{"id":1172220,"date":"2026-05-15T23:41:03","date_gmt":"2026-05-16T06:41:03","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-research-item&p=1172220"},"modified":"2026-05-15T23:41:04","modified_gmt":"2026-05-16T06:41:04","slug":"hobit-hardness-optimized-batch-sampling-for-infonce-training","status":"publish","type":"msr-research-item","link":"https:\/\/www.microsoft.com\/en-us\/research\/publication\/hobit-hardness-optimized-batch-sampling-for-infonce-training\/","title":{"rendered":"HOBIT: Hardness Optimized Batch Sampling for InfoNCE Training"},"content":{"rendered":"
\n
\n
\n
\n
\n

Contrastive training with\u00a0<\/span>InfoNCE<\/span>\u00a0loss and in-batch negatives is the\u00a0<\/span>standard <\/span>approach for learning dual-encoder models. Its effectiveness,\u00a0<\/span>however, <\/span>critically depends on the availability of hard negatives; in their\u00a0<\/span>absence, <\/span>learning quickly saturates. Existing methods address this via\u00a0<\/span>explicit hard-<\/span>negative mining, which is often costly or heuristic-driven. We\u00a0<\/span>introduce <\/span>HOBIT<\/span><\/b>, a principled mini-batch construction method that\u00a0<\/span>improves in-batch <\/span>negative quality by reordering training examples at every\u00a0<\/span>epoch. <\/span>HOBIT\u00a0solves an optimization problem motivated by the\u00a0<\/span>InfoNCE<\/span>\u00a0<\/span>objective <\/span>to yield mini-batches such that each query in the batch is exposed\u00a0<\/span>to hard <\/span>yet non-contradictory, informative negative examples. We show that\u00a0<\/span>the <\/span>optimization objective is monotone and submodular which in turn leads\u00a0<\/span>us <\/span>to a greedy algorithm that admits the standard\u00a0<\/span>O(1-1\/e)<\/span><\/i> approximation\u00a0<\/span>guarantee.<\/span>\u200b Empirically, we show that HOBIT incurs negligible computational overhead while significantly outperforming state-of-the-art batching methods, and remains complementary to existing hard negative mining techniques.<\/span><\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

Contrastive training with\u00a0InfoNCE\u00a0loss and in-batch negatives is the\u00a0standard approach for learning dual-encoder models. Its effectiveness,\u00a0however, critically depends on the availability of hard negatives; in their\u00a0absence, learning quickly saturates. Existing methods address this via\u00a0explicit hard-negative mining, which is often costly or heuristic-driven. We\u00a0introduce HOBIT, a principled mini-batch construction method that\u00a0improves in-batch negative quality by reordering training […]<\/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":null,"msr_publishername":"","msr_publisher_other":"","msr_booktitle":"","msr_chapter":"","msr_edition":"","msr_editors":"","msr_how_published":"","msr_isbn":"","msr_issue":"","msr_journal":"","msr_number":"","msr_organization":"","msr_pages_string":"","msr_page_range_start":"","msr_page_range_end":"","msr_series":"","msr_volume":"","msr_copyright":"","msr_conference_name":"","msr_doi":"","msr_arxiv_id":"","msr_s2_paper_id":"","msr_mag_id":"","msr_pubmed_id":"","msr_other_authors":"","msr_other_contributors":"","msr_speaker":"","msr_award":"","msr_affiliation":"","msr_institution":"","msr_host":"","msr_version":"","msr_duration":"","msr_original_fields_of_study":"","msr_release_tracker_id":"","msr_s2_match_type":"","msr_citation_count_updated":"","msr_published_date":"2026-07-01","msr_highlight_text":"Spotlight Acceptance","msr_notes":"","msr_longbiography":"","msr_publicationurl":"","msr_external_url":"","msr_secondary_video_url":"","msr_conference_url":"https:\/\/icml.cc\/","msr_journal_url":"","msr_s2_pdf_url":"","msr_year":0,"msr_citation_count":0,"msr_influential_citations":0,"msr_reference_count":0,"msr_s2_match_confidence":0,"msr_microsoftintellectualproperty":false,"msr_s2_open_access":false,"msr_s2_author_ids":[],"msr_pub_ids":[],"msr_hide_image_in_river":null,"footnotes":""},"msr-research-highlight":[246574],"research-area":[13556,13555],"msr-publication-type":[193716],"msr-publisher":[],"msr-focus-area":[],"msr-locale":[268875],"msr-post-option":[269148,269142],"msr-field-of-study":[249610,248503,268410],"msr-conference":[260284],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-1172220","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-highlight-award","msr-research-area-artificial-intelligence","msr-research-area-search-information-retrieval","msr-locale-en_us","msr-post-option-approved-for-river","msr-post-option-include-in-river","msr-field-of-study-dense-retrieval","msr-field-of-study-information-retrieval","msr-field-of-study-representation-learning"],"msr_publishername":"","msr_edition":"","msr_affiliation":"","msr_published_date":"2026-07-01","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"","msr_volume":"","msr_number":"","msr_editors":"","msr_series":"","msr_issue":"","msr_organization":"","msr_how_published":"","msr_notes":"","msr_highlight_text":"Spotlight Acceptance","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":"","msr_publication_uploader":[{"type":"url","viewUrl":"false","id":"false","title":"https:\/\/icml.cc\/virtual\/2026\/poster\/64085","label_id":"243109","label":0}],"msr_related_uploader":"","msr_citation_count":0,"msr_citation_count_updated":"","msr_s2_paper_id":"","msr_influential_citations":0,"msr_reference_count":0,"msr_arxiv_id":"","msr_s2_author_ids":[],"msr_s2_open_access":false,"msr_s2_pdf_url":null,"msr_attachments":[],"msr-author-ordering":[{"type":"user_nicename","value":"Himanshu Dutta","user_id":44162,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Himanshu Dutta"},{"type":"text","value":"Lokesh Nagalapatti","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Yashoteja Prabhu","user_id":41203,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Yashoteja Prabhu"}],"msr_impact_theme":[],"msr_research_lab":[199562],"msr_event":[],"msr_group":[],"msr_project":[],"publication":[],"video":[],"msr-tool":[],"msr_publication_type":"inproceedings","related_content":[],"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1172220","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":3,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1172220\/revisions"}],"predecessor-version":[{"id":1172223,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1172220\/revisions\/1172223"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1172220"}],"wp:term":[{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=1172220"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1172220"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=1172220"},{"taxonomy":"msr-publisher","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publisher?post=1172220"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=1172220"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1172220"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1172220"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=1172220"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=1172220"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=1172220"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=1172220"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=1172220"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}