[{"quality_controlled":"1","citation":{"ista":"Nicolicioiu A, Iofinova EB, Jovanovic A, Kurtic E, Nikdan M, Panferov A, Markov I, Shavit N, Alistarh D-A. 2026. Panza: Investigating the feasibility of fully-local personalized text generation, OpenReview,p.","mla":"Nicolicioiu, Armand, et al. “Panza: Investigating the Feasibility of Fully-Local Personalized Text Generation.” <i>Third Conference on Parsimony and Learning (Proceedings Track)</i>, 81, OpenReview, 2026.","chicago":"Nicolicioiu, Armand, Eugenia B Iofinova, Andrej Jovanovic, Eldar Kurtic, Mahdi Nikdan, Andrei Panferov, Ilia Markov, Nir Shavit, and Dan-Adrian Alistarh. <i>Panza: Investigating the Feasibility of Fully-Local Personalized Text Generation</i>. <i>Third Conference on Parsimony and Learning (Proceedings Track)</i>. OpenReview, 2026.","short":"A. Nicolicioiu, E.B. Iofinova, A. Jovanovic, E. Kurtic, M. Nikdan, A. Panferov, I. Markov, N. Shavit, D.-A. Alistarh, Panza: Investigating the Feasibility of Fully-Local Personalized Text Generation, OpenReview, 2026.","apa":"Nicolicioiu, A., Iofinova, E. B., Jovanovic, A., Kurtic, E., Nikdan, M., Panferov, A., … Alistarh, D.-A. (2026). <i>Panza: Investigating the feasibility of fully-local personalized text generation</i>. <i>Third Conference on Parsimony and Learning (Proceedings Track)</i>. Tübíngen, Germany: OpenReview.","ieee":"A. Nicolicioiu <i>et al.</i>, <i>Panza: Investigating the feasibility of fully-local personalized text generation</i>. OpenReview, 2026.","ama":"Nicolicioiu A, Iofinova EB, Jovanovic A, et al. <i>Panza: Investigating the Feasibility of Fully-Local Personalized Text Generation</i>. OpenReview; 2026."},"oa_version":"Accepted Version","article_number":"81","corr_author":"1","title":"Panza: Investigating the feasibility of fully-local personalized text generation","publication":"Third Conference on Parsimony and Learning (Proceedings Track)","month":"03","date_published":"2026-03-06T00:00:00Z","related_material":{"record":[{"status":"public","relation":"dissertation_contains","id":"21854"}]},"abstract":[{"lang":"eng","text":"The availability of powerful open-source large language models (LLMs) opens exciting use cases, such as using personal data to fine-tune these models to imitate a user’s unique writing style. Two key requirements for this functionality are personalization–in the sense that the output should recognizably reflect the user’s own writing style—and privacy–users may justifiably be wary of uploading extremely personal data, such as their email archive, to a third-party service. In this paper, we demonstrate the feasibility of training and running such an assistant, which we call Panza, on commodity hardware, for the specific use case of email generation. Panza’s personalization features are based on a combination of parameter-efficient fine-tuning using a variant of the Reverse Instructions technique [1] and Retrieval-Augmented Generation (RAG) [2]. We demonstrate that this combination allows us to fine-tune an LLM to reflect a user’s writing style using limited data, while executing on extremely limited resources, e.g. on a free Google Colab instance. Our key methodological contribution is the first detailed study of evaluation metrics for this task, and\r\nof how different choices of system components–the use of RAG and of different fine-tuning approaches–impact the system’s performance. Additionally, we demonstrate that very little data - under 100 email samples - are sufficient to create models that convincingly imitate humans, showcasing a previously unknown attack vector in language models. We are releasing the full Panza code as well as three new email datasets licensed for research use."}],"department":[{"_id":"GradSch"},{"_id":"DaAl"}],"tmp":{"short":"CC BY (4.0)","legal_code_url":"https://creativecommons.org/licenses/by/4.0/legalcode","image":"/images/cc_by.png","name":"Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)"},"_id":"21857","publication_status":"published","language":[{"iso":"eng"}],"keyword":["LLMs","PEFT","LoRA","personalization","efficient ML"],"date_created":"2026-05-11T08:50:28Z","year":"2026","conference":{"start_date":"2026-03-23","location":"Tübíngen, Germany","name":"CPAL: Conference on Parsimony and Learning","end_date":"2026-03-26"},"oa":1,"status":"public","type":"conference_poster","user_id":"8b945eb4-e2f2-11eb-945a-df72226e66a9","publisher":"OpenReview","day":"06","date_updated":"2026-05-19T11:20:27Z","OA_type":"green","author":[{"full_name":"Nicolicioiu, Armand","last_name":"Nicolicioiu","first_name":"Armand"},{"id":"f9a17499-f6e0-11ea-865d-fdf9a3f77117","last_name":"Iofinova","first_name":"Eugenia B","orcid":"0000-0002-7778-3221","full_name":"Iofinova, Eugenia B"},{"full_name":"Jovanovic, Andrej","last_name":"Jovanovic","first_name":"Andrej"},{"last_name":"Kurtic","first_name":"Eldar","id":"47beb3a5-07b5-11eb-9b87-b108ec578218","full_name":"Kurtic, Eldar"},{"full_name":"Nikdan, Mahdi","last_name":"Nikdan","first_name":"Mahdi","id":"66374281-f394-11eb-9cf6-869147deecc0"},{"full_name":"Panferov, Andrei","last_name":"Panferov","first_name":"Andrei","id":"2c18daae-4dbe-11ef-8491-98ce2d960f09"},{"full_name":"Markov, Ilia","last_name":"Markov","first_name":"Ilia","id":"D0CF4148-C985-11E9-8066-0BDEE5697425"},{"full_name":"Shavit, Nir","last_name":"Shavit","first_name":"Nir"},{"full_name":"Alistarh, Dan-Adrian","id":"4A899BFC-F248-11E8-B48F-1D18A9856A87","last_name":"Alistarh","orcid":"0000-0003-3650-940X","first_name":"Dan-Adrian"}],"article_processing_charge":"No","OA_place":"publisher","main_file_link":[{"open_access":"1","url":"https://openreview.net/pdf?id=soFWnTqd23"}]},{"type":"conference","page":"12284-12303","status":"public","external_id":{"arxiv":["2401.06118"]},"conference":{"end_date":"2024-07-27","location":"Vienna, Austria","start_date":"2024-07-21","name":"ICML: International Conference on Machine Learning"},"oa":1,"year":"2024","scopus_import":"1","date_created":"2024-09-22T22:01:43Z","language":[{"iso":"eng"}],"alternative_title":["PMLR"],"volume":235,"main_file_link":[{"url":" https://doi.org/10.48550/arXiv.2401.06118","open_access":"1"}],"article_processing_charge":"No","author":[{"full_name":"Egiazarian, Vage","last_name":"Egiazarian","first_name":"Vage"},{"full_name":"Panferov, Andrei","first_name":"Andrei","last_name":"Panferov","id":"2c18daae-4dbe-11ef-8491-98ce2d960f09"},{"full_name":"Kuznedelev, Denis","last_name":"Kuznedelev","first_name":"Denis"},{"first_name":"Elias","last_name":"Frantar","id":"09a8f98d-ec99-11ea-ae11-c063a7b7fe5f","full_name":"Frantar, Elias"},{"full_name":"Babenko, Artem","last_name":"Babenko","first_name":"Artem"},{"id":"4A899BFC-F248-11E8-B48F-1D18A9856A87","first_name":"Dan-Adrian","orcid":"0000-0003-3650-940X","last_name":"Alistarh","full_name":"Alistarh, Dan-Adrian"}],"date_updated":"2024-10-01T08:13:05Z","day":"01","acknowledgement":"Authors would like to thank Ruslan Svirschevski for his help in solving technical issues with AQLM and baselines. We also thank Tim Dettmers for helpful discussions on the structure of weights in modern LLMs and size-accuracy trade-offs. The authors would also like to thank Daniil Pavlov for his assistance with CPU benchmarking. Finally, authors would like to thank the communities of ML enthusiasts known as LocalLLaMA5 and Petals community on discord6\r\nfor the crowd wisdom about running LLMs on consumer devices. Egiazarian Vage and Denis Kuznedelev and Andrei Panferov were supported by the grant for research centers in the field of AI provided by the Analytical Center for the Government of the Russian Federation (ACRF) in\r\naccordance with the agreement on the provision of subsidies (identifier of the agreement 000000D730321P5Q0002) and the agreement with HSE University No. 70-2021-00139.","publisher":"ML Research Press","user_id":"2DF688A6-F248-11E8-B48F-1D18A9856A87","publication":"Proceedings of the 41st International Conference on Machine Learning","title":"Extreme compression of large language models via additive quantization","arxiv":1,"corr_author":"1","oa_version":"Preprint","citation":{"chicago":"Egiazarian, Vage, Andrei Panferov, Denis Kuznedelev, Elias Frantar, Artem Babenko, and Dan-Adrian Alistarh. “Extreme Compression of Large Language Models via Additive Quantization.” In <i>Proceedings of the 41st International Conference on Machine Learning</i>, 235:12284–303. ML Research Press, 2024.","mla":"Egiazarian, Vage, et al. “Extreme Compression of Large Language Models via Additive Quantization.” <i>Proceedings of the 41st International Conference on Machine Learning</i>, vol. 235, ML Research Press, 2024, pp. 12284–303.","ista":"Egiazarian V, Panferov A, Kuznedelev D, Frantar E, Babenko A, Alistarh D-A. 2024. Extreme compression of large language models via additive quantization. Proceedings of the 41st International Conference on Machine Learning. ICML: International Conference on Machine Learning, PMLR, vol. 235, 12284–12303.","apa":"Egiazarian, V., Panferov, A., Kuznedelev, D., Frantar, E., Babenko, A., &#38; Alistarh, D.-A. (2024). Extreme compression of large language models via additive quantization. In <i>Proceedings of the 41st International Conference on Machine Learning</i> (Vol. 235, pp. 12284–12303). Vienna, Austria: ML Research Press.","ieee":"V. Egiazarian, A. Panferov, D. Kuznedelev, E. Frantar, A. Babenko, and D.-A. Alistarh, “Extreme compression of large language models via additive quantization,” in <i>Proceedings of the 41st International Conference on Machine Learning</i>, Vienna, Austria, 2024, vol. 235, pp. 12284–12303.","ama":"Egiazarian V, Panferov A, Kuznedelev D, Frantar E, Babenko A, Alistarh D-A. Extreme compression of large language models via additive quantization. In: <i>Proceedings of the 41st International Conference on Machine Learning</i>. Vol 235. ML Research Press; 2024:12284-12303.","short":"V. Egiazarian, A. Panferov, D. Kuznedelev, E. Frantar, A. Babenko, D.-A. Alistarh, in:, Proceedings of the 41st International Conference on Machine Learning, ML Research Press, 2024, pp. 12284–12303."},"quality_controlled":"1","publication_identifier":{"eissn":["2640-3498"]},"intvolume":"       235","publication_status":"published","_id":"18113","department":[{"_id":"DaAl"},{"_id":"GradSch"}],"abstract":[{"text":"The emergence of accurate open large language models (LLMs) has led to a race towards performant quantization techniques which can enable their execution on end-user devices. In this paper, we revisit the problem of “extreme” LLM compression—defined as targeting extremely low bit counts, such as 2 to 3 bits per parameter—from the point of view of classic methods in Multi-Codebook Quantization (MCQ). Our algorithm, called AQLM, generalizes the classic Additive Quantization (AQ) approach for information retrieval to advance the state-of-the-art in LLM compression, via two innovations: 1) learned additive quantization of weight matrices in input-adaptive fashion, and 2) joint optimization of codebook parameters across each transformer blocks. Broadly, AQLM is the first scheme that is Pareto optimal in terms of accuracy-vs-model-size when compressing to less than 3 bits per parameter, and significantly improves upon all known schemes in the extreme compression (2bit) regime. In addition, AQLM is practical: we provide fast GPU and CPU implementations of AQLM for token generation, which enable us to match or outperform optimized FP16 implementations for speed, while executing in a much smaller memory footprint.","lang":"eng"}],"date_published":"2024-09-01T00:00:00Z","month":"09"}]