{"paper":{"title":"What Structural Inductive Bias Helps Transformers Reason Over Knowledge Graphs? A Study with Tabula RASA","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Sparse adjacency masking alone supplies the main inductive bias that lets transformers perform multi-hop reasoning over knowledge graphs.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Camilla Mazzoleni, Federico Martelli, Gian-Alessandro Lombardi, Jonas Petersen, Riccardo Maggioni","submitted_at":"2026-02-02T21:35:39Z","abstract_excerpt":"What structural inductive bias helps transformers reason over knowledge graphs? Through controlled ablations of a minimal transformer modification with four independently removable components (sparse adjacency masking, edge-type biases, query scaling, value gating), we isolate which structural signals drive multi-hop reasoning. Our finding is sharp: sparse adjacency masking alone accounts for the dominant share of improvement over unmasked transformers (+72.5pp on 3-hop MetaQA, +45.5pp on WebQSP, +53.9pp on CWQ), while learned relation parameters add only modest refinement and can actively hur"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"sparse adjacency masking alone accounts for the dominant share of improvement over unmasked transformers (+72.5pp on 3-hop MetaQA, +45.5pp on WebQSP, +53.9pp on CWQ)","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The four components can be removed independently without introducing implementation-specific interactions that confound the measured contributions of each.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Sparse adjacency masking alone explains the bulk of transformer gains on multi-hop KGQA tasks, showing that topological structure matters far more than relation-specific parameters.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Sparse adjacency masking alone supplies the main inductive bias that lets transformers perform multi-hop reasoning over knowledge graphs.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c7dc286a33c4f737d6afbbebade15dd427879e9455dc357fdd44cb8c832ac7d8"},"source":{"id":"2602.02834","kind":"arxiv","version":4},"verdict":{"id":"f13561b3-dde1-42da-9a34-7109c06b4935","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T07:56:17.606894Z","strongest_claim":"sparse adjacency masking alone accounts for the dominant share of improvement over unmasked transformers (+72.5pp on 3-hop MetaQA, +45.5pp on WebQSP, +53.9pp on CWQ)","one_line_summary":"Sparse adjacency masking alone explains the bulk of transformer gains on multi-hop KGQA tasks, showing that topological structure matters far more than relation-specific parameters.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The four components can be removed independently without introducing implementation-specific interactions that confound the measured contributions of each.","pith_extraction_headline":"Sparse adjacency masking alone supplies the main inductive bias that lets transformers perform multi-hop reasoning over knowledge graphs."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2602.02834/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":2,"snapshot_sha256":"e466d27186fbd3d41566d28962c5d73f48203f676d237cf2482d41f77efbe9d4"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}