussions, each annotated with one of five intervention types such as factual correction or concept definition and marked by an explicit trigger point. Models are trained to emit a special silent token when no intervention is appropriate, teaching them to stay quiet until they can contribute helpfully. Two model architectures are tested: an integrated end-to-end system and a decoupled classifier-generator setup for lower latency. The work focuses on accurate timing of interventions alongside generation of useful replies.

Core claim

By creating a large-scale dataset of realistic multi-turn human discussions annotated with five intervention types and explicit conversational triggers where AI input adds value, models can be trained to predict a special silent token when no intervention is needed, allowing them to remain quiet until a helpful contribution can be made.

What carries the argument

A scalable two-stage data generation pipeline that synthesizes multi-turn discussions annotated with intervention types and explicit triggers, paired with training on a silent token to signal no intervention.

If this is right

  • Models learn to output a silent token and avoid speaking when no helpful intervention is possible.
  • Accurate timing of contributions makes LLMs more effective as collaborative partners.
  • The decoupled architecture supports low-latency inference suitable for live conversations.
  • Evaluation on timing accuracy and response quality provides concrete benchmarks for proactive behavior.

Where Pith is reading between the lines

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

  • The same silent-token approach could be adapted to other interactive settings such as live coding sessions or classroom tutoring.
  • Testing the models directly in unscripted human-AI chats would reveal how well the synthesized triggers generalize beyond the generated data.
  • Adding signals for speaker intent or topic shifts might refine the decision of when an intervention is truly valuable.

Load-bearing premise

The synthesized discussions with annotated triggers accurately reflect real-world human conversations where the intervention types add value without disrupting natural flow.

What would settle it

Placing the trained model into unannotated real human group discussions and measuring whether its chosen intervention points and responses match independent human ratings of timing and helpfulness.

Figures

Figures reproduced from arXiv: 2508.18167 by Christopher Malon, Deep Anil Patel, Iain Melvin, Martin Renqiang Min.

Figure 1
Figure 1. Figure 1: High level overview of our data generation pipeline. At each stage, an example output is [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The prompt used in Stage 1 to synthesize a structured scenario from a Yahoo! Answers [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The main prompt used in Stage 2 to generate a full discussion transcript from a scenario. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: An example of a final generated data point from the DiscussLLM dataset. The AI, Nexus, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Architectural overview of the unified End-to-End baseline. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Architectural overview of the Decoupled baseline. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
read the original abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities in understanding and generating human-like text, yet they largely operate as reactive agents, responding only when directly prompted. This passivity creates an "awareness gap," limiting their potential as truly collaborative partners in dynamic human discussions. We introduce $\textit{DiscussLLM}$, a framework designed to bridge this gap by training models to proactively decide not just $\textit{what}$ to say, but critically, $\textit{when}$ to speak. Our primary contribution is a scalable two-stage data generation pipeline that synthesizes a large-scale dataset of realistic multi-turn human discussions. Each discussion is annotated with one of five intervention types (e.g., Factual Correction, Concept Definition) and contains an explicit conversational trigger where an AI intervention adds value. By training models to predict a special silent token when no intervention is needed, they learn to remain quiet until a helpful contribution can be made. We explore two architectural baselines: an integrated end-to-end model and a decoupled classifier-generator system optimized for low-latency inference. We evaluate these models on their ability to accurately time interventions and generate helpful responses, paving the way for more situationally aware and proactive conversational AI.

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

2 major / 2 minor

Summary. The manuscript introduces DiscussLLM, a framework for training LLMs to decide not only what to say but when to speak in multi-turn human discussions. Its primary contribution is a scalable two-stage data generation pipeline that synthesizes large-scale annotated discussions, each labeled with one of five intervention types (e.g., Factual Correction, Concept Definition) and containing an explicit trigger where an AI intervention adds value. Models are trained to emit a special silent token until such a trigger occurs. Two architectural baselines are explored: an integrated end-to-end model and a decoupled classifier-generator system. The work claims to evaluate the models on intervention timing accuracy and response helpfulness.

Significance. If the synthetic triggers prove to be natural value-adding points that generalize beyond the generated distribution, this could meaningfully advance proactive, collaborative conversational agents and close the 'awareness gap' in current LLMs. The scalable pipeline itself is a constructive engineering contribution that could be reused in related dialogue tasks.

major comments (2)
  1. [Data Generation Pipeline] Data Generation section: The manuscript asserts that the two-stage pipeline produces 'realistic multi-turn human discussions' with triggers where intervention 'adds value,' yet provides no human ratings of naturalness, no comparison against existing dialogue corpora (e.g., MultiWOZ or DailyDialog), and no explicit mechanism for enforcing conversational flow. This assumption is load-bearing for the central claim that the learned timing policy will transfer to unscripted conversations.
  2. [Experiments] Evaluation section: The abstract states that the models are evaluated on 'ability to accurately time interventions and generate helpful responses,' but the manuscript supplies no quantitative metrics, baselines, error analysis, or ablation results. Without these, it is impossible to determine whether the end-to-end or decoupled approach supports the timing claims.
minor comments (2)
  1. [Abstract] Abstract and introduction: The five intervention types are listed but not illustrated with even one concrete example of a trigger and the corresponding helpful response; adding a short example would clarify the task.
  2. [Method] Notation: The special 'silent token' is introduced without specifying its exact token ID, training objective weight, or how it interacts with the standard next-token loss.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. We address each of the major comments below and describe the changes we will make in the revised version.

read point-by-point responses

  1. Referee: [Data Generation Pipeline] Data Generation section: The manuscript asserts that the two-stage pipeline produces 'realistic multi-turn human discussions' with triggers where intervention 'adds value,' yet provides no human ratings of naturalness, no comparison against existing dialogue corpora (e.g., MultiWOZ or DailyDialog), and no explicit mechanism for enforcing conversational flow. This assumption is load-bearing for the central claim that the learned timing policy will transfer to unscripted conversations.

    Authors: We agree that additional validation is needed to support the realism of the synthetic data. In the revised manuscript, we will include results from a human evaluation where participants rate the naturalness of the generated discussions and the value of the intervention points. We will also provide comparisons of dialogue statistics to DailyDialog and clarify the conversational flow enforcement through the LLM-based generation process in stage one. We acknowledge that proving transfer to completely unscripted conversations is challenging and will discuss this as a limitation. revision: yes

  2. Referee: [Experiments] Evaluation section: The abstract states that the models are evaluated on 'ability to accurately time interventions and generate helpful responses,' but the manuscript supplies no quantitative metrics, baselines, error analysis, or ablation results. Without these, it is impossible to determine whether the end-to-end or decoupled approach supports the timing claims.

    Authors: We thank the referee for highlighting this gap. The current manuscript primarily presents the framework and qualitative examples. In the revised version, we will add a comprehensive evaluation section with quantitative metrics for timing accuracy (e.g., precision and recall for intervention decisions), baselines including reactive LLMs, ablation results for the two architectures, and error analysis of failure cases. Human ratings for helpfulness will also be reported. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical data synthesis and training pipeline

full rationale

This is an applied empirical ML paper whose core contribution is a two-stage synthetic data generation pipeline for multi-turn discussions annotated with intervention triggers, followed by training and evaluation of end-to-end or decoupled models on timing and response quality. No closed-form derivations, equations, or first-principles results are present that could reduce to fitted inputs by construction. The pipeline and baselines are described as engineering choices for scalability and low-latency inference, with evaluation against the generated data distribution; these steps do not invoke self-citations as load-bearing uniqueness theorems or rename known results as novel predictions. The framework is therefore self-contained against external benchmarks of data realism and model performance.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unverified quality of the synthetic dataset and the assumption that the five intervention types correspond to valuable real-world moments; no free parameters or invented entities are specified in the abstract.

axioms (1)
  • domain assumption Synthesized multi-turn human discussions can accurately model real conversations and contain explicit triggers where AI interventions add value.
    The two-stage pipeline assumes generated data reflects realistic scenarios for the five intervention types.

pith-pipeline@v0.9.0 · 5746 in / 1281 out tokens · 63953 ms · 2026-05-21T22:04:34.023591+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

42 extracted references · 42 canonical work pages · 7 internal anchors

  1. [1]

    GPT-4 Technical Report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with ad- vanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    The Llama 3 Herd of Models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    The claude 3 model family: Opus, sonnet, haiku

    AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 1(1):4, 2024

    work page 2024

  5. [5]

    RoBERTa: A Robustly Optimized BERT Pretraining Approach

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1907

  6. [6]

    Mixed-initiative interaction

    James E Allen, Curry I Guinn, and Eric Horvtz. Mixed-initiative interaction. IEEE Intelligent Systems and their Applications, 14(5):14–23, 1999

    work page 1999

  7. [7]

    Principles of mixed-initiative user interfaces

    Eric Horvitz. Principles of mixed-initiative user interfaces. In Proceedings of the SIGCHI conference on Human Factors in Computing Systems, pages 159–166, 1999

    work page 1999

  8. [8]

    Reflections on challenges and promises of mixed-initiative interaction

    Eric J Horvitz. Reflections on challenges and promises of mixed-initiative interaction. AI Magazine, 28(2): 3–3, 2007

    work page 2007

  9. [9]

    Synergi: A mixed-initiative system for scholarly synthesis and sensemaking

    Hyeonsu B Kang, Tongshuang Wu, Joseph Chee Chang, and Aniket Kittur. Synergi: A mixed-initiative system for scholarly synthesis and sensemaking. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages 1–19, 2023

    work page 2023

  10. [10]

    Mixed-initiative interaction with computational generative systems

    Florian Lehmann. Mixed-initiative interaction with computational generative systems. In Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1–6, 2023

    work page 2023

  11. [11]

    Mixed-initiative systems for training and decision-aid applications

    Jaime R Carbonell and Allan M Collins. Mixed-initiative systems for training and decision-aid applications. Technical report, 1970

    work page 1970

  12. [12]

    Tunable llm-based proactive recommendation agent

    Mingze Wang, Chongming Gao, Wenjie Wang, Yangyang Li, and Fuli Feng. Tunable llm-based proactive recommendation agent. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 19262–19276, 2025

    work page 2025

  13. [13]

    Cod- inggenie: A proactive llm-powered programming assistant

    Sebastian Zhao, Alan Zhu, Hussein Mozannar, David Sontag, Ameet Talwalkar, and Valerie Chen. Cod- inggenie: A proactive llm-powered programming assistant. In Proceedings of the 33rd ACM International Conference on the Foundations of Software Engineering, pages 1168–1172, 2025

    work page 2025

  14. [14]

    Proactive agent: Shifting llm agents from reactive responses to active assistance

    Yaxi Lu, Shenzhi Yang, Cheng Qian, Guirong Chen, Qinyu Luo, Yesai Wu, Huadong Wang, Xin Cong, Zhong Zhang, Yankai Lin, et al. Proactive agent: Shifting llm agents from reactive responses to active assistance. arXiv preprint arXiv:2410.12361, 2024

    work page arXiv 2024

  15. [15]

    arXiv preprint arXiv:2305.02750 , year=

    Yang Deng, Wenqiang Lei, Wai Lam, and Tat-Seng Chua. A survey on proactive dialogue systems: Problems, methods, and prospects. arXiv preprint arXiv:2305.02750, 2023

    work page arXiv 2023

  16. [16]

    Proactive conversa- tional ai: A comprehensive survey of advancements and opportunities

    Yang Deng, Lizi Liao, Wenqiang Lei, Grace Hui Yang, Wai Lam, and Tat-Seng Chua. Proactive conversa- tional ai: A comprehensive survey of advancements and opportunities. ACM Transactions on Information Systems, 43(3):1–45, 2025

    work page 2025

  17. [17]

    Socialmind: Llm-based proactive ar social assistive system with human-like perception for in-situ live interactions

    Bufang Yang, Yunqi Guo, Lilin Xu, Zhenyu Yan, Hongkai Chen, Guoliang Xing, and Xiaofan Jiang. Socialmind: Llm-based proactive ar social assistive system with human-like perception for in-situ live interactions. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 9(1): 1–30, 2025

    work page 2025

  18. [18]

    Talk to the hand: an llm-powered chatbot with visual pointer as proactive companion for on-screen tasks

    Thanawit Prasongpongchai, Pat Pataranutaporn, Monchai Lertsutthiwong, and Pattie Maes. Talk to the hand: an llm-powered chatbot with visual pointer as proactive companion for on-screen tasks. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, pages 1–16, 2025

    work page 2025

  19. [19]

    Developing a proactive programming assistant leveraging an llm for personalized real-time feedback

    Jussi Impiö. Developing a proactive programming assistant leveraging an llm for personalized real-time feedback. 2025. 9

    work page 2025

  20. [20]

    Redefining proactivity for information seeking dialogue

    Jing Yang Lee, Seokhwan Kim, Kartik Mehta, Jiun-Yu Kao, Yu-Hsiang Lin, and Arpit Gupta. Redefining proactivity for information seeking dialogue. arXiv preprint arXiv:2410.15297, 2024

    work page arXiv 2024

  21. [21]

    Proactive conversational agents with inner thoughts

    Xingyu Bruce Liu, Shitao Fang, Weiyan Shi, Chien-Sheng Wu, Takeo Igarashi, and Xiang’Anthony’ Chen. Proactive conversational agents with inner thoughts. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, pages 1–19, 2025

    work page 2025

  22. [22]

    Videollm-online: Online video large language model for streaming video

    Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407–18418, 2024

    work page 2024

  23. [23]

    Streaming dense video captioning

    Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243–18252, 2024

    work page 2024

  24. [24]

    arXiv:2405.13003 [cs.CL]

    Heydar Soudani, Roxana Petcu, Evangelos Kanoulas, and Faegheh Hasibi. A survey on recent advances in conversational data generation. arXiv preprint arXiv:2405.13003, 2024

    work page internal anchor Pith review arXiv 2024

  25. [25]

    Synthetic dialogue data generation: A comprehensive survey of methods, evaluation, and challenges

    Anshul Chavda and Pushpak Bhattacharyya. Synthetic dialogue data generation: A comprehensive survey of methods, evaluation, and challenges

    work page

  26. [26]

    Diverse and effective synthetic data generation for adaptable zero-shot dialogue state tracking

    James D Finch and Jinho D Choi. Diverse and effective synthetic data generation for adaptable zero-shot dialogue state tracking. arXiv preprint arXiv:2405.12468, 2024

    work page arXiv 2024

  27. [27]

    A framework for synthetic audio conversations generation using large language models

    Kaung Myat Kyaw and Jonathan Hoyin Chan. A framework for synthetic audio conversations generation using large language models. In 2024 IEEE/WIC International Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), pages 355–359. IEEE, 2024

    work page 2024

  28. [28]

    Synthetic patient– physician conversations simulated by large language models: A multi-dimensional evaluation

    Syed Ali Haider, Srinivasagam Prabha, Cesar Abraham Gomez-Cabello, Sahar Borna, Ariana Genovese, Maissa Trabilsy, Bernardo G Collaco, Nadia G Wood, Sanjay Bagaria, Cui Tao, et al. Synthetic patient– physician conversations simulated by large language models: A multi-dimensional evaluation. Sensors, 25 (14):4305, 2025

    work page 2025

  29. [29]

    A synthetic data generation framework for grounded dialogues

    Jianzhu Bao, Rui Wang, Yasheng Wang, Aixin Sun, Yitong Li, Fei Mi, and Ruifeng Xu. A synthetic data generation framework for grounded dialogues. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10866–10882, 2023

    work page 2023

  30. [30]

    Omnichat: Enhancing spoken dialogue systems with scalable synthetic data for diverse scenarios

    Xize Cheng, Dongjie Fu, Xiaoda Yang, Minghui Fang, Ruofan Hu, Jingyu Lu, Bai Jionghao, Zehan Wang, Shengpeng Ji, Rongjie Huang, et al. Omnichat: Enhancing spoken dialogue systems with scalable synthetic data for diverse scenarios. arXiv preprint arXiv:2501.01384, 2025

    work page arXiv 2025

  31. [31]

    Streaming detection of queried event start

    Cristóbal Eyzaguirre, Eric Tang, Shyamal Buch, Adrien Gaidon, Jiajun Wu, and Juan C Niebles. Streaming detection of queried event start. Advances in Neural Information Processing Systems, 37:100698–100733, 2024

    work page 2024

  32. [32]

    Convogen: Enhancing conversational ai with synthetic data: A multi-agent approach

    Reem Gody, Mahmoud Goudy, and Ahmed Y Tawfik. Convogen: Enhancing conversational ai with synthetic data: A multi-agent approach. arXiv preprint arXiv:2503.17460, 2025

    work page arXiv 2025

  33. [33]

    Artificial conversations, real results: Fostering language detection with synthetic data

    Fatemeh Mohammadi, Tommaso Romano, Samira Maghool, and Paolo Ceravolo. Artificial conversations, real results: Fostering language detection with synthetic data. arXiv preprint arXiv:2503.24062, 2025

    work page arXiv 2025

  34. [34]

    Synthetic dialogue dataset generation using llm agents

    Yelaman Abdullin, Diego Molla-Aliod, Bahadorreza Ofoghi, John Yearwood, and Qingyang Li. Synthetic dialogue dataset generation using llm agents. arXiv preprint arXiv:2401.17461, 2024

    work page arXiv 2024

  35. [35]

    Scaling Synthetic Data Creation with 1,000,000,000 Personas

    Tao Ge, Xin Chan, Xiaoyang Wang, Dian Yu, Haitao Mi, and Dong Yu. Scaling synthetic data creation with 1,000,000,000 personas. arXiv preprint arXiv:2406.20094, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    LLaVA-Video: Video Instruction Tuning With Synthetic Data

    Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  37. [37]

    A large-scale sentiment analysis for yahoo! answers

    Onur Kucuktunc, B Barla Cambazoglu, Ingmar Weber, and Hakan Ferhatosmanoglu. A large-scale sentiment analysis for yahoo! answers. In Proceedings of the fifth ACM international conference on Web search and data mining, pages 633–642, 2012

    work page 2012

  38. [38]

    Llama 3 model card

    AI@Meta. Llama 3 model card. 2024. URL https://github.com/meta-llama/llama3/blob/ main/MODEL_CARD.md. 10

    work page 2024

  39. [39]

    Perplexity—a measure of the difficulty of speech recognition tasks

    Fred Jelinek, Robert L Mercer, Lalit R Bahl, and James K Baker. Perplexity—a measure of the difficulty of speech recognition tasks. The Journal of the Acoustical Society of America, 62(S1):S63–S63, 1977

    work page 1977

  40. [40]

    A neural probabilistic language model

    Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin. A neural probabilistic language model. Journal of machine learning research, 3(Feb):1137–1155, 2003

    work page 2003

  41. [41]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022

    work page 2022

  42. [42]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017. 11

    work page 2017