{"paper":{"title":"RAT+: Train Dense, Infer Sparse -- Recurrence Augmented Attention for Dilated Inference","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"RAT+ lets one densely pretrained model switch to dilated sparse attention at inference with only short adaptation.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Caglar Gulcehre, Xiuying Wei","submitted_at":"2026-02-20T13:09:49Z","abstract_excerpt":"Structured dilated attention has an appealing inference-time efficiency knob: it reduces the FLOPs of attention and the KV cache size by a factor of the dilation size D, while preserving long-range connectivity. While prior work studies it by training each configuration from scratch, directly sparsifying a pretrained attention model into a dilated pattern leads to severe accuracy degradation, preventing flexible reuse across inference scenarios. We introduce RAT+, a dense-pretraining architecture that augments attention with full-sequence recurrence and active recurrence learning. A single RAT"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"A single RAT+ model is pretrained densely once and can then be flexibly switched at inference time to dilated attention (optionally with local windows) or hybrid layer/head compositions, requiring only a short 1B-token resolution adaptation rather than retraining separate sparse models.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That adding full-sequence recurrence and active recurrence learning during dense pretraining creates representations that transfer to dilated sparse patterns with only short adaptation and limited accuracy loss.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"RAT+ pretrains a single dense recurrent-augmented attention model that supports flexible dilated sparse inference after short adaptation, matching dense accuracy at moderate dilation and losing only 1-3 points at high dilation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"RAT+ lets one densely pretrained model switch to dilated sparse attention at inference with only short adaptation.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6c02010c9b714beb8dd8d24366b528de6c1b16a7bcc888b9ae1fadc04da7b1e6"},"source":{"id":"2602.18196","kind":"arxiv","version":4},"verdict":{"id":"2b217652-3625-4931-8ddd-6ce0045843c6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T20:56:31.911959Z","strongest_claim":"A single RAT+ model is pretrained densely once and can then be flexibly switched at inference time to dilated attention (optionally with local windows) or hybrid layer/head compositions, requiring only a short 1B-token resolution adaptation rather than retraining separate sparse models.","one_line_summary":"RAT+ pretrains a single dense recurrent-augmented attention model that supports flexible dilated sparse inference after short adaptation, matching dense accuracy at moderate dilation and losing only 1-3 points at high dilation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That adding full-sequence recurrence and active recurrence learning during dense pretraining creates representations that transfer to dilated sparse patterns with only short adaptation and limited accuracy loss.","pith_extraction_headline":"RAT+ lets one densely pretrained model switch to dilated sparse attention at inference with only short adaptation."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2602.18196/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"}