{"paper":{"title":"Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"A latent mediation method using sparse autoencoders delivers reliable token-level influence attribution for LLM predictions on any task.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Kyra Gan, Promit Ghosal, Shixing Yu","submitted_at":"2026-05-12T23:01:29Z","abstract_excerpt":"A critical step for reliable large language models (LLMs) use in healthcare is to attribute predictions to their training data, akin to a medical case study. This requires token-level precision: pinpointing not just which training examples influence a decision, but which tokens within them are responsible. While influence functions offer a principled framework for this, prior work is restricted to autoregressive settings and relies on an implicit assumption of token independence, rendering their identified influences unreliable. We introduce a flexible framework that infers token-level influen"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We introduce a flexible framework that infers token-level influence through a latent mediation approach for general prediction tasks... Token-level influence is obtained by propagating latent attributions back to the input space via token activation patterns.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that sparse autoencoders attached to LLM layers learn a basis of approximately independent latent features whose influence can be accurately propagated via Jacobian-vector products without introducing new biases.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A latent mediation method using sparse autoencoders delivers reliable token-level influence attribution for LLM predictions on any task.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"35dbf689b3acf125df8d4bb6ccbb9ef6005c9f1f5168ab2d2995135279e05403"},"source":{"id":"2605.12809","kind":"arxiv","version":1},"verdict":{"id":"d614feb7-4442-45a3-bb37-9da913b1d86f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:14:06.178610Z","strongest_claim":"We introduce a flexible framework that infers token-level influence through a latent mediation approach for general prediction tasks... Token-level influence is obtained by propagating latent attributions back to the input space via token activation patterns.","one_line_summary":"A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that sparse autoencoders attached to LLM layers learn a basis of approximately independent latent features whose influence can be accurately propagated via Jacobian-vector products without introducing new biases.","pith_extraction_headline":"A latent mediation method using sparse autoencoders delivers reliable token-level influence attribution for LLM predictions on any task."},"references":{"count":300,"sample":[{"doi":"","year":null,"title":"International Conference on Learning Representations , year=","work_id":"f975b9b8-13c7-4593-9830-1887eee9d305","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2013,"title":"Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=","work_id":"dd6121d7-f0e7-43ad-9da1-25608f600c48","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"Philosophical transactions of the royal society A: Mathematical, Physical and Engineering Sciences , volume=","work_id":"41da284d-638e-40b8-bf3e-19af20c218ed","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"arXiv preprint arXiv:2411.07618 , year=","work_id":"c0e4d0c8-7272-40d3-83e2-44de8b252f9f","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Transformer Circuits Thread , volume=","work_id":"bb66d71e-4735-49aa-af5f-610b0164a701","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":300,"snapshot_sha256":"147f2defe4e18329597fa089ffd4c2c7fe5e3b8ef3f87dc2d21434a56197bc3f","internal_anchors":50},"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"}