{"id":"https://openalex.org/W4416036176","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.822","title":"CARMA: Enhanced Compositionality in LLMs via Advanced Regularisation and Mutual Information Alignment","display_name":"CARMA: Enhanced Compositionality in LLMs via Advanced Regularisation and Mutual Information Alignment","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4416036176","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.822"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2025.emnlp-main.822","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.822","pdf_url":"https://aclanthology.org/2025.emnlp-main.822.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"type":"article","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.emnlp-main.822.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5114337749","display_name":"Nura Aljaafari","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Nura Aljaafari","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Danilo Carvalho","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Danilo Carvalho","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5022138909","display_name":"Andr\u00e9 Victor Lucci Freitas","orcid":"https://orcid.org/0000-0002-5763-4990"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Andre Freitas","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":3,"corresponding_author_ids":["https://openalex.org/A5114337749"],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.17488149,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"16250","last_page":"16270"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.05909999832510948,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.05909999832510948,"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/T11273","display_name":"Advanced Graph Neural Networks","score":0.04619999974966049,"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/T13702","display_name":"Machine Learning in Healthcare","score":0.040699999779462814,"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/mutual-information","display_name":"Mutual information","score":0.5942999720573425},{"id":"https://openalex.org/keywords/principle-of-compositionality","display_name":"Principle of compositionality","score":0.4526999890804291},{"id":"https://openalex.org/keywords/complete-information","display_name":"Complete information","score":0.3192000091075897},{"id":"https://openalex.org/keywords/information-system","display_name":"Information system","score":0.28769999742507935},{"id":"https://openalex.org/keywords/feature","display_name":"Feature (linguistics)","score":0.2743000090122223}],"concepts":[{"id":"https://openalex.org/C152139883","wikidata":"https://www.wikidata.org/wiki/Q252973","display_name":"Mutual information","level":2,"score":0.5942999720573425},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.559499979019165},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5236999988555908},{"id":"https://openalex.org/C121375916","wikidata":"https://www.wikidata.org/wiki/Q936559","display_name":"Principle of compositionality","level":2,"score":0.4526999890804291},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.39480000734329224},{"id":"https://openalex.org/C113336015","wikidata":"https://www.wikidata.org/wiki/Q574010","display_name":"Complete information","level":2,"score":0.3192000091075897},{"id":"https://openalex.org/C180198813","wikidata":"https://www.wikidata.org/wiki/Q121182","display_name":"Information system","level":2,"score":0.28769999742507935},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.2743000090122223},{"id":"https://openalex.org/C7797323","wikidata":"https://www.wikidata.org/wiki/Q3798612","display_name":"Pointwise mutual information","level":3,"score":0.25929999351501465},{"id":"https://openalex.org/C2778334786","wikidata":"https://www.wikidata.org/wiki/Q1586270","display_name":"Variation (astronomy)","level":2,"score":0.2567000091075897},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.2551000118255615}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.emnlp-main.822","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.822","pdf_url":"https://aclanthology.org/2025.emnlp-main.822.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.emnlp-main.822","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.822","pdf_url":"https://aclanthology.org/2025.emnlp-main.822.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[{"id":"https://openalex.org/G1467557333","display_name":"Etude de la 's\u00e9mantique syntaxique' en URSS","funder_award_id":"29196","funder_id":"https://openalex.org/F4320320924","funder_display_name":"Schweizerischer Nationalfonds zur F\u00f6rderung der Wissenschaftlichen Forschung"},{"id":"https://openalex.org/G4836148144","display_name":null,"funder_award_id":"20002","funder_id":"https://openalex.org/F4320320924","funder_display_name":"Schweizerischer Nationalfonds zur F\u00f6rderung der Wissenschaftlichen Forschung"},{"id":"https://openalex.org/G598684370","display_name":"RATIONAL: Robust Augmented Retrieval for Natural Language Inference over Transformer-based Models","funder_award_id":"229196","funder_id":"https://openalex.org/F4320320924","funder_display_name":"Schweizerischer Nationalfonds zur F\u00f6rderung der Wissenschaftlichen Forschung"},{"id":"https://openalex.org/G993971353","display_name":null,"funder_award_id":"200021","funder_id":"https://openalex.org/F4320320924","funder_display_name":"Schweizerischer Nationalfonds zur F\u00f6rderung der Wissenschaftlichen Forschung"}],"funders":[{"id":"https://openalex.org/F4320319985","display_name":"Cancer Research UK","ror":"https://ror.org/054225q67"},{"id":"https://openalex.org/F4320319990","display_name":"National Institute for Health and Care Research","ror":"https://ror.org/0187kwz08"},{"id":"https://openalex.org/F4320320924","display_name":"Schweizerischer Nationalfonds zur F\u00f6rderung der Wissenschaftlichen Forschung","ror":"https://ror.org/00yjd3n13"},{"id":"https://openalex.org/F4320333687","display_name":"Manchester Biomedical Research Centre","ror":null}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416036176.pdf","grobid_xml":"https://content.openalex.org/works/W4416036176.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2],"(LLMs)":[3],"struggle":[4],"with":[5,152],"compositional":[6,57,121,156],"generalisation,":[7],"limiting":[8],"their":[9],"ability":[10],"to":[11,16,73,104],"systematically":[12],"combine":[13],"learned":[14,134],"components":[15],"interpret":[17],"novel":[18],"inputs.While":[19],"architectural":[20],"modifications,":[21],"fine-tuning,":[22,115],"and":[23,54,69,82,91,102,119],"data":[24],"augmentation":[25],"improve":[26,155],"compositionality,":[27],"they":[28],"often":[29],"have":[30],"limited":[31],"adaptability,":[32],"face":[33],"scalability":[34],"constraints,":[35],"or":[36],"yield":[37],"diminishing":[38],"returns":[39],"on":[40,87,97],"real":[41],"data.To":[42],"address":[43],"this,":[44],"we":[45],"propose":[46],"CARMA,":[47],"an":[48],"intervention":[49],"that":[50,108,149],"enhances":[51],"the":[52,111],"stability":[53,71],"robustness":[55,103],"of":[56],"reasoning":[58],"in":[59,132,162],"LLMs":[60],"while":[61,158],"preserving":[62],"fine-tuned":[63],"performance.CARMA":[64],"employs":[65],"mutual":[66],"information":[67],"regularisation":[68],"layer-wise":[70],"constraints":[72],"mitigate":[74],"feature":[75],"fragmentation,":[76],"ensuring":[77],"structured":[78],"representations":[79],"persist":[80],"across":[81,126],"within":[83],"layers.We":[84],"evaluate":[85],"CARMA":[86,109,151,167],"inverse":[88],"dictionary":[89],"modelling":[90],"sentiment":[92],"classification,":[93],"measuring":[94],"its":[95,123],"impact":[96],"semantic":[98],"consistency,":[99],"performance":[100,161],"stability,":[101],"lexical":[105],"perturbations.Results":[106],"show":[107],"reduces":[110],"variability":[112],"introduced":[113],"by":[114],"stabilises":[116],"token":[117],"representations,":[118],"improves":[120],"reasoning.While":[122],"effectiveness":[124],"varies":[125],"architectures,":[127],"CARMA's":[128],"key":[129],"strength":[130],"lies":[131],"reinforcing":[133],"structures":[135],"rather":[136],"than":[137],"introducing":[138],"new":[139],"capabilities,":[140],"making":[141],"it":[142],"a":[143],"scalable":[144],"auxiliary":[145],"method.These":[146],"findings":[147],"suggest":[148],"integrating":[150],"fine-tuning":[153],"can":[154],"generalisation":[157],"maintaining":[159],"task-specific":[160],"LLMs.":[163],"Original":[164],"ModelFT":[165],"Model":[166,168]},"counts_by_year":[],"updated_date":"2026-04-10T15:06:20.359241","created_date":"2025-11-08T00:00:00"}