{"paper":{"title":"The Platonic Representation Hypothesis","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Representations learned by different neural networks are converging toward a shared statistical model of reality.","cross_cats":["cs.AI","cs.CV","cs.NE"],"primary_cat":"cs.LG","authors_text":"Brian Cheung, Minyoung Huh, Phillip Isola, Tongzhou Wang","submitted_at":"2024-05-13T17:58:30Z","abstract_excerpt":"We argue that representations in AI models, particularly deep networks, are converging. First, we survey many examples of convergence in the literature: over time and across multiple domains, the ways by which different neural networks represent data are becoming more aligned. Next, we demonstrate convergence across data modalities: as vision models and language models get larger, they measure distance between datapoints in a more and more alike way. We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato's concept of an ideal reality. We ter"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato's concept of an ideal reality. We term such a representation the platonic representation.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That observed similarities in representation spaces reflect convergence to an objective underlying model of reality rather than shared inductive biases, training data overlap, or architectural similarities.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Representations learned by large AI models are converging toward a shared statistical model of reality.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Representations learned by different neural networks are converging toward a shared statistical model of reality.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"9614a7ec0c066c07f3d25e426453e0d680af5b23a9b18f3cb70ba84788696f54"},"source":{"id":"2405.07987","kind":"arxiv","version":5},"verdict":{"id":"39eb710b-4ad3-45fa-af51-5756119544b8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T05:58:39.603476Z","strongest_claim":"We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato's concept of an ideal reality. We term such a representation the platonic representation.","one_line_summary":"Representations learned by large AI models are converging toward a shared statistical model of reality.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That observed similarities in representation spaces reflect convergence to an objective underlying model of reality rather than shared inductive biases, training data overlap, or architectural similarities.","pith_extraction_headline":"Representations learned by different neural networks are converging toward a shared statistical model of reality."},"references":{"count":272,"sample":[{"doi":"","year":null,"title":"Cognitive Systems Research , volume =","work_id":"2c395a6c-6e7f-4254-9ad1-4bc5f08b6ca0","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Communications of the ACM , volume=","work_id":"6ceafbf5-fa63-4493-b4f6-a0298754841e","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Nature communications , volume=","work_id":"3c961a44-3a00-4a0a-94c8-de2554fa7702","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Text and Code Embeddings by Contrastive Pre-Training","work_id":"b5a7ebcc-85b7-4a3f-8495-8d4b5220f949","ref_index":4,"cited_arxiv_id":"2201.10005","is_internal_anchor":true},{"doi":"","year":null,"title":"Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=","work_id":"d199e356-1cd3-4a1e-8c3c-b04d6cbed803","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":272,"snapshot_sha256":"c6d4ca73fedde0da0b5184c9930ab6e7bbd10bcae37532c1b337107d6b28ee3c","internal_anchors":38},"formal_canon":{"evidence_count":2,"snapshot_sha256":"3529eecbb8053c09d21662f7a80ab5d8b1be84d33c9402f3378db5598ac53d95"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}