{"id":"https://openalex.org/W7127303646","doi":"https://doi.org/10.48550/arxiv.2602.00161","title":"LLM Compression by Block Removal with Constrained Binary Optimization","display_name":"LLM Compression by Block Removal with Constrained Binary Optimization","publication_year":2026,"publication_date":"2026-01-29","ids":{"openalex":"https://openalex.org/W7127303646","doi":"https://doi.org/10.48550/arxiv.2602.00161"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2602.00161","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2602.00161","pdf_url":"https://arxiv.org/pdf/2602.00161","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by-sa","license_id":"https://openalex.org/licenses/cc-by-sa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2602.00161","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5032726536","display_name":"D. J. Jansen","orcid":"https://orcid.org/0000-0003-4220-5592","observed_orcids":["https://orcid.org/0000-0003-4220-5592"]},"institutions":[{"id":"https://openalex.org/I4210120949","display_name":"Multiverse Computing (Spain)","ror":"https://ror.org/02p1vs657","country_code":"ES","type":"company","lineage":["https://openalex.org/I4210120949"]}],"countries":["ES"],"is_corresponding":false,"raw_author_name":"Jansen, David","raw_affiliation_strings":["Multiverse Computing"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Multiverse Computing","institution_ids":["https://openalex.org/I4210120949"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5002260238","display_name":"Roman Rausch","orcid":"https://orcid.org/0000-0002-0427-3634","observed_orcids":["https://orcid.org/0000-0002-0427-3634"]},"institutions":[{"id":"https://openalex.org/I4210120949","display_name":"Multiverse Computing (Spain)","ror":"https://ror.org/02p1vs657","country_code":"ES","type":"company","lineage":["https://openalex.org/I4210120949"]}],"countries":["ES"],"is_corresponding":false,"raw_author_name":"Rausch, Roman","raw_affiliation_strings":["Multiverse Computing"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Multiverse Computing","institution_ids":["https://openalex.org/I4210120949"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5144412958","display_name":"Ali Hashemi","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4210120949","display_name":"Multiverse Computing (Spain)","ror":"https://ror.org/02p1vs657","country_code":"ES","type":"company","lineage":["https://openalex.org/I4210120949"]}],"countries":["ES"],"is_corresponding":false,"raw_author_name":"Hashemi, Ali","raw_affiliation_strings":["Multiverse Computing"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Multiverse Computing","institution_ids":["https://openalex.org/I4210120949"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5144425230","display_name":"David Montero","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Montero, David","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5020314133","display_name":"Rom\u00e1n Or\u00fas","orcid":"https://orcid.org/0000-0002-4496-8115","observed_orcids":["https://orcid.org/0000-0002-4496-8115"]},"institutions":[{"id":"https://openalex.org/I110594554","display_name":"Ikerbasque","ror":"https://ror.org/01cc3fy72","country_code":"ES","type":"other","lineage":["https://openalex.org/I110594554"]},{"id":"https://openalex.org/I47686490","display_name":"Donostia International Physics Center","ror":"https://ror.org/02e24yw40","country_code":"ES","type":"other","lineage":["https://openalex.org/I47686490"]}],"countries":["ES"],"is_corresponding":false,"raw_author_name":"Or\u00fas, Rom\u00e1n","raw_affiliation_strings":["Multiverse Computing Donostia International Physics Center Ikerbasque Foundation for Science"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Multiverse Computing Donostia International Physics Center Ikerbasque Foundation for Science","institution_ids":["https://openalex.org/I110594554","https://openalex.org/I47686490"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":5,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.9477999806404114,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.9477999806404114,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.032600000500679016,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T12535","display_name":"Machine Learning and Data Classification","score":0.003800000064074993,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/large-language-model-compression","display_name":"large language model compression","score":0.9810000061988831},{"id":"https://openalex.org/keywords/transformer-layers","display_name":"transformer layers","score":0.4779999852180481},{"id":"https://openalex.org/keywords/model-pruning","display_name":"model pruning","score":0.40700000524520874},{"id":"https://openalex.org/keywords/heuristic-solver","display_name":"heuristic solver","score":0.6190000176429749},{"id":"https://openalex.org/keywords/mixture-of-experts","display_name":"mixture of experts","score":0.6690000295639038},{"id":"https://openalex.org/keywords/llama-3","display_name":"Llama 3","score":0.5989999771118164},{"id":"https://openalex.org/keywords/qwen3","display_name":"Qwen3","score":0.6890000104904175},{"id":"https://openalex.org/keywords/mmlu","display_name":"MMLU","score":0.8029999732971191}],"concepts":[{"id":"https://openalex.org/C48372109","wikidata":"https://www.wikidata.org/wiki/Q3913","display_name":"Binary number","level":2,"score":0.6191999912261963},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.6190000176429749},{"id":"https://openalex.org/C2777210771","wikidata":"https://www.wikidata.org/wiki/Q4927124","display_name":"Block (permutation group theory)","level":2,"score":0.6154999732971191},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6057999730110168},{"id":"https://openalex.org/C2780767217","wikidata":"https://www.wikidata.org/wiki/Q5532421","display_name":"Generality","level":2,"score":0.5949000120162964},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.5266000032424927},{"id":"https://openalex.org/C2780586882","wikidata":"https://www.wikidata.org/wiki/Q7520643","display_name":"Simple (philosophy)","level":2,"score":0.43320000171661377},{"id":"https://openalex.org/C2993597952","wikidata":"https://www.wikidata.org/wiki/Q108123644","display_name":"Block model","level":2,"score":0.40610000491142273},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.39660000801086426},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3944000005722046},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.3653999865055084},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.3375999927520752},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.3116999864578247},{"id":"https://openalex.org/C126980161","wikidata":"https://www.wikidata.org/wiki/Q863783","display_name":"Simulated annealing","level":2,"score":0.2874000072479248},{"id":"https://openalex.org/C197855036","wikidata":"https://www.wikidata.org/wiki/Q380172","display_name":"Binary tree","level":2,"score":0.2849000096321106},{"id":"https://openalex.org/C55660270","wikidata":"https://www.wikidata.org/wiki/Q5164377","display_name":"Constrained optimization","level":2,"score":0.28220000863075256},{"id":"https://openalex.org/C151376022","wikidata":"https://www.wikidata.org/wiki/Q168698","display_name":"Exponential function","level":2,"score":0.2759000062942505},{"id":"https://openalex.org/C41431624","wikidata":"https://www.wikidata.org/wiki/Q1053357","display_name":"Block size","level":3,"score":0.2651999890804291},{"id":"https://openalex.org/C2776436953","wikidata":"https://www.wikidata.org/wiki/Q5163215","display_name":"Consistency (knowledge bases)","level":2,"score":0.26440000534057617},{"id":"https://openalex.org/C75235859","wikidata":"https://www.wikidata.org/wiki/Q582659","display_name":"Exponential growth","level":2,"score":0.26019999384880066},{"id":"https://openalex.org/C49937458","wikidata":"https://www.wikidata.org/wiki/Q2599292","display_name":"Probabilistic logic","level":2,"score":0.25459998846054077}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2602.00161","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2602.00161","pdf_url":"https://arxiv.org/pdf/2602.00161","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by-sa","license_id":"https://openalex.org/licenses/cc-by-sa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2602.00161","doi":"https://doi.org/10.48550/arxiv.2602.00161","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2602.00161","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2602.00161","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2602.00161","pdf_url":"https://arxiv.org/pdf/2602.00161","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by-sa","license_id":"https://openalex.org/licenses/cc-by-sa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320321705","display_name":"Eusko Jaurlaritza","ror":"https://ror.org/00pz2fp31"},{"id":"https://openalex.org/F4320324498","display_name":"Institute for Advanced Studies in Basic Sciences","ror":"https://ror.org/00bzsst90"},{"id":"https://openalex.org/F4320338448","display_name":"HORIZON EUROPE European Innovation Council","ror":"https://ror.org/05cx8cy07"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7127303646.pdf","grobid_xml":"https://content.openalex.org/works/W7127303646.grobid-xml"},"referenced_works_count":15,"referenced_works":["https://openalex.org/W1568345435","https://openalex.org/W2996908057","https://openalex.org/W3083410900","https://openalex.org/W4306808908","https://openalex.org/W4307934016","https://openalex.org/W4327810158","https://openalex.org/W4379260375","https://openalex.org/W4381586827","https://openalex.org/W4389650671","https://openalex.org/W4401306886","https://openalex.org/W4404990197","https://openalex.org/W4416043746","https://openalex.org/W4417024993","https://openalex.org/W7103754027","https://openalex.org/W7124165156"],"related_works":["https://openalex.org/W4412676021","https://openalex.org/W4406204726","https://openalex.org/W4403661492","https://openalex.org/W4406841407","https://openalex.org/W4409364387","https://openalex.org/W4405564330","https://openalex.org/W4415796567","https://openalex.org/W4387595586","https://openalex.org/W7124118818","https://openalex.org/W4396822602"],"abstract_inverted_index":{"In":[0],"this":[1,196],"paper,":[2],"we":[3,88,154,214],"formulate":[4],"the":[5,76,97,162,182,199],"compression":[6,78,84],"of":[7,52,56,85],"large":[8,54],"language":[9],"models":[10],"(LLMs)":[11],"by":[12],"optimally":[13],"deleting":[14],"transformer":[15],"blocks":[16],"(``block":[17],"removal'')":[18],"as":[19,81,128,130],"a":[20,31,39,53,145,149,205],"constrained":[21],"binary":[22],"optimization":[23],"(CBO)":[24],"problem":[25,164,183],"that":[26,156,167],"can":[27,187],"be":[28,188],"mapped":[29],"to":[30,101,180,191],"physical":[32],"system":[33],"(Ising":[34],"glass),":[35],"whose":[36],"energies":[37],"are":[38],"strong":[40],"proxy":[41],"for":[42,82,120,148,161,217],"downstream":[43,171],"model":[44],"performance.":[45],"This":[46],"formulation":[47],"enables":[48],"an":[49,90],"efficient":[50,136],"ranking":[51],"number":[55],"candidate":[57],"block-removal":[58,105],"configurations":[59],"yielding":[60],"many":[61],"high-quality,":[62],"non-trivial":[63],"solutions":[64,166],"beyond":[65],"those":[66,115],"only":[67,139],"removing":[68,222],"consecutive":[69],"regions.":[70],"Our":[71],"method":[72,186],"performs":[73,111],"strongly":[74],"in":[75,173],"deep":[77],"regime,":[79],"such":[80],"50%":[83],"Llama-3.3-70B-Instruct,":[86],"where":[87,213],"achieve":[89],"almost":[91],"23":[92],"percentage":[93],"point":[94],"increase":[95],"on":[96,112,144,170,198],"MMLU":[98],"benchmark":[99],"compared":[100],"other":[102],"state-of-the-art":[103],"(SOTA)":[104],"methods.":[106],"For":[107],"lighter":[108],"compression,":[109],"it":[110,177],"par":[113],"with":[114],"methods":[116],"across":[117],"several":[118],"benchmarks":[119],"Llama-3.1-8B-Instruct,":[121],"Qwen3-14B":[122],"(both":[123],"before":[124],"and":[125,137,141,208,212,219],"after":[126],"retraining),":[127],"well":[129,169],"Llama-3.3-70B-Instruct.":[131],"The":[132,185],"approach":[133],"is":[134,178],"computationally":[135],"requires":[138],"forward":[140],"backward":[142],"passes":[143],"calibration":[146],"dataset":[147],"few":[150],"active":[151],"parameters.":[152],"Additionally,":[153],"demonstrate":[155],"using":[157],"good":[158],"heuristic":[159],"solvers":[160],"CBO":[163],"provides":[165],"perform":[168],"tasks":[172],"negligible":[174],"runtime":[175],"when":[176,221],"unfeasible":[179],"solve":[181],"exactly.":[184],"readily":[189],"applied":[190],"any":[192],"architecture.":[193],"We":[194],"illustrate":[195],"generality":[197],"recent":[200],"NVIDIA-Nemotron-3-Nano-30B-A3B-FP8":[201],"model,":[202],"which":[203],"exhibits":[204],"highly":[206],"inhomogeneous":[207],"challenging":[209],"block":[210],"structure,":[211],"outperform":[215],"SOTA":[216],"AIME25":[218],"GPQA":[220],"either":[223],"2":[224],"attention":[225],"layers":[226],"or":[227],"3":[228],"mixture-of-experts":[229],"layers.":[230]},"counts_by_year":[{"year":2026,"cited_by_count":4},{"year":2025,"cited_by_count":1}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-02-04T00:00:00"}