{"paper":{"title":"Soft-TransFormers for Continual Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Task-specific multiplicative masks on attention projections enable continual learning in frozen pre-trained transformers with minimal added parameters.","cross_cats":["cs.AI","cs.CV"],"primary_cat":"cs.LG","authors_text":"Chang D. Yoo, Haeyong Kang","submitted_at":"2024-11-25T03:52:47Z","abstract_excerpt":"Inspired by the Well-initialized Lottery Ticket Hypothesis (WLTH), we introduce Soft-TransFormers (Soft-TF), a continual learning framework that adapts a frozen pre-trained Transformer through task-specific soft subnetworks: real-valued multiplicative masks over the query, key, value, and output projections of selected self-attention layers. The masks are initialized at one, so optimization starts exactly at the pre-trained solution, and mask-space gradient descent is intrinsically biased toward modulating the backbone's dominant pathways; we prove that, under standard convex-Lipschitz assumpt"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across multiple continual learning benchmarks, Soft-TF achieves state-of-the-art performance, consistently outperforming prompt-based, adapter-based, and LoRA-style baselines while requiring minimal additional parameters.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The central claim depends on the premise that task-specific multiplicative masks applied to the key, query, value, and output projections in self-attention, together with a lightweight dual-prompt mechanism, will enable smooth task adaptation while preserving shared representations in the frozen pre-trained transformer.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Soft-TF learns task-specific multiplicative masks on transformer self-attention layers combined with dual prompts to achieve SOTA continual learning performance with minimal added parameters.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Task-specific multiplicative masks on attention projections enable continual learning in frozen pre-trained transformers with minimal added parameters.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c4ff11323261ab90278de6465da1485fce8de7d6da76c6ef7ef88fdc6cd48316"},"source":{"id":"2411.16073","kind":"arxiv","version":4},"verdict":{"id":"6f61dbe1-c6ba-4b2a-b130-f5d32a313076","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-23T08:11:08.840576Z","strongest_claim":"Across multiple continual learning benchmarks, Soft-TF achieves state-of-the-art performance, consistently outperforming prompt-based, adapter-based, and LoRA-style baselines while requiring minimal additional parameters.","one_line_summary":"Soft-TF learns task-specific multiplicative masks on transformer self-attention layers combined with dual prompts to achieve SOTA continual learning performance with minimal added parameters.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The central claim depends on the premise that task-specific multiplicative masks applied to the key, query, value, and output projections in self-attention, together with a lightweight dual-prompt mechanism, will enable smooth task adaptation while preserving shared representations in the frozen pre-trained transformer.","pith_extraction_headline":"Task-specific multiplicative masks on attention projections enable continual learning in frozen pre-trained transformers with minimal added parameters."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2411.16073/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"}