[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-a-new-way-to-align-ai-models-across-languages-via-their-experts":10,"sections":35},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},9771,"a-new-way-to-align-ai-models-across-languages-via-their-experts","A New Way to Align AI Models Across Languages, Via Their Experts","A new technique aligns AI models' mixture-of-experts routers across languages instead of hidden states, and it boosts multilingual performance in early tests.","A new training trick helps AI language models treat the same idea consistently across languages by comparing which internal experts handle each language, not raw neural activity.\n\nResearchers built a method for cross-lingual alignment in decoder-only large language models that use a mixture-of-experts (MoE) architecture, where specialized subnetworks called experts process different inputs. Aligning meaning across languages in these models is normally hard because tokenization splits the same sentence into different pieces depending on the language, so comparing raw hidden states token by token is unreliable. Instead, the team matched the outputs of the MoE router - the component that decides which experts handle which tokens - since those outputs can be pooled across a whole sequence rather than compared token by token. In continual pre-training experiments on four open-source MoE models, this routing-based alignment loss also pulled the underlying hidden representations closer together across languages and improved scores on the researchers' own multilingual evaluation suite.\n\nMost multilingual alignment techniques were built for encoder models, where sentence representations across languages are comparatively easy to compare directly. Decoder-only LLMs - the architecture behind nearly every modern chatbot - have resisted that approach because of tokenization mismatches, so this result extends a known trick to the mixture-of-experts systems that already power several open-weight model families.\n\nIt's four models and one custom benchmark, not a shipped product feature, so the fair read is a promising early signal, not a solved problem.","[\"multilingual-ai\",\"mixture-of-experts\",\"llm-training\",\"ai-research\"]","2026-10-02T04:00:00.000Z","2026-10-03T09:50:25.653Z","2026-10-03T09:50:31.344Z","published",null,[],"ai",[26,27,28,29],"multilingual-ai","mixture-of-experts","llm-training","ai-research",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.01921",0,{"sections":36},[37,40,44,48,53,57,61,66,71,76,81,86,91,96],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",6058,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",848,{"name":45,"slug":46,"count":47,"latest_published_at":18},"Policy","policy",439,{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",317,"2026-10-01T22:00:00.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":18},"Hardware","hardware",199,{"name":58,"slug":59,"count":60,"latest_published_at":18},"Science","science",176,{"name":62,"slug":63,"count":64,"latest_published_at":65},"Consumer Tech","consumer-tech",155,"2026-10-01T19:54:10.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Dev Tools","dev-tools",96,"2026-10-01T16:57:03.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Software","software",93,"2026-09-30T21:41:11.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",90,"2026-10-01T21:55:22.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",50,"2026-09-30T21:37:54.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]