{"id":"https://openalex.org/W7224010208","doi":"https://doi.org/10.48550/arxiv.2610.34886","title":"ThinkBit: Search Codewords, Recalibrate Scales, Select by Reasoning","display_name":"ThinkBit: Search Codewords, Recalibrate Scales, Select by Reasoning","publication_year":2026,"publication_date":"2026-10-26","ids":{"openalex":"https://openalex.org/W7224010208","doi":"https://doi.org/10.48550/arxiv.2610.34886"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2610.34886","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.34886","pdf_url":"https://arxiv.org/pdf/2610.34886","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2610.34886","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5027477000","display_name":"Rhea Sanjay Sukthanker","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I1290206253","display_name":"Microsoft (United States)","ror":"https://ror.org/00d0nc645","country_code":"US","type":"company","lineage":["https://openalex.org/I1290206253"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Sukthanker, Rhea Sanjay","raw_affiliation_strings":["Microsoft Applied Sciences Group"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Microsoft Applied Sciences Group","institution_ids":["https://openalex.org/I1290206253"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5080222030","display_name":"Pashmina Cameron","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I1290206253","display_name":"Microsoft (United States)","ror":"https://ror.org/00d0nc645","country_code":"US","type":"company","lineage":["https://openalex.org/I1290206253"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Cameron, Pashmina","raw_affiliation_strings":["Microsoft Applied Sciences Group"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Microsoft Applied Sciences Group","institution_ids":["https://openalex.org/I1290206253"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I1290206253"],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":0,"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":0.9496999979019165,"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":0.9496999979019165,"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":0.0066999997943639755,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.003700000001117587,"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/knowledge-distillation","display_name":"knowledge distillation","score":0.7820000052452087},{"id":"https://openalex.org/keywords/quantization-aware-training","display_name":"quantization-aware training","score":0.984000027179718},{"id":"https://openalex.org/keywords/reasoning-language-models","display_name":"reasoning language models","score":0.6549999713897705},{"id":"https://openalex.org/keywords/math-500","display_name":"MATH-500","score":0.6859999895095825},{"id":"https://openalex.org/keywords/model-compression","display_name":"model compression","score":0.5839999914169312},{"id":"https://openalex.org/keywords/qwen3","display_name":"Qwen3","score":0.5230000019073486}],"concepts":[{"id":"https://openalex.org/C100279451","wikidata":"https://www.wikidata.org/wiki/Q372193","display_name":"Perplexity","level":3,"score":0.7181000113487244},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7027000188827515},{"id":"https://openalex.org/C176217482","wikidata":"https://www.wikidata.org/wiki/Q860554","display_name":"Metric (unit)","level":2,"score":0.6010000109672546},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5587999820709229},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.5149000287055969},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.511900007724762},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.41339999437332153},{"id":"https://openalex.org/C153207627","wikidata":"https://www.wikidata.org/wiki/Q863873","display_name":"Code word","level":3,"score":0.4124000072479248},{"id":"https://openalex.org/C42058472","wikidata":"https://www.wikidata.org/wiki/Q810214","display_name":"Base (topology)","level":2,"score":0.4117000102996826},{"id":"https://openalex.org/C2780898871","wikidata":"https://www.wikidata.org/wiki/Q860554","display_name":"Performance metric","level":2,"score":0.4068000018596649},{"id":"https://openalex.org/C77618280","wikidata":"https://www.wikidata.org/wiki/Q1155772","display_name":"Scheme (mathematics)","level":2,"score":0.39309999346733093},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.3508000075817108},{"id":"https://openalex.org/C21200559","wikidata":"https://www.wikidata.org/wiki/Q7451068","display_name":"Sensitivity (control systems)","level":2,"score":0.3497999906539917},{"id":"https://openalex.org/C2777210771","wikidata":"https://www.wikidata.org/wiki/Q4927124","display_name":"Block (permutation group theory)","level":2,"score":0.34779998660087585},{"id":"https://openalex.org/C2778755073","wikidata":"https://www.wikidata.org/wiki/Q10858537","display_name":"Scale (ratio)","level":2,"score":0.3287000060081482},{"id":"https://openalex.org/C97137487","wikidata":"https://www.wikidata.org/wiki/Q729138","display_name":"Integer (computer science)","level":2,"score":0.32850000262260437},{"id":"https://openalex.org/C57691317","wikidata":"https://www.wikidata.org/wiki/Q1289248","display_name":"Scalar (mathematics)","level":2,"score":0.32089999318122864},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2865000069141388},{"id":"https://openalex.org/C149441793","wikidata":"https://www.wikidata.org/wiki/Q200726","display_name":"Probability distribution","level":2,"score":0.28290000557899475},{"id":"https://openalex.org/C90805587","wikidata":"https://www.wikidata.org/wiki/Q10944557","display_name":"Word (group theory)","level":2,"score":0.2800999879837036},{"id":"https://openalex.org/C50817715","wikidata":"https://www.wikidata.org/wiki/Q79895177","display_name":"Regret","level":2,"score":0.2718000113964081},{"id":"https://openalex.org/C2779227376","wikidata":"https://www.wikidata.org/wiki/Q6505497","display_name":"Layer (electronics)","level":2,"score":0.27079999446868896},{"id":"https://openalex.org/C159032336","wikidata":"https://www.wikidata.org/wiki/Q2488768","display_name":"Non-monotonic logic","level":2,"score":0.26919999718666077},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.26589998602867126},{"id":"https://openalex.org/C149629883","wikidata":"https://www.wikidata.org/wiki/Q660926","display_name":"Fraction (chemistry)","level":2,"score":0.25099998712539673}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2610.34886","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.34886","pdf_url":"https://arxiv.org/pdf/2610.34886","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2610.34886","doi":"https://doi.org/10.48550/arxiv.2610.34886","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2610.34886","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":"cc-by-sa","license_id":"https://openalex.org/licenses/cc-by-sa","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2610.34886","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.34886","pdf_url":"https://arxiv.org/pdf/2610.34886","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7224010208.pdf","grobid_xml":"https://content.openalex.org/works/W7224010208.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7163155315","https://openalex.org/W7167595835","https://openalex.org/W7223097342","https://openalex.org/W7125390417","https://openalex.org/W7164006465","https://openalex.org/W7165942897","https://openalex.org/W7214092670","https://openalex.org/W7223241439","https://openalex.org/W7160904519","https://openalex.org/W7172447540"],"abstract_inverted_index":{"Reasoning":[0],"language":[1],"models":[2],"generate":[3],"thousands":[4],"of":[5,24,44,87,238],"tokens":[6],"per":[7],"answer,":[8],"and":[9,39,195,220,252],"generation":[10],"speed":[11],"is":[12,75,112,231],"limited":[13],"by":[14,37,84,114,127,176],"memory":[15],"bandwidth,":[16],"so":[17],"storing":[18],"weights":[19],"in":[20,71],"two":[21],"bits":[22],"instead":[23],"bfloat16":[25,81,270],"can":[26,52],"potentially":[27],"make":[28],"inference":[29],"8$\\times$":[30],"cheaper.":[31],"Two-bit":[32],"methods":[33],"are":[34],"usually":[35],"judged":[36],"perplexity":[38],"multiple-choice":[40],"question-answering":[41],"accuracy,":[42],"both":[43],"which":[45,72],"say":[46],"little":[47],"about":[48],"whether":[49],"a":[50,62,80,85,92,99,115,130,151,161,180,256],"model":[51,170,240,259],"still":[53],"reason":[54],"through":[55],"more":[56,242],"difficult":[57],"problems.":[58],"We":[59,89,158,246],"present":[60],"ThinkBit,":[61],"strictly":[63],"scalar":[64,118],"two-bit":[65],"framework":[66,250],"built":[67],"to":[68,95,105,254],"preserve":[69],"reasoning,":[70],"every":[73,234],"weight":[74],"an":[76],"integer":[77],"codeword":[78],"times":[79],"scale":[82],"shared":[83],"group":[86],"weights.":[88],"first":[90],"use":[91],"sensitivity":[93],"metric":[94],"assign":[96],"each":[97,141],"layer":[98],"ternary,":[100],"2-bit,":[101],"or":[102],"3-bit":[103],"tier":[104],"allocate":[106],"bitwidths":[107],"optimally":[108],"among":[109],"layers.":[110],"This":[111],"followed":[113],"layerwise":[116,131],"Gumbel-softmax-based":[117],"quantization":[119],"algorithm":[120],"that":[121,167,248],"searches":[122],"for":[123,213,266],"optimal":[124],"codewords":[125,148],"block":[126],"block.":[128],"Since":[129],"scheme":[132],"does":[133],"not":[134],"capture":[135],"how":[136],"quantized":[137],"layers":[138],"interact":[139],"with":[140,146],"other,":[142],"we":[143],"recalibrate":[144],"scales,":[145],"the":[147,155,169,172,188,192,196,217,239,267],"frozen,":[149],"using":[150],"knowledge-distillation-based":[152],"loss":[153],"against":[154,211,264],"teacher":[156],"model.":[157,271],"then":[159],"perform":[160],"short":[162],"quantization-aware":[163],"training":[164,185],"(QAT)":[165],"stage":[166,194],"lets":[168],"mimic":[171],"base":[173],"model's":[174],"distribution":[175],"reassigning":[177],"codewords.":[178],"Finally,":[179],"paired":[181],"reasoning":[182,225],"test":[183],"on":[184,209,216,224,244,262],"problems":[186],"selects":[187],"final":[189],"checkpoint":[190,206],"between":[191],"recalibration":[193],"QAT":[197,223],"stage.":[198],"On":[199],"Qwen3-1.7B":[200],"at":[201,227],"matched":[202],"size,":[203],"our":[204],"recalibrated":[205],"scores":[207,260],"46.7":[208],"MATH-500":[210],"30.0":[212],"EfficientQAT":[214],"trained":[215],"same":[218],"data,":[219],"73.0":[221],"after":[222],"traces;":[226],"0.52":[228],"GB":[229,258,269],"it":[230],"smaller":[232],"than":[233],"published":[235],"2-bit":[236],"GGUF":[237],"yet":[241],"accurate":[243],"MATH-500.":[245],"show":[247],"this":[249],"scales":[251],"up":[253],"14B,":[255],"3.74":[257],"94.8":[261],"MATH-500,":[263],"97.5":[265],"29.5":[268]},"counts_by_year":[],"updated_date":"2026-10-30T12:39:25.321706","created_date":"2026-10-28T00:00:00"}