Speculative decoding has emerged as a promising approach to accelerating large language model (LLM) generation using a fast drafter while maintaining alignment with the target model’s distribution. However, existing approaches face a trade-off: external drafters offer flexibility but can suffer from slower drafting, while self-speculation methods use drafters tailored to the target model but require re-training. In this paper, we introduce novel drafters based on Mamba, a state-of-the-art state space model (SSM), as a solution that combines the best aspects of both approaches. By leveraging the linear structure of SSMs, our approach avoids the quadratic complexity inherent in traditional Transformer-based methods, enabling faster drafting and lower memory usage while maintaining the flexibility to work across different target models. We further enhance efficiency with a novel test-time tree search algorithm for generating high-quality draft candidates. Our empirical evaluation demonstrates that Mamba-based drafters not only outperform existing external drafting methods but are also comparable to state-of-the-art self-speculation approaches while using less memory and maintaining their cross-model adaptability.<\/p>\n","protected":false},"excerpt":{"rendered":"
Speculative decoding has emerged as a promising approach to accelerating large language model (LLM) generation using a fast drafter while maintaining alignment with the target model’s distribution. However, existing approaches face a trade-off: external drafters offer flexibility but can suffer from slower drafting, while self-speculation methods use drafters tailored to the target model but require […]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-author-ordering":[{"type":"text","value":"Daewon Choi","user_id":0},{"type":"text","value":"Seunghyuk Oh","user_id":0},{"type":"text","value":"Saket Dingliwal","user_id":0},{"type":"user_nicename","value":"Jihoon Tack","user_id":"44058"},{"type":"text","value":"Kyuyoung Kim","user_id":0},{"type":"text","value":"Woomin Song","user_id":0},{"type":"text","value":"Seojin Kim","user_id":0},{"type":"text","value":"Insu Han","user_id":0},{"type":"text","value":"Jinwoo Shin","user_id":0},{"type":"text","value":"A. Galstyan","user_id":0},{"type":"text","value":"Shubham Katiyar","user_id":0},{"type":"text","value":"S. Bodapati","user_id":0}],"msr_publishername":"","msr_publisher_other":"","msr_booktitle":"","msr_chapter":"","msr_edition":"Conference on Empirical Methods in Natural Language Processing","msr_editors":"","msr_how_published":"","msr_isbn":"","msr_issue":"","msr_journal":"ArXiv","msr_number":"","msr_organization":"","msr_pages_string":"","msr_page_range_start":"","msr_page_range_end":"","msr_series":"","msr_volume":"abs\/2506.01206","msr_copyright":"","msr_conference_name":"Conference on Empirical Methods in Natural Language Processing","msr_doi":"10.48550\/arXiv.2506.01206","msr_arxiv_id":"2506.01206","msr_mag_id":"","msr_other_authors":"","msr_other_contributors":"","msr_speaker":"","msr_award":"","msr_affiliation":"","msr_institution":"","msr_host":"","msr_version":"","msr_duration":"","msr_release_tracker_id":"","msr_highlight_type":"","msr_date_display_format":"","msr_main_download_label":"","msr_external_link_label":"","msr_doi_label":"","msr_published_date":"2025-06-01","msr_startdate":"","msr_presentation_date":"","msr_highlight_text":"","msr_notes":"","msr_longbiography":"","msr_publicationurl":"","msr_external_url":"","msr_secondary_video_url":"","msr_conference_url":"","msr_journal_url":"","msr_year":2025,"msr_month":6,"msr_day":1,"msr_microsoftintellectualproperty":false,"msr_pub_id":"458679121f151b721b73e07441d32b02ae481b0e","msr_publication_uploader":[{"type":"url","title":"https:\/\/dblp.org\/rec\/conf\/emnlp\/ChoiODTKSKHSGKB25.html","label_id":243109,"id":false,"viewUrl":false},{"type":"url","title":"https:\/\/arxiv.org\/abs\/2506.01206","label_id":243109,"id":false,"viewUrl":false}],"msr_related_uploader":[],"msr_original_fields_of_study":[],"msr_s2_paper_id":"458679121f151b721b73e07441d32b02ae481b0e","msr_s2_pdf_url":"","msr_citation_count_updated":"","msr_citation_count":3,"msr_influential_citations":0,"msr_reference_count":41,"msr_s2_open_access":false,"msr_s2_author_ids":[],"msr_pub_ids":[{"provider":"s2","id":"458679121f151b721b73e07441d32b02ae481b0e"},{"provider":"doi","id":"10.48550\/arXiv.2506.01206"},{"provider":"arxiv","id":"2506.01206"},{"provider":"dblp","id":"conf\/emnlp\/ChoiODTKSKHSGKB25"},{"provider":"corpusid","id":"279075136"}],"msr_hide_image_in_river":0,"footnotes":""},"msr-research-highlight":[],"research-area":[13556],"msr-publication-type":[193715,193716],"msr-publisher":[],"msr-publication-cta":[],"msr-focus-area":[],"msr-locale":[268875],"msr-post-option":[],"msr-field-of-study":[246691],"msr-conference":[270479],"msr-journal":[268188],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-1184498","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-area-artificial-intelligence","msr-locale-en_us","msr-field-of-study-computer-science"],"msr_publishername":"","msr_edition":"Conference on Empirical Methods in Natural Language Processing","msr_affiliation":"","msr_published_date":"2025-06-01","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"ArXiv","msr_volume":"abs\/2506.01206","msr_number":"","msr_editors":"","msr_series":"","msr_issue":"","msr_organization":"","msr_how_published":"","msr_notes":"","msr_highlight_text":"","msr_release_tracker_id":"","msr_original_fields_of_study":"","msr_download_urls":"","msr_external_url":"","msr_secondary_video_url":"","msr_longbiography":"","msr_microsoftintellectualproperty":0,"msr_main_download":"","msr_publicationurl":"","msr_doi":"10.48550\/arXiv.2506.01206","msr_publication_uploader":[{"type":"url","title":"https:\/\/dblp.org\/rec\/conf\/emnlp\/ChoiODTKSKHSGKB25.html","label_id":243109,"id":false,"viewUrl":false},{"type":"url","title":"https:\/\/arxiv.org\/abs\/2506.01206","label_id":243109,"id":false,"viewUrl":false}],"msr_related_uploader":[],"msr_citation_count":3,"msr_citation_count_updated":"","msr_s2_paper_id":"458679121f151b721b73e07441d32b02ae481b0e","msr_influential_citations":0,"msr_reference_count":41,"msr_arxiv_id":"2506.01206","msr_s2_author_ids":[],"msr_s2_open_access":false,"msr_s2_pdf_url":null,"msr_attachments":[],"msr-author-ordering":[{"type":"text","value":"Daewon Choi","user_id":0,"rest_url":false},{"type":"text","value":"Seunghyuk Oh","user_id":0,"rest_url":false},{"type":"text","value":"Saket Dingliwal","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Jihoon Tack","user_id":44058,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Jihoon Tack"},{"type":"text","value":"Kyuyoung Kim","user_id":0,"rest_url":false},{"type":"text","value":"Woomin Song","user_id":0,"rest_url":false},{"type":"text","value":"Seojin Kim","user_id":0,"rest_url":false},{"type":"text","value":"Insu Han","user_id":0,"rest_url":false},{"type":"text","value":"Jinwoo Shin","user_id":0,"rest_url":false},{"type":"text","value":"A. Galstyan","user_id":0,"rest_url":false},{"type":"text","value":"Shubham Katiyar","user_id":0,"rest_url":false},{"type":"text","value":"S. Bodapati","user_id":0,"rest_url":false}],"msr_impact_theme":[],"msr_research_lab":[],"msr_event":[],"msr_group":[],"msr_project":[],"publication":[],"video":[],"msr-tool":[],"msr_publication_type":"article","related_content":[],"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1184498","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-research-item"}],"version-history":[{"count":2,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1184498\/revisions"}],"predecessor-version":[{"id":1184744,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1184498\/revisions\/1184744"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1184498"}],"wp:term":[{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=1184498"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1184498"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=1184498"},{"taxonomy":"msr-publisher","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publisher?post=1184498"},{"taxonomy":"msr-publication-cta","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-cta?post=1184498"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=1184498"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1184498"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1184498"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=1184498"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=1184498"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=1184498"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=1184498"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=1184498"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}