{"paper":{"title":"Evaluation Drift in LLM Personality Induction: Are We Moving the Goalpost?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Fine-tuning stabilizes LLM personality questionnaire scores but full-profile accuracy stays near chance.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Iyiola E. Olatunji, Jacques Klein, Prateek Rajput, Tegawend\\'e F. Bissyand\\'e, Yewei Song","submitted_at":"2026-05-16T13:44:06Z","abstract_excerpt":"Can large language models reliably express a human-like personality, or are they merely mimicking surface cues without a stable underlying profile? To investigate this, we induce personality in LLMs by fine-tuning them on the long-form essays, where each essay is associated with a target Big Five personality profile. We then evaluate the stability and fidelity of the induced personality using the IPIP-NEO questionnaire. Specifically, we ask: (i) does post-training (SFT, DPO, ORPO) stabilize questionnaire scores under prompt rephrasings, and (ii) can it induce target Big Five profiles from ungu"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our results demonstrate that fine-tuning consistently reduces variance in questionnaire responses across five models, directly mitigating the evaluation fragility reported in pre-trained models. However, this newfound stability reveals a more fundamental limitation: accuracy on the full five-dimensional profile remains near chance, even when single-trait scores improve.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The IPIP-NEO questionnaire responses from LLMs validly and comprehensively measure the induced personality profile, and that the unguided essays contain sufficient trait-relevant cues to support faithful induction of the target five-dimensional profile.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Fine-tuning LLMs on essays reduces variance in IPIP-NEO responses across models but does not raise full five-trait profile accuracy above near-chance levels from unguided text.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Fine-tuning stabilizes LLM personality questionnaire scores but full-profile accuracy stays near chance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"a60b6640618cbedef44a9a55d1a94d12a4eb898ad9db8ddfabb27798975a88f4"},"source":{"id":"2605.16996","kind":"arxiv","version":1},"verdict":{"id":"ad810bb2-9a0a-4c0c-b352-f170501a5933","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T20:37:32.521866Z","strongest_claim":"Our results demonstrate that fine-tuning consistently reduces variance in questionnaire responses across five models, directly mitigating the evaluation fragility reported in pre-trained models. However, this newfound stability reveals a more fundamental limitation: accuracy on the full five-dimensional profile remains near chance, even when single-trait scores improve.","one_line_summary":"Fine-tuning LLMs on essays reduces variance in IPIP-NEO responses across models but does not raise full five-trait profile accuracy above near-chance levels from unguided text.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The IPIP-NEO questionnaire responses from LLMs validly and comprehensively measure the induced personality profile, and that the unguided essays contain sufficient trait-relevant cues to support faithful induction of the target five-dimensional profile.","pith_extraction_headline":"Fine-tuning stabilizes LLM personality questionnaire scores but full-profile accuracy stays near chance."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.16996/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T21:01:19.050961Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T20:50:49.377930Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"citation_quote_validity","ran_at":"2026-05-19T19:49:57.560600Z","status":"skipped","version":"0.1.0","findings_count":0},{"name":"cited_work_retraction","ran_at":"2026-05-19T19:23:36.116463Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T18:41:56.202536Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.291934Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"3f553111af421dbb59a5183c17b8c6c9e2917a692f49d0ca6111a77be026d30d"},"references":{"count":16,"sample":[{"doi":"","year":2023,"title":"InFindings of the Associa- tion for Computational Linguistics: EMNLP 2023, pages 2370–2386, Singapore","work_id":"26cf2f73-15a2-4856-a148-b4929d09762c","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Hans Christian, Derwin Suhartono, Andry Chowanda, and Kamal Z Zamli","work_id":"423d01ee-88a4-44d0-8336-3471bde8ab87","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","ref_index":3,"cited_arxiv_id":"2407.21783","is_internal_anchor":true},{"doi":"","year":null,"title":"Matej Gjurković and Jan Šnajder","work_id":"91f3b0b4-c23e-4a07-aeb0-5b3c98b33f0c","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2011,"title":"In2011 IEEE third international conference on privacy, security, risk and trust and 2011 IEEE third international conference on social computing, pages 149–156","work_id":"46248671-7c5f-40ad-b37b-b18746ecd5cd","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":16,"snapshot_sha256":"613212c98550714faef8b1d1e7ac3c071986de75adab7d1d7bf8cc35774b360b","internal_anchors":3},"formal_canon":{"evidence_count":2,"snapshot_sha256":"ed57fe63a88412f21d67b18bda2db6e0af8deb990c25bca0ab02661362c1d5a3"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}