{"id":"https://openalex.org/W7225447100","doi":null,"title":"Concept Decodability Across Chess Architectures: Testing Whether Human-Aligned Models Encode Concepts More Linearly","display_name":"Concept Decodability Across Chess Architectures: Testing Whether Human-Aligned Models Encode Concepts More Linearly","publication_year":2026,"publication_date":"2026-09-30","ids":{"openalex":"https://openalex.org/W7225447100"},"language":"en","primary_location":{"id":"pmh:oai:eprints.gla.ac.uk:397716","doi":null,"is_oa":true,"landing_page_url":"https://eprints.gla.ac.uk/view/author/83773.html>,","pdf_url":"https://eprints.gla.ac.uk/397716/1/397716.pdf","source":{"id":"https://openalex.org/S4210235606","display_name":"ENLIGHTEN (Jurnal Bimbingan dan Konseling Islam)","issn_l":"2622-8912","issn":["2622-8912","2622-8920"],"is_oa":true,"is_in_doaj":true,"is_core":false,"listed_in":["doaj"],"host_organization":"https://openalex.org/I4405259220","host_organization_name":"Institut Agama Islam Negeri Langsa","host_organization_lineage":["https://openalex.org/I4405259220"],"host_organization_lineage_names":[],"type":"journal"},"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":"Conference or Workshop Item"},"volume":null,"type":"conference-paper","is_traditional":true,"indexed_in":[],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://eprints.gla.ac.uk/397716/1/397716.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5113398991","display_name":"Antonella M A Marsella","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Marsella, Antonella M. A.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5068323517","display_name":"Tiffany Vlaar","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Vlaar, Tiffany Joyce","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5018297803","display_name":"Stuart Grey","orcid":"https://orcid.org/0000-0001-9274-0601","observed_orcids":["https://orcid.org/0000-0001-9274-0601"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Grey, Stuart","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5082776788","display_name":"Fani Deligianni","orcid":"https://orcid.org/0000-0003-1306-5017","observed_orcids":["https://orcid.org/0000-0003-1306-5017"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Deligianni, Fani","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":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.82025482,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"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":0.6449999809265137,"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":0.6449999809265137,"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/T11574","display_name":"Artificial Intelligence in Games","score":0.3449999988079071,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.00139999995008111,"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/chess-engines","display_name":"chess engines","score":0.7429999709129333},{"id":"https://openalex.org/keywords/maia-2","display_name":"MAIA-2","score":0.781000018119812},{"id":"https://openalex.org/keywords/interpretability","display_name":"interpretability","score":0.6259999871253967},{"id":"https://openalex.org/keywords/linear-probing","display_name":"linear probing","score":0.5049999952316284},{"id":"https://openalex.org/keywords/materials-evaluation","display_name":"material evaluation","score":0.6179999709129333},{"id":"https://openalex.org/keywords/model-architecture-comparison","display_name":"model architecture comparison","score":0.5559999942779541},{"id":"https://openalex.org/keywords/hybrid-cnn-transformer-model","display_name":"CNN-Transformer hybrid","score":0.527999997138977}],"concepts":[{"id":"https://openalex.org/C66746571","wikidata":"https://www.wikidata.org/wiki/Q1134833","display_name":"ENCODE","level":3,"score":0.7746999859809875},{"id":"https://openalex.org/C184898388","wikidata":"https://www.wikidata.org/wiki/Q1435712","display_name":"Pairwise comparison","level":2,"score":0.6662999987602234},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6140999794006348},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5442000031471252},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.42899999022483826},{"id":"https://openalex.org/C206041023","wikidata":"https://www.wikidata.org/wiki/Q1751970","display_name":"Wilcoxon signed-rank test","level":3,"score":0.42800000309944153},{"id":"https://openalex.org/C48372109","wikidata":"https://www.wikidata.org/wiki/Q3913","display_name":"Binary number","level":2,"score":0.3540000021457672},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.3379000127315521},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.30410000681877136},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.30390000343322754},{"id":"https://openalex.org/C2777267654","wikidata":"https://www.wikidata.org/wiki/Q3519023","display_name":"Test (biology)","level":2,"score":0.28349998593330383},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.2736000120639801},{"id":"https://openalex.org/C25343380","wikidata":"https://www.wikidata.org/wiki/Q277521","display_name":"Relation (database)","level":2,"score":0.2676999866962433}],"mesh":[],"locations_count":1,"locations":[{"id":"pmh:oai:eprints.gla.ac.uk:397716","doi":null,"is_oa":true,"landing_page_url":"https://eprints.gla.ac.uk/view/author/83773.html>,","pdf_url":"https://eprints.gla.ac.uk/397716/1/397716.pdf","source":{"id":"https://openalex.org/S4210235606","display_name":"ENLIGHTEN (Jurnal Bimbingan dan Konseling Islam)","issn_l":"2622-8912","issn":["2622-8912","2622-8920"],"is_oa":true,"is_in_doaj":true,"is_core":false,"listed_in":["doaj"],"host_organization":"https://openalex.org/I4405259220","host_organization_name":"Institut Agama Islam Negeri Langsa","host_organization_lineage":["https://openalex.org/I4405259220"],"host_organization_lineage_names":[],"type":"journal"},"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":"Conference or Workshop Item"}],"best_oa_location":{"id":"pmh:oai:eprints.gla.ac.uk:397716","doi":null,"is_oa":true,"landing_page_url":"https://eprints.gla.ac.uk/view/author/83773.html>,","pdf_url":"https://eprints.gla.ac.uk/397716/1/397716.pdf","source":{"id":"https://openalex.org/S4210235606","display_name":"ENLIGHTEN (Jurnal Bimbingan dan Konseling Islam)","issn_l":"2622-8912","issn":["2622-8912","2622-8920"],"is_oa":true,"is_in_doaj":true,"is_core":false,"listed_in":["doaj"],"host_organization":"https://openalex.org/I4405259220","host_organization_name":"Institut Agama Islam Negeri Langsa","host_organization_lineage":["https://openalex.org/I4405259220"],"host_organization_lineage_names":[],"type":"journal"},"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":"Conference or Workshop Item"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7225447100.pdf","grobid_xml":"https://content.openalex.org/works/W7225447100.grobid-xml"},"referenced_works_count":1,"referenced_works":["https://openalex.org/W7160903411"],"related_works":["https://openalex.org/W4412082334","https://openalex.org/W7214510316","https://openalex.org/W7222986731","https://openalex.org/W7161802264","https://openalex.org/W7165614246","https://openalex.org/W4415798814","https://openalex.org/W4413025194","https://openalex.org/W7225139677","https://openalex.org/W7103148622","https://openalex.org/W4400701072"],"abstract_inverted_index":{"Introduction:":[0],"Recent":[1],"work":[2,47,165,428],"has":[3],"pushed":[4],"toward":[5,443],"human-aligned":[6,70,81],"chess":[7,58,382],"engines":[8,42],"capable":[9],"of":[10,63,127,200,212,260,351,372],"predicting":[11],"moves":[12],"at":[13,149,396,419],"a":[14,19,25,124,192,218,444],"specific":[15],"skill":[16],"level.Maia-2":[17],"[1],":[18],"hybrid":[20],"CNN-transformer,":[21],"and":[22,85,130,142,176,250,264,291,343,362,439],"Allie":[23,251],"[2],":[24],"GPT-2-style":[26],"transformer,":[27],"exemplify":[28],"this":[29,73,167,261,304,431],"shift:":[30],"both":[31,174,319,352],"are":[32,66,101,384],"trained":[33],"on":[34,355],"human":[35],"games":[36],"rather":[37],"than":[38,318],"self-play,":[39],"unlike":[40],"superhuman":[41,50],"such":[43],"as":[44,197],"Stockfish":[45,255,300],"[3].Prior":[46],"shows":[48],"that":[49],"models":[51,178],"also":[52],"encode":[53],"linearly":[54,317,395],"interpretable":[55],"concepts":[56,65,100,196,206,315,383,393],"in":[57,69,161,173,276,337,378],"[4].This":[59],"raised":[60],"the":[61,96,198,213,247,252,258,265,296,325,366,397,401,405,415,422],"question":[62],"whether":[64,93],"more":[67],"decodable":[68,275,385],"models.We":[71],"test":[72,220,323],"directly":[74],"by":[75,228,271],"comparing":[76],"concept":[77,110,243,376,423],"decodability":[78,249],"across":[79,108,203,240,386],"two":[80],"architectures":[82,354],"(Maia-2,":[83],"Allie)":[84],"one":[86],"purely":[87],"evaluation-driven":[88],"architecture":[89,262],"(Stockfish's":[90],"NNUE),":[91],"asking":[92],"human-alignment":[94],"predicts":[95],"extent":[97],"to":[98,311],"which":[99],"decodable.Methodology:":[102],"Using":[103],"an":[104],"81-concept":[105],"binary":[106],"battery":[107],"six":[109,214],"groups":[111],"(king":[112],"safety,":[113,357],"tactical":[114,116,282,285,358],"forks,":[115],"pins,":[117,359],"positional":[118,288,360],"structure,":[119,121,361],"pawn":[120,292,363],"material),":[122],"with":[123,163,191,221,254],"synthetic":[125],"dataset":[126],"10,000":[128,131],"positive":[129],"negative":[132],"entries":[133],"per":[134],"concept,":[135],"we":[136],"extract":[137],"activations":[138],"via":[139,182,436],"forward":[140],"hooks":[141],"train":[143],"linear":[144],"support":[145],"vector":[146],"machine":[147],"probes":[148],"each":[150,211,268],"model's":[151],"pre-output":[152,398],"layer,":[153],"our":[154],"anchor":[155],"point":[156],"for":[157,224,233,408,421,449],"crossarchitecture":[158],"comparability.This":[159],"is":[160,180,189],"accordance":[162],"prior":[164],"showing":[166],"layer":[168,399],"gives":[169],"consistent,":[170],"comparable":[171],"results":[172],"CNN":[175],"transformer":[177],"[5].Decodability":[179],"scored":[181],"F1":[183],"under":[184],"stratified":[185],"60/20/20":[186],"splits.Statistical":[187],"significance":[188],"assessed":[190],"repeated-measures":[193],"framework":[194],"treating":[195],"unit":[199],"analysis,":[201],"first":[202,367],"all":[204,242],"81":[205],"jointly,":[207],"then":[208],"separately":[209],"within":[210,267,414],"semantic":[215],"groups,":[216,244],"using":[217,433],"Friedman":[219,322],"Kendall's":[222],"W":[223,340],"effect":[225],"size,":[226],"followed":[227],"Holm-corrected":[229,344],"Wilcoxon":[230],"signedrank":[231],"tests":[232],"pairwise":[234,345],"comparisons.Results:":[235],"As":[236],"Figure":[237],"1":[238],"shows,":[239],"nearly":[241],"Maia-2":[245,348,391],"achieves":[246],"highest":[248],"lowest,":[253],"(SF)":[256],"intermediate.However,":[257],"size":[259],"gap,":[263],"variance":[266],"model,":[269],"differ":[270],"concept.Maia-2":[272],"was":[273,295,306],"most":[274,394],"king":[277,356],"safety":[278],"(F1":[279],"=":[280,331,341],"0.772),":[281],"forks":[283],"(0.794),":[284],"pins":[286],"(0.723),":[287],"structure":[289,293],"(0.829),":[290],"(0.717).Material":[294],"only":[297],"category":[298],"where":[299],"outperformed":[301],"(0.733),":[302],"though":[303],"difference":[305],"not":[307],"statistically":[308,369],"significant":[309,329],"due":[310],"low":[312],"power.Allie":[313],"decoded":[314],"less":[316],"other":[320,353],"architectures.The":[321],"confirmed":[324],"overall":[326],"differences":[327],"were":[328],"(\u03c7\u00b2(2)":[330],"99.56,p":[332],"<":[333],".001)with":[334],"high":[335],"concordance":[336],"ordering":[338],"(Kendall's":[339],"0.615),":[342],"comparisons":[346],"showed":[347],"significantly":[349],"ahead":[350],"structure.This":[364],"offers":[365],"systematic,":[368],"grounded":[370],"map":[371],"shared":[373],"versus":[374],"architecture-specific":[375],"representation":[377],"chess-playing":[379],"networks.Conclusion:":[380],"Human-interpretable":[381],"fundamentally":[387],"different":[388],"architectures,":[389],"but":[390],"encodes":[392],"among":[400],"three,":[402],"making":[403],"it":[404,425],"strongest":[406],"candidate":[407],"translating":[409],"internal":[410],"representations":[411],"into":[412],"explanations":[413],"intelligent":[416,446],"tutoring":[417,447],"system,":[418],"least":[420],"classes":[424],"captures":[426],"best.Future":[427],"will":[429],"extend":[430],"analysis":[432],"causal":[434],"validation":[435],"activation":[437],"patching":[438],"sparse-autoencoder-based":[440],"feature":[441],"discovery,":[442],"skill-adaptive":[445],"system":[448],"chess.":[450]},"counts_by_year":[],"updated_date":"2026-11-08T12:12:28.789561","created_date":"2026-11-07T00:00:00"}