{"id":"https://openalex.org/W7163155315","doi":"https://doi.org/10.48550/arxiv.2606.02011","title":"Extreme Low-Bit Inference in Reasoning Models: Failure Modes and Targeted Recovery","display_name":"Extreme Low-Bit Inference in Reasoning Models: Failure Modes and Targeted Recovery","publication_year":2026,"publication_date":"2026-06-01","ids":{"openalex":"https://openalex.org/W7163155315","doi":"https://doi.org/10.48550/arxiv.2606.02011"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.02011","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.02011","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,"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","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2606.02011","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5135925354","display_name":"Ekaterina Alimaskina","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Alimaskina, Ekaterina","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137675804","display_name":"Darya Rudas","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Rudas, Darya","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137701843","display_name":"Denis Shveykin","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shveykin, Denis","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135921521","display_name":"Gleb Molodtsov","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Molodtsov, Gleb","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137674931","display_name":"Pavel Vasiliev","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Vasiliev, Pavel","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137622845","display_name":"Aleksandr Beznosikov","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Beznosikov, Aleksandr","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"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/T10502","display_name":"Advanced Memory and Neural Computing","score":0.2574999928474426,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10502","display_name":"Advanced Memory and Neural Computing","score":0.2574999928474426,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T12808","display_name":"Ferroelectric and Negative Capacitance Devices","score":0.13689999282360077,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.0625,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.8493000268936157},{"id":"https://openalex.org/keywords/speedup","display_name":"Speedup","score":0.6328999996185303},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.6248000264167786},{"id":"https://openalex.org/keywords/quantization","display_name":"Quantization (signal processing)","score":0.5906000137329102},{"id":"https://openalex.org/keywords/decoding-methods","display_name":"Decoding methods","score":0.5781000256538391},{"id":"https://openalex.org/keywords/model-based-reasoning","display_name":"Model-based reasoning","score":0.5131000280380249},{"id":"https://openalex.org/keywords/opportunistic-reasoning","display_name":"Opportunistic reasoning","score":0.44269999861717224},{"id":"https://openalex.org/keywords/process","display_name":"Process (computing)","score":0.4316999912261963}],"concepts":[{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.8493000268936157},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6880000233650208},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.6328999996185303},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6248000264167786},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.5906000137329102},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.5781000256538391},{"id":"https://openalex.org/C37335422","wikidata":"https://www.wikidata.org/wiki/Q6888134","display_name":"Model-based reasoning","level":3,"score":0.5131000280380249},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5130000114440918},{"id":"https://openalex.org/C86827895","wikidata":"https://www.wikidata.org/wiki/Q7098582","display_name":"Opportunistic reasoning","level":4,"score":0.44269999861717224},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.4316999912261963},{"id":"https://openalex.org/C9616225","wikidata":"https://www.wikidata.org/wiki/Q3929429","display_name":"Semantic reasoner","level":2,"score":0.42010000348091125},{"id":"https://openalex.org/C193221554","wikidata":"https://www.wikidata.org/wiki/Q5153664","display_name":"Commonsense reasoning","level":2,"score":0.40299999713897705},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.37619999051094055},{"id":"https://openalex.org/C115086926","wikidata":"https://www.wikidata.org/wiki/Q17004651","display_name":"Causal reasoning","level":3,"score":0.36390000581741333},{"id":"https://openalex.org/C89288958","wikidata":"https://www.wikidata.org/wiki/Q7301504","display_name":"Reasoning system","level":2,"score":0.34200000762939453},{"id":"https://openalex.org/C20162079","wikidata":"https://www.wikidata.org/wiki/Q1151406","display_name":"Case-based reasoning","level":2,"score":0.33799999952316284},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.335999995470047},{"id":"https://openalex.org/C83725634","wikidata":"https://www.wikidata.org/wiki/Q7268699","display_name":"Qualitative reasoning","level":2,"score":0.26809999346733093},{"id":"https://openalex.org/C174839445","wikidata":"https://www.wikidata.org/wiki/Q1134386","display_name":"Lock (firearm)","level":2,"score":0.2615000009536743},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.2605000138282776}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.02011","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.02011","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,"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","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2606.02011","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.02011","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,"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","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"Reasoning":[1],"Models":[2],"(LRMs)":[3],"rely":[4],"on":[5],"long":[6],"reasoning":[7,63,68,72,158],"traces,":[8],"making":[9],"inference":[10,24,178],"expensive.":[11],"While":[12],"low-bit":[13,157],"quantization":[14,48],"reduces":[15],"per-token":[16],"decoding":[17],"cost,":[18],"we":[19,94],"show":[20,80,154],"that":[21,81,155],"aggressive":[22],"2-bit":[23,47,104,177],"can":[25,179],"fail":[26],"to":[27,87,120,127,138,149],"deliver":[28],"end-to-end":[29,185],"speedup":[30],"because":[31],"instability":[32],"in":[33],"the":[34,103],"generation":[35,168],"process":[36],"inflates":[37],"total":[38],"token":[39],"count.":[40],"Instead":[41],"of":[42,70],"merely":[43],"lowering":[44],"answer":[45,123],"accuracy,":[46],"often":[49],"produces":[50],"much":[51],"longer":[52],"traces":[53,69,116],"with":[54,170],"repetitive":[55,115],"loops,":[56],"budget":[57],"exhaustion,":[58],"delayed":[59],"commitment,":[60],"and":[61,76,79,110,117,173],"unclosed":[62],"segments.":[64],"We":[65],"analyze":[66],"full":[67],"Qwen3":[71],"models":[73],"across":[74],"mathematical":[75],"commonsense":[77],"benchmarks":[78],"accuracy":[82,135,181],"degradation":[83],"is":[84,189],"tightly":[85],"linked":[86],"these":[88],"process-level":[89],"failures.":[90],"To":[91],"address":[92],"them,":[93],"introduce":[95],"two":[96],"lightweight":[97,171],"controls:":[98],"FP16":[99,175],"planning,":[100],"which":[101,113],"gives":[102],"model":[105],"a":[106],"short":[107],"high-precision":[108],"outline,":[109],"loop":[111,131,143],"rescue,":[112],"detects":[114],"either":[118],"commits":[119],"an":[121],"earlier":[122],"or":[124],"falls":[125],"back":[126],"FP16.":[128],"On":[129],"MATH-500,":[130],"rescue":[132,144],"improves":[133,145],"Qwen3-8B":[134],"from":[136,147],"17.2%":[137],"74.2%,":[139],"while":[140,182],"planning":[141],"plus":[142],"Qwen3-32B":[146],"65.0%":[148],"87.2%.":[150],"Overall,":[151],"our":[152],"results":[153],"extreme":[156],"becomes":[159],"practical":[160],"when":[161],"its":[162],"failures":[163],"are":[164],"treated":[165],"as":[166],"controllable":[167],"pathologies:":[169],"detection":[172],"selective":[174],"support,":[176],"recover":[180],"preserving":[183],"real":[184],"speed.":[186],"Our":[187],"code":[188],"available":[190],"at:":[191],"https://github.com/brain-lab-research/quantized-reasoning.":[192]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-06-03T00:00:00"}