{"paper":{"title":"Post-Training Language Models for Crosslingual Consistency","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Direct Consistency Optimization derives a reward from the LLM itself to produce consistent answers to the same question across languages.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Arianna Bisazza, Jirui Qi, Mrinmaya Sachan, Raquel Fern\\'andez, Ryan Cotterell, Tianyu Liu","submitted_at":"2026-03-04T23:36:55Z","abstract_excerpt":"Language models often respond inconsistently to translation-equivalent prompts across languages, undermining the reliability of multilingual systems. To quantify this, we give an information-theoretic definition of crosslingual consistency as a divergence bound between a model's response distribution and its round-trip pushforward across languages. We then introduce penalized consistency optimization (PCO), a post-training procedure that couples this divergence with a Kullback-Leibler penalty to a fixed reference language model. Because direct optimization of PCO requires expensive on-policy r"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"DCO significantly improves crosslingual consistency across diverse LLMs and outperforms existing methods when training with samples of multiple languages, while complementing DPO when gold labels are available.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a structured reward function derived directly from the LLM itself can reliably enforce true crosslingual knowledge consistency without introducing new inconsistencies or biases during optimization.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DCO optimizes LLMs for crosslingual knowledge consistency via a structured reward derived directly from the model itself, outperforming prior methods in experiments.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Direct Consistency Optimization derives a reward from the LLM itself to produce consistent answers to the same question across languages.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"23b41dd4fd1d4743d6821102c7648dea214a93948081ab4b459126d9f4079a7f"},"source":{"id":"2603.04678","kind":"arxiv","version":3},"verdict":{"id":"61f392d0-127c-43ad-b170-703d36fdd0a3","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T15:56:49.533431Z","strongest_claim":"DCO significantly improves crosslingual consistency across diverse LLMs and outperforms existing methods when training with samples of multiple languages, while complementing DPO when gold labels are available.","one_line_summary":"DCO optimizes LLMs for crosslingual knowledge consistency via a structured reward derived directly from the model itself, outperforming prior methods in experiments.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a structured reward function derived directly from the LLM itself can reliably enforce true crosslingual knowledge consistency without introducing new inconsistencies or biases during optimization.","pith_extraction_headline":"Direct Consistency Optimization derives a reward from the LLM itself to produce consistent answers to the same question across languages."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2603.04678/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"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"}