{"paper":{"title":"Are Compact Rationales Free? Measuring Tile Selection Headroom in Frozen WSI-MIL","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"FOCI reveals that compact rationales for frozen WSI-MIL predictions depend on the choice of backbone aggregator.","cross_cats":["cs.AI","cs.CV"],"primary_cat":"eess.IV","authors_text":"Hwiyoung Kim, Hyun Do Jung, Jungwon Choi, Soojung Choi, Yujin Oh","submitted_at":"2026-05-12T12:15:41Z","abstract_excerpt":"Whole-slide image (WSI) multiple instance learning (MIL) classifiers can achieve strong slide-level AUC while leaving the full-bag prediction opaque. Attention scores are widely reused as post-hoc explanations, but high attention can reflect aggregation preference rather than a compact, model-sufficient rationale. We study post-hoc rationale highlighting for frozen WSI-MIL: given a trained classifier, can its slide-level prediction be recovered from a compact, output-consistent tile subset without retraining the backbone? We instantiate this with Finding Optimal Contextual Instances (FOCI), a "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across three WSI benchmarks and seven MIL backbones, FOCI reveals that compact rationales are selection-headroom dependent: transformer and multi-branch attention aggregators can admit compact rationales, near-minimal attention-pooling baselines enter a selection-saturation regime, and hard-selection backbones can conflict with an external readout. For TransMIL, relative to its documented CLS-proxy ranking, FOCI reduces the Minimum Sufficient K (MSK) tile count by 32-56% across benchmarks, while ACMIL+FOCI attains the highest mean SHI (+0.465).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the sufficiency and exclusion objectives used to train FOCI produce tile subsets that are genuinely model-sufficient and free of readout-induced artifacts, and that the Sequential Reveal Protocol accurately isolates selection headroom without confounding effects from the specific keep/drop training procedure.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"FOCI adds a post-hoc readout to frozen WSI-MIL models to find compact output-consistent tile subsets and measures selection headroom with SHI, showing transformer-based models allow smaller rationales than attention-pooling baselines.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"FOCI reveals that compact rationales for frozen WSI-MIL predictions depend on the choice of backbone aggregator.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"56502401941b1f087f2f46c46f7a8a30549c1e1272395da7c0690c2672dd5952"},"source":{"id":"2605.12575","kind":"arxiv","version":1},"verdict":{"id":"d0d55ff1-8d7f-4010-90db-c7f5c4cef935","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:35:07.746936Z","strongest_claim":"Across three WSI benchmarks and seven MIL backbones, FOCI reveals that compact rationales are selection-headroom dependent: transformer and multi-branch attention aggregators can admit compact rationales, near-minimal attention-pooling baselines enter a selection-saturation regime, and hard-selection backbones can conflict with an external readout. For TransMIL, relative to its documented CLS-proxy ranking, FOCI reduces the Minimum Sufficient K (MSK) tile count by 32-56% across benchmarks, while ACMIL+FOCI attains the highest mean SHI (+0.465).","one_line_summary":"FOCI adds a post-hoc readout to frozen WSI-MIL models to find compact output-consistent tile subsets and measures selection headroom with SHI, showing transformer-based models allow smaller rationales than attention-pooling baselines.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the sufficiency and exclusion objectives used to train FOCI produce tile subsets that are genuinely model-sufficient and free of readout-induced artifacts, and that the Sequential Reveal Protocol accurately isolates selection headroom without confounding effects from the specific keep/drop training procedure.","pith_extraction_headline":"FOCI reveals that compact rationales for frozen WSI-MIL predictions depend on the choice of backbone aggregator."},"references":{"count":41,"sample":[{"doi":"","year":2019,"title":"Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J","work_id":"9c9d09a0-b465-413e-b9f4-4b6815b21e87","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Attention-based deep multiple instance learning","work_id":"8ef34d56-4419-46b3-ba9c-ec64c047299a","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Data-efficient and weakly supervised computational pathology on whole-slide images.Nature biomedical engineering, 5(6):555–570","work_id":"d10a268e-9baa-47ed-bd74-87bb2719f105","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Towards a general-purpose foundation model for computational pathology.Nature medicine, 30(3):850–862","work_id":"3630134d-7234-4fdb-a23f-689393e31801","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Deep learning for whole slide image analysis: an overview.Frontiers in medicine, 6:264, 2019","work_id":"2a5e3d11-c406-47b5-9eb8-ed322351536f","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":41,"snapshot_sha256":"89e2d24ff836abadcd54d43675b028df970358857dd063067defab1b13bf0d56","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"}