{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:LY6PQ5M3CSJK5Q4JACPDVWJQBF","short_pith_number":"pith:LY6PQ5M3","schema_version":"1.0","canonical_sha256":"5e3cf8759b1492aec389009e3ad9300975d9cb2a8113ea8cf86e031f303fe230","source":{"kind":"arxiv","id":"2601.21731","version":2},"attestation_state":"computed","paper":{"title":"Mechanistic Evidence for Spectral Structures in Prior-Data Fitted Networks","license":"http://creativecommons.org/licenses/by/4.0/","headline":"PFNs encode spectral information in attention scores that is causally used for predictions and extractable as explicit kernels.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Kaustubh Sharma, Ojasva Nema, Parikshit Pareek, Srijan Tiwari","submitted_at":"2026-01-29T13:51:26Z","abstract_excerpt":"Prior-Data Fitted Networks (PFNs) enable amortized Bayesian inference in a single forward pass, yet their internal representations remain opaque. It is unknown whether PFNs encode identifiable Bayesian structure or merely memorize input-output mappings. We provide mechanistic evidence that PFNs learn structured spectral representations and that these can be extracted as explicit kernels. First, probing experiments across three architectures, including the publicly released TabPFN, show that spectral information is linearly decodable from the latent attention score and organized along a dominan"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2601.21731","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-01-29T13:51:26Z","cross_cats_sorted":[],"title_canon_sha256":"42153520386bf76a0bd4e24f28e20ccc51f0c504cebf39a24a2524f09f5730f7","abstract_canon_sha256":"5fbf04465a93c6ba16c6d132c13f342ab55f5a780b29a04ded37b2e7d4b731f6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:09:24.118817Z","signature_b64":"OwRvxKQtcFp/jJApGZKejUhx3Ab4YX6+y8f0VD7+Pk35FduCcgFsPzfmCV/lwb9Cdb5r5bpfAiZI7+mup6QyAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5e3cf8759b1492aec389009e3ad9300975d9cb2a8113ea8cf86e031f303fe230","last_reissued_at":"2026-05-18T03:09:24.118112Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:09:24.118112Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mechanistic Evidence for Spectral Structures in Prior-Data Fitted Networks","license":"http://creativecommons.org/licenses/by/4.0/","headline":"PFNs encode spectral information in attention scores that is causally used for predictions and extractable as explicit kernels.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Kaustubh Sharma, Ojasva Nema, Parikshit Pareek, Srijan Tiwari","submitted_at":"2026-01-29T13:51:26Z","abstract_excerpt":"Prior-Data Fitted Networks (PFNs) enable amortized Bayesian inference in a single forward pass, yet their internal representations remain opaque. It is unknown whether PFNs encode identifiable Bayesian structure or merely memorize input-output mappings. We provide mechanistic evidence that PFNs learn structured spectral representations and that these can be extracted as explicit kernels. First, probing experiments across three architectures, including the publicly released TabPFN, show that spectral information is linearly decodable from the latent attention score and organized along a dominan"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Probing, activation patching, and subspace interventions establish that spectral information is linearly decodable from PFN latent attention scores, causally used for prediction, concentrated in a low-dimensional subspace, and extractable via a Filter Bank Decoder as explicit stationary kernels that support competitive GP regression in a single forward pass.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the linearly decodable spectral directions identified by probing and interventions are the actual mechanism driving the PFN's Bayesian predictions rather than a correlated side effect of training on continuous regression tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PFNs learn linearly decodable spectral information in attention latents that is causally used for prediction and extractable as explicit kernels via a Filter Bank Decoder supporting competitive one-pass GP regression.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"PFNs encode spectral information in attention scores that is causally used for predictions and extractable as explicit kernels.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"a3aea810cf91aff976903b6b9a4622bb5f5254bd18e76ef6a2fa9be6391ee1f0"},"source":{"id":"2601.21731","kind":"arxiv","version":2},"verdict":{"id":"12e4a26b-ebf3-4c0e-b092-261ef0baa0cc","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T10:18:29.920562Z","strongest_claim":"Probing, activation patching, and subspace interventions establish that spectral information is linearly decodable from PFN latent attention scores, causally used for prediction, concentrated in a low-dimensional subspace, and extractable via a Filter Bank Decoder as explicit stationary kernels that support competitive GP regression in a single forward pass.","one_line_summary":"PFNs learn linearly decodable spectral information in attention latents that is causally used for prediction and extractable as explicit kernels via a Filter Bank Decoder supporting competitive one-pass GP regression.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the linearly decodable spectral directions identified by probing and interventions are the actual mechanism driving the PFN's Bayesian predictions rather than a correlated side effect of training on continuous regression tasks.","pith_extraction_headline":"PFNs encode spectral information in attention scores that is causally used for predictions and extractable as explicit kernels."},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2601.21731","created_at":"2026-05-18T03:09:24.118218+00:00"},{"alias_kind":"arxiv_version","alias_value":"2601.21731v2","created_at":"2026-05-18T03:09:24.118218+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2601.21731","created_at":"2026-05-18T03:09:24.118218+00:00"},{"alias_kind":"pith_short_12","alias_value":"LY6PQ5M3CSJK","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"LY6PQ5M3CSJK5Q4J","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"LY6PQ5M3","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LY6PQ5M3CSJK5Q4JACPDVWJQBF","json":"https://pith.science/pith/LY6PQ5M3CSJK5Q4JACPDVWJQBF.json","graph_json":"https://pith.science/api/pith-number/LY6PQ5M3CSJK5Q4JACPDVWJQBF/graph.json","events_json":"https://pith.science/api/pith-number/LY6PQ5M3CSJK5Q4JACPDVWJQBF/events.json","paper":"https://pith.science/paper/LY6PQ5M3"},"agent_actions":{"view_html":"https://pith.science/pith/LY6PQ5M3CSJK5Q4JACPDVWJQBF","download_json":"https://pith.science/pith/LY6PQ5M3CSJK5Q4JACPDVWJQBF.json","view_paper":"https://pith.science/paper/LY6PQ5M3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2601.21731&json=true","fetch_graph":"https://pith.science/api/pith-number/LY6PQ5M3CSJK5Q4JACPDVWJQBF/graph.json","fetch_events":"https://pith.science/api/pith-number/LY6PQ5M3CSJK5Q4JACPDVWJQBF/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LY6PQ5M3CSJK5Q4JACPDVWJQBF/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LY6PQ5M3CSJK5Q4JACPDVWJQBF/action/storage_attestation","attest_author":"https://pith.science/pith/LY6PQ5M3CSJK5Q4JACPDVWJQBF/action/author_attestation","sign_citation":"https://pith.science/pith/LY6PQ5M3CSJK5Q4JACPDVWJQBF/action/citation_signature","submit_replication":"https://pith.science/pith/LY6PQ5M3CSJK5Q4JACPDVWJQBF/action/replication_record"}},"created_at":"2026-05-18T03:09:24.118218+00:00","updated_at":"2026-05-18T03:09:24.118218+00:00"}