{"paper":{"title":"Macro: Enhancing Multilingual Counterfactual Explanations through Alignment-as-Preference Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A preference alignment method called Macro improves the validity of multilingual self-generated counterfactual explanations by 12.55 percent on average while maintaining minimality.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Bohao Chu, Jing Yang, Qianli Wang, Simon Ostermann, Yihong Liu, Yilong Wang","submitted_at":"2026-05-12T06:56:18Z","abstract_excerpt":"Self-generated counterfactual explanations (SCEs) are minimally modified inputs (minimality) generated by large language models (LLMs) that flip their own predictions (validity), offering a causally grounded approach to unraveling black-box LLM behavior. Yet extending them beyond English remains challenging: existing methods struggle to produce valid SCEs in non-dominant languages, and a persistent trade-off between validity and minimality undermines explanation quality. We introduce Macro, a preference alignment framework that applies Direct Preference Optimization (DPO) to multilingual SCE g"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Experiments across four LLMs and seven typologically diverse languages show that Macro improves validity by 12.55% on average over the chain-of-thought baseline without degrading minimality, while avoiding the severe minimality violations of the translation-based baseline.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The composite scoring function used to construct preference pairs accurately and unbiasedly captures the validity-minimality trade-off across typologically diverse languages and different LLMs.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Macro uses Direct Preference Optimization on composite-scored preference pairs to improve validity of multilingual self-generated counterfactual explanations by 12.55% on average without degrading minimality.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A preference alignment method called Macro improves the validity of multilingual self-generated counterfactual explanations by 12.55 percent on average while maintaining minimality.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"28e00cdb8af15cb3ccd9b8b0554f11a65c4b9b02c3c90dd8579829e53e2b0b6e"},"source":{"id":"2605.11632","kind":"arxiv","version":2},"verdict":{"id":"907fb810-b009-4583-a604-482d01beb33c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T01:13:58.101109Z","strongest_claim":"Experiments across four LLMs and seven typologically diverse languages show that Macro improves validity by 12.55% on average over the chain-of-thought baseline without degrading minimality, while avoiding the severe minimality violations of the translation-based baseline.","one_line_summary":"Macro uses Direct Preference Optimization on composite-scored preference pairs to improve validity of multilingual self-generated counterfactual explanations by 12.55% on average without degrading minimality.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The composite scoring function used to construct preference pairs accurately and unbiasedly captures the validity-minimality trade-off across typologically diverse languages and different LLMs.","pith_extraction_headline":"A preference alignment method called Macro improves the validity of multilingual self-generated counterfactual explanations by 12.55 percent on average while maintaining minimality."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.11632/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-21T00:31:32.115026Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-20T14:16:48.159725Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-20T04:02:00.401849Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T11:40:41.514511Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"c56bffa2b253209226b768852087a08432da90077ba9ddd798026b1ffe2beb6d"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}