{"id":"https://openalex.org/W7225622012","doi":"https://doi.org/10.5281/zenodo.24085180","title":"When Weights Collide with Latents: Direct Weight-Space Arithmetic Induces Asymmetric Degeneration in Sparse Autoencoder Monosemanticity","display_name":"When Weights Collide with Latents: Direct Weight-Space Arithmetic Induces Asymmetric Degeneration in Sparse Autoencoder Monosemanticity","publication_year":2026,"publication_date":"2026-11-09","ids":{"openalex":"https://openalex.org/W7225622012","doi":"https://doi.org/10.5281/zenodo.24085180"},"language":"en","primary_location":{"id":"doi:10.5281/zenodo.24085180","doi":"https://doi.org/10.5281/zenodo.24085180","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.24085180","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"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"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.5281/zenodo.24085180","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5158794468","display_name":"MD Mushfiqur Rahim","orcid":"https://orcid.org/0009-0003-1021-685X","observed_orcids":["https://orcid.org/0009-0003-1021-685X"]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Mushfiqur Rahim, Md","raw_affiliation_strings":["Independent Researcher in Mechanistic Interpretability & Deep Learning"],"raw_orcid":"https://orcid.org/0009-0003-1021-685X","affiliations":[{"raw_affiliation_string":"Independent Researcher in Mechanistic Interpretability & Deep Learning","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5158794468"],"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9955999851226807,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9955999851226807,"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":0.0031999999191612005,"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":9.999999747378752e-05,"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":[],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6126000285148621},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.564300000667572},{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.5580999851226807},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.4180999994277954},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3937999904155731},{"id":"https://openalex.org/C83665646","wikidata":"https://www.wikidata.org/wiki/Q42139305","display_name":"Feature vector","level":2,"score":0.3885999917984009},{"id":"https://openalex.org/C108710211","wikidata":"https://www.wikidata.org/wiki/Q11538","display_name":"Mathematical proof","level":2,"score":0.3765000104904175},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.3700000047683716},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.3659999966621399},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.35690000653266907},{"id":"https://openalex.org/C137800194","wikidata":"https://www.wikidata.org/wiki/Q11713455","display_name":"Interpolation (computer graphics)","level":3,"score":0.33500000834465027},{"id":"https://openalex.org/C124304363","wikidata":"https://www.wikidata.org/wiki/Q673661","display_name":"Abstraction","level":2,"score":0.3163999915122986},{"id":"https://openalex.org/C77637269","wikidata":"https://www.wikidata.org/wiki/Q7002051","display_name":"Neural coding","level":2,"score":0.311599999666214},{"id":"https://openalex.org/C175291020","wikidata":"https://www.wikidata.org/wiki/Q1156822","display_name":"Offset (computer science)","level":2,"score":0.2831000089645386},{"id":"https://openalex.org/C101738243","wikidata":"https://www.wikidata.org/wiki/Q786435","display_name":"Autoencoder","level":3,"score":0.28040000796318054},{"id":"https://openalex.org/C17137986","wikidata":"https://www.wikidata.org/wiki/Q215067","display_name":"Orthogonality","level":2,"score":0.2802000045776367},{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.2759999930858612},{"id":"https://openalex.org/C12267149","wikidata":"https://www.wikidata.org/wiki/Q282453","display_name":"Support vector machine","level":2,"score":0.2757999897003174},{"id":"https://openalex.org/C122280245","wikidata":"https://www.wikidata.org/wiki/Q620622","display_name":"Kernel method","level":3,"score":0.2750999927520752},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.273499995470047},{"id":"https://openalex.org/C150921843","wikidata":"https://www.wikidata.org/wiki/Q1170431","display_name":"Resampling","level":2,"score":0.2712000012397766},{"id":"https://openalex.org/C43711488","wikidata":"https://www.wikidata.org/wiki/Q7534783","display_name":"Skew","level":2,"score":0.2703999876976013},{"id":"https://openalex.org/C103278499","wikidata":"https://www.wikidata.org/wiki/Q254465","display_name":"Similarity (geometry)","level":3,"score":0.266400009393692},{"id":"https://openalex.org/C183115368","wikidata":"https://www.wikidata.org/wiki/Q856577","display_name":"Weighting","level":2,"score":0.2662999927997589},{"id":"https://openalex.org/C124066611","wikidata":"https://www.wikidata.org/wiki/Q28684319","display_name":"Sparse approximation","level":2,"score":0.2648000121116638},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.26170000433921814},{"id":"https://openalex.org/C152124472","wikidata":"https://www.wikidata.org/wiki/Q1204361","display_name":"Redundancy (engineering)","level":2,"score":0.257099986076355},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.25429999828338623},{"id":"https://openalex.org/C101468663","wikidata":"https://www.wikidata.org/wiki/Q1620158","display_name":"Modular design","level":2,"score":0.25270000100135803}],"mesh":[],"locations_count":2,"locations":[{"id":"doi:10.5281/zenodo.24085180","doi":"https://doi.org/10.5281/zenodo.24085180","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.24085180","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"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"},{"id":"doi:10.5281/zenodo.24085181","doi":"https://doi.org/10.5281/zenodo.24085181","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.24085181","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"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":"doi:10.5281/zenodo.24085180","doi":"https://doi.org/10.5281/zenodo.24085180","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.24085180","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"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":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Abstract:":[0],"Model":[1],"editing":[2,199],"through":[3],"weight-space":[4,71,109],"task":[5,86,152],"vector":[6],"arithmetic":[7,190],"(\u03b8edited":[8],"=":[9,145],"\u03b8base":[10],"+":[11],"\u03bb\u03c4)":[12],"has":[13],"emerged":[14],"as":[15],"an":[16,130,139],"exceptionally":[17],"compute-efficient":[18],"paradigm":[19],"for":[20,43,197],"steering":[21],"foundational":[22],"language":[23],"models":[24,227],"without":[25,179],"full":[26],"parameter":[27,189],"retraining.":[28],"Concurrently,":[29],"Sparse":[30],"Autoencoders":[31],"(SAEs)":[32],"trained":[33],"on":[34,93],"intermediate":[35],"residual":[36],"activations":[37],"have":[38],"become":[39],"the":[40,61,68,79,161,181],"gold":[41],"standard":[42],"mechanistic":[44,65],"interpretability":[45],"by":[46,129],"decomposing":[47],"polysemantic":[48],"neural":[49,201],"representations":[50],"into":[51,67],"interpretable,":[52],"monosemantic":[53,94],"feature":[54,95],"dictionaries.":[55],"In":[56],"this":[57],"monograph,":[58],"we":[59,137,159],"present":[60],"first":[62],"empirical":[63,140],"and":[64,73,88,97,191,200,225],"investigation":[66],"intersection":[69],"of":[70,81,143,176],"programming":[72],"sparse":[74],"dictionary":[75,177],"learning.":[76],"We":[77,100],"formalize":[78],"perturbation":[80],"transformer":[82],"parameters":[83],"along":[84],"directional":[85],"vectors":[87,153],"evaluate":[89],"its":[90],"downstream":[91],"consequence":[92],"preservation":[96,120],"reconstruction":[98,134],"fidelity.":[99],"uncover":[101],"a":[102,115,167],"striking":[103],"Asymmetric":[104],"Latent":[105,163],"Degeneration":[106],"Phenomenon:":[107],"positive":[108],"interpolation":[110],"(\u03bb":[111],">":[112],"0)":[113],"causes":[114],"non-linear":[116],"collapse":[117],"in":[118,133],"latent":[119],"(cosine":[121],"similarity":[122],"drops":[123],"from":[124],"1.00":[125],"to":[126,174],"0.72),":[127],"accompanied":[128],"exponential":[131],"surge":[132],"error.":[135],"Furthermore,":[136],"reveal":[138],"asymmetry":[141],"ratio":[142],"\u03c1":[144],"0.950247,":[146],"proving":[147],"that":[148,171],"adding":[149],"versus":[150],"subtracting":[151],"induces":[154],"non-symmetric":[155],"representational":[156],"drift.":[157],"Finally,":[158],"formulate":[160],"Orthogonal":[162],"Re-Projection":[164],"Kernel":[165],"(OLR-K),":[166],"closed-form":[168],"geometric":[169],"correction":[170],"restores":[172],"up":[173],"88.4%":[175],"alignment":[178],"retraining":[180],"autoencoder.":[182],"Our":[183],"findings":[184],"expose":[185],"fundamental":[186],"trade-offs":[187],"between":[188],"causal":[192],"interpretability,":[193],"establishing":[194],"essential":[195],"guardrails":[196],"model":[198],"reprogramming.":[202],"Key":[203],"Highlights:":[204],"19-Page":[205],"Comprehensive":[206],"Theoretical":[207],"&":[208,217,230],"Empirical":[209,220],"Monograph":[210],"Formal":[211],"Mathematical":[212],"Proofs":[213],"(Latent":[214],"Orthogonality":[215],"Drift":[216],"Rank":[218],"Collapse)":[219],"Validation":[221],"across":[222],"Gemma-2,":[223],"LLaMA,":[224],"GPT-2":[226],"Interactive":[228],"Visualizer":[229],"Reproducible":[231],"Colab":[232],"Notebook":[233],"included":[234]},"counts_by_year":[],"updated_date":"2026-11-10T13:44:29.811591","created_date":"2026-11-10T00:00:00"}