{"paper":{"title":"RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"RoPE-based attention loses locality bias and token relevance consistency as context length grows, with failure probability approaching 0.5.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Aram Galstyan, Eliu A Huerta, Hao Peng, Minyang Tian, Phillip Harris, Srikanth Ronanki, Subendhu Rongali, Yufeng Du","submitted_at":"2026-05-15T01:16:16Z","abstract_excerpt":"We identify intrinsic limitations of Rotary Positional Embeddings (RoPE) in Transformer-based long-context language models. Our theoretical analysis abstracts away from the specific content of the context and depends only on its length. We prove that as context length increases, RoPE-based attention becomes unpredictable and loses two properties that are central to its effectiveness. First, it loses its locality bias: RoPE is no more likely to favor nearer positions than substantially farther ones. Second, it loses consistency in token relevance: a key vector that receives a higher attention s"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We prove that as context length increases, RoPE-based attention becomes unpredictable and loses its locality bias and consistency in token relevance, with the probability of failure approaching 0.5.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The theoretical analysis abstracts away from the specific content of the context and depends only on its length.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Proves that RoPE attention loses locality bias and token distinction in long contexts, approaching random behavior independent of content.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"RoPE-based attention loses locality bias and token relevance consistency as context length grows, with failure probability approaching 0.5.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b6baae0e772faef81fc052e87cc2c52be3a23dad539069ac6baa7526a0eaa034"},"source":{"id":"2605.15514","kind":"arxiv","version":1},"verdict":{"id":"24f6a120-bea3-469a-ac35-f047796697a4","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T15:29:44.440825Z","strongest_claim":"We prove that as context length increases, RoPE-based attention becomes unpredictable and loses its locality bias and consistency in token relevance, with the probability of failure approaching 0.5.","one_line_summary":"Proves that RoPE attention loses locality bias and token distinction in long contexts, approaching random behavior independent of content.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The theoretical analysis abstracts away from the specific content of the context and depends only on its length.","pith_extraction_headline":"RoPE-based attention loses locality bias and token relevance consistency as context length grows, with failure probability approaching 0.5."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.15514/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T16:01:17.893729Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T15:40:36.322934Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"cited_work_retraction","ran_at":"2026-05-19T14:51:54.755432Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T14:21:54.053291Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"shingle_duplication","ran_at":"2026-05-19T13:49:41.847002Z","status":"skipped","version":"0.1.0","findings_count":0},{"name":"citation_quote_validity","ran_at":"2026-05-19T13:49:41.385938Z","status":"skipped","version":"0.1.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T13:33:22.633228Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"bc7cfd4c2122fea322863127a0fa9246126f23702411932a9c86123cf8db369c"},"references":{"count":55,"sample":[{"doi":"","year":2021,"title":"RoFormer: Enhanced Transformer with Rotary Position Embedding , author=. 2021 , eprint=","work_id":"5fc636be-88c8-4c86-a321-9d1b76f63f13","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context , author=. 2024 , eprint=","work_id":"11af042d-6a6a-489e-9608-119eb98fa8f6","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Proceedings of the 41st International Conference on Machine Learning , articleno =","work_id":"98fe2912-080b-4d44-bb7c-7737e89f9d0b","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"2026 , institution =","work_id":"d1f73ff4-2701-4e73-acc7-f3f9f51430c6","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"International Conference on Learning Representations , year=","work_id":"ba153a1a-8735-43e0-a86c-f7c5a0b27705","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":55,"snapshot_sha256":"bf670194bf65bec3d315f50c230105ed6ae2bf9f6aa5d7c32ff0cbb5806bb531","internal_anchors":2},"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"}