{"id":"https://openalex.org/W7223339489","doi":"https://doi.org/10.48550/arxiv.2610.30191","title":"Why Constrained Updates Still Forget: A Functional-Path Analysis of Continual SFT","display_name":"Why Constrained Updates Still Forget: A Functional-Path Analysis of Continual SFT","publication_year":2026,"publication_date":"2026-10-21","ids":{"openalex":"https://openalex.org/W7223339489","doi":"https://doi.org/10.48550/arxiv.2610.30191"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2610.30191","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.30191","pdf_url":"https://arxiv.org/pdf/2610.30191","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","license_id":"https://openalex.org/licenses/cc-by","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/2610.30191","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5159356837","display_name":"Qingsen Ma","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I177725633","display_name":"Chinese University of Hong Kong","ror":"https://ror.org/00t33hh48","country_code":"HK","type":"education","lineage":["https://openalex.org/I177725633"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Ma, Qingsen","raw_affiliation_strings":["The Chinese University of Hong Kong"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"The Chinese University of Hong Kong","institution_ids":["https://openalex.org/I177725633"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5159327368","display_name":"Dianyun Wang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I139759216","display_name":"Beijing University of Posts and Telecommunications","ror":"https://ror.org/04w9fbh59","country_code":"CN","type":"education","lineage":["https://openalex.org/I139759216"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wang, Dianyun","raw_affiliation_strings":["Beijing University of Posts and Telecommunications"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Beijing University of Posts and Telecommunications","institution_ids":["https://openalex.org/I139759216"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5159356105","display_name":"Jiahong Liu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I177725633","display_name":"Chinese University of Hong Kong","ror":"https://ror.org/00t33hh48","country_code":"HK","type":"education","lineage":["https://openalex.org/I177725633"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Liu, Jiahong","raw_affiliation_strings":["The Chinese University of Hong Kong"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"The Chinese University of Hong Kong","institution_ids":["https://openalex.org/I177725633"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5123516107","display_name":"Irwin King","orcid":"https://orcid.org/0000-0001-8106-6447","observed_orcids":["https://orcid.org/0000-0001-8106-6447"]},"institutions":[{"id":"https://openalex.org/I177725633","display_name":"Chinese University of Hong Kong","ror":"https://ror.org/00t33hh48","country_code":"HK","type":"education","lineage":["https://openalex.org/I177725633"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"King, Irwin","raw_affiliation_strings":["The Chinese University of Hong Kong"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"The Chinese University of Hong Kong","institution_ids":["https://openalex.org/I177725633"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.9794999957084656,"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"}},"topics":[{"id":"https://openalex.org/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.9794999957084656,"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/T10028","display_name":"Topic Modeling","score":0.01590000092983246,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.000699999975040555,"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/catastrophic-forgetting","display_name":"catastrophic forgetting","score":0.8939999938011169},{"id":"https://openalex.org/keywords/qlora","display_name":"QLoRA","score":0.7509999871253967},{"id":"https://openalex.org/keywords/backward-transfer","display_name":"backward transfer","score":0.5979999899864197},{"id":"https://openalex.org/keywords/language-model-adaptation","display_name":"language model adaptation","score":0.3700000047683716}],"concepts":[{"id":"https://openalex.org/C177774035","wikidata":"https://www.wikidata.org/wiki/Q1246948","display_name":"Granularity","level":2,"score":0.6985999941825867},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6973999738693237},{"id":"https://openalex.org/C32834561","wikidata":"https://www.wikidata.org/wiki/Q660730","display_name":"Subspace topology","level":2,"score":0.6140999794006348},{"id":"https://openalex.org/C2776036281","wikidata":"https://www.wikidata.org/wiki/Q48769818","display_name":"Constraint (computer-aided design)","level":2,"score":0.4788999855518341},{"id":"https://openalex.org/C57493831","wikidata":"https://www.wikidata.org/wiki/Q3134666","display_name":"Projection (relational algebra)","level":2,"score":0.4749999940395355},{"id":"https://openalex.org/C13662910","wikidata":"https://www.wikidata.org/wiki/Q193139","display_name":"Trajectory","level":2,"score":0.4607999920845032},{"id":"https://openalex.org/C2776135515","wikidata":"https://www.wikidata.org/wiki/Q17143721","display_name":"Regularization (linguistics)","level":2,"score":0.4528999924659729},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.4499000012874603},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4228000044822693},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.37389999628067017},{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.3352000117301941},{"id":"https://openalex.org/C22367795","wikidata":"https://www.wikidata.org/wiki/Q7625208","display_name":"Structured prediction","level":2,"score":0.3330000042915344},{"id":"https://openalex.org/C147764199","wikidata":"https://www.wikidata.org/wiki/Q6865248","display_name":"Minification","level":2,"score":0.2711000144481659},{"id":"https://openalex.org/C18555067","wikidata":"https://www.wikidata.org/wiki/Q8375051","display_name":"Joint (building)","level":2,"score":0.2637999951839447},{"id":"https://openalex.org/C189237950","wikidata":"https://www.wikidata.org/wiki/Q2500758","display_name":"Stationary point","level":2,"score":0.26080000400543213},{"id":"https://openalex.org/C8505890","wikidata":"https://www.wikidata.org/wiki/Q605095","display_name":"Budget constraint","level":2,"score":0.2538999915122986},{"id":"https://openalex.org/C136643341","wikidata":"https://www.wikidata.org/wiki/Q1361526","display_name":"Reachability","level":2,"score":0.251800000667572}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2610.30191","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.30191","pdf_url":"https://arxiv.org/pdf/2610.30191","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","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2610.30191","doi":"https://doi.org/10.48550/arxiv.2610.30191","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2610.30191","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","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2610.30191","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.30191","pdf_url":"https://arxiv.org/pdf/2610.30191","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","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7223339489.pdf","grobid_xml":"https://content.openalex.org/works/W7223339489.grobid-xml"},"referenced_works_count":8,"referenced_works":["https://openalex.org/W2902625698","https://openalex.org/W2908510526","https://openalex.org/W3168867926","https://openalex.org/W4403796339","https://openalex.org/W4406032137","https://openalex.org/W6929549612","https://openalex.org/W7117724335","https://openalex.org/W7134931384"],"related_works":["https://openalex.org/W7162795067","https://openalex.org/W7221545612","https://openalex.org/W7162750378","https://openalex.org/W6966871545","https://openalex.org/W7160599108","https://openalex.org/W7161974287","https://openalex.org/W7127612843","https://openalex.org/W7220692321","https://openalex.org/W7222526522","https://openalex.org/W7220403910"],"abstract_inverted_index":{"Supervised":[0],"fine-tuning":[1],"adapts":[2],"language":[3],"models":[4],"to":[5,131,170],"specialized":[6],"tasks":[7],"but":[8],"can":[9],"disrupt":[10],"previously":[11],"acquired":[12],"behavior.":[13],"Existing":[14],"approaches":[15],"constrain":[16],"parameters,":[17],"gradients,":[18],"or":[19,85],"replay":[20,116,148],"objectives,":[21],"yet":[22],"the":[23,26,29,50,69,93,98,103],"connection":[24],"between":[25],"protected":[27],"object,":[28],"intervening":[30],"trajectory,":[31],"and":[32,57,67,114,136,173,182],"retained":[33],"outputs":[34],"remains":[35],"unclear.":[36],"We":[37],"develop":[38],"a":[39,63],"functional-path":[40],"account":[41,61],"that":[42],"decomposes":[43],"old-loss":[44],"changes":[45],"along":[46],"realized":[47],"updates":[48],"into":[49],"controlled":[51],"direction,":[52],"gradient":[53],"staleness,":[54],"probe":[55],"mismatch,":[56],"finite-step":[58],"effects.":[59],"The":[60],"yields":[62],"conditional":[64],"refresh":[65],"budget":[66],"identifies":[68],"restoring":[70],"direction":[71],"supplied":[72],"by":[73,122],"informative":[74],"replay.":[75],"Matched":[76],"QLoRA":[77],"interventions":[78],"separate":[79],"constraint":[80],"granularity":[81,99],"from":[82,110,162],"placement":[83,88],"before":[84],"after":[86],"Adam:":[87],"strongly":[89],"affects":[90],"retention":[91,140],"across":[92],"studied":[94],"code-label":[95],"audits,":[96],"whereas":[97],"interaction":[100],"depends":[101],"on":[102],"endpoint.":[104],"Checkpointed":[105],"generation":[106],"separates":[107],"endpoint":[108],"recovery":[109],"higher":[111],"sampled-path":[112],"accuracy,":[113],"matched":[115],"schedules":[117],"reveal":[118],"loss":[119],"exposure":[120],"concealed":[121],"final":[123],"scores.":[124],"A":[125],"sequential-task":[126],"comparison":[127],"extends":[128],"this":[129,154],"distinction":[130],"backward":[132],"transfer:":[133],"subspace":[134],"regularization":[135],"projection":[137],"improve":[138],"average":[139],"while":[141],"trading":[142],"off":[143],"different":[144],"old":[145],"tasks;":[146],"supervised":[147],"protects":[149],"both":[150],"more":[151],"consistently":[152],"in":[153],"setting.":[155],"Additional":[156],"response":[157],"interfaces":[158],"distinguish":[159],"content":[160],"gains":[161],"exact-response":[163],"retention.":[164],"These":[165],"findings":[166],"connect":[167],"update-level":[168],"protection":[169],"task-wise":[171],"behavior":[172],"motivate":[174],"mitigation":[175],"through":[176],"explicit":[177],"functional":[178],"anchors,":[179],"trajectory":[180],"evaluation,":[181],"joint":[183],"old/new":[184],"auditing.":[185]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-10-24T00:00:00"}