{"paper":{"title":"TTT3R: 3D Reconstruction as Test-Time Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Framing 3D reconstruction as test-time training yields a closed-form learning rate from alignment confidence that doubles global pose accuracy on long sequences.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Andreas Geiger, Anpei Chen, Xingyu Chen, Yue Chen, Yuliang Xiu","submitted_at":"2025-09-30T17:59:51Z","abstract_excerpt":"Modern Recurrent Neural Networks have become a competitive architecture for 3D reconstruction due to their linear-time complexity. However, their performance degrades significantly when applied beyond the training context length, revealing limited length generalization. In this work, we revisit the 3D reconstruction foundation models from a Test-Time Training perspective, framing their designs as an online learning problem. Building on this perspective, we leverage the alignment confidence between the memory state and incoming observations to derive a closed-form learning rate for memory updat"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"This training-free intervention, termed TTT3R, substantially improves length generalization, achieving a 2× improvement in global pose estimation over baselines, while operating at 20 FPS with just 6 GB of GPU memory to process thousands of images.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That alignment confidence between the memory state and incoming observations can be computed reliably and directly yields a closed-form learning rate that correctly balances retention of history with adaptation to new data without introducing instability or bias.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"TTT3R derives a closed-form learning rate from memory-observation alignment confidence to boost length generalization in RNN-based 3D reconstruction by 2x in global pose estimation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Framing 3D reconstruction as test-time training yields a closed-form learning rate from alignment confidence that doubles global pose accuracy on long sequences.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7c70ed132bc01dfcaafee50c054a19a61e361953e5e3792002f4c25e4cb22fa6"},"source":{"id":"2509.26645","kind":"arxiv","version":4},"verdict":{"id":"eca97549-12f1-47ef-b8be-f9650a31484c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T06:35:59.893810Z","strongest_claim":"This training-free intervention, termed TTT3R, substantially improves length generalization, achieving a 2× improvement in global pose estimation over baselines, while operating at 20 FPS with just 6 GB of GPU memory to process thousands of images.","one_line_summary":"TTT3R derives a closed-form learning rate from memory-observation alignment confidence to boost length generalization in RNN-based 3D reconstruction by 2x in global pose estimation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That alignment confidence between the memory state and incoming observations can be computed reliably and directly yields a closed-form learning rate that correctly balances retention of history with adaptation to new data without introducing instability or bias.","pith_extraction_headline":"Framing 3D reconstruction as test-time training yields a closed-form learning rate from alignment confidence that doubles global pose accuracy on long sequences."},"references":{"count":112,"sample":[{"doi":"","year":2010,"title":"Bundle adjustment in the large","work_id":"98511d9d-34e5-46fc-b399-cfed80b316d7","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2011,"title":"Building rome in a day.ACM Communications, 2011","work_id":"66df7549-0359-4b87-b4aa-d40cc0986570","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Cross-view completion models are zero-shot correspondence estimators","work_id":"1b5feeac-d1ba-40fc-b2af-03fce6a3e07e","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2008,"title":"Speeded-up robust features (surf).Computer vision and image understanding, 2008","work_id":"3620411b-df89-473c-ba6d-5c6184b527bb","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"xlstm: Extended long short-term memory","work_id":"44e150f9-d858-4df1-9dc2-1cae87bc20fc","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":112,"snapshot_sha256":"e0089e70acc92b31e4aaafba12806174d7da72bf50922815eebee91a3e593664","internal_anchors":18},"formal_canon":{"evidence_count":1,"snapshot_sha256":"a204f5ebae18cf3a9b3b7135dc56cb4c5a24e002eb0a68948a28b7e507b242d8"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}