{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:XFEMU7DJWV2HNNVHAYGKXNFR3Y","short_pith_number":"pith:XFEMU7DJ","schema_version":"1.0","canonical_sha256":"b948ca7c69b57476b6a7060cabb4b1de335261c2e720089c36947842b60d90e8","source":{"kind":"arxiv","id":"2606.21906","version":1},"attestation_state":"computed","paper":{"title":"Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"An Yang, Chujie Zheng, Dayiheng Liu, Fei Huang, Gao Huang, Jingren Zhou, Sean Du, Sining Zhoubian, Tianyi Tang, Xuanming Zhang, Yuxuan Chen","submitted_at":"2026-06-20T07:03:26Z","abstract_excerpt":"Autoregressive generation in large language models (LLMs) conventionally decodes from the final layer, assuming that deeper representations yield more reliable next-token predictions. We revisit this assumption by revealing a recurring Guess-Refine-Perturb dynamic: early layers form coarse guesses, intermediate layers refine reasoning-relevant semantics, and final layers can perturb these refined predictions toward generic or alignment-preferred tokens. We introduce Confident Decoding, a training-free decoding strategy that dynamically selects the most reliable near-final layer through entropy"},"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":"2606.21906","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-06-20T07:03:26Z","cross_cats_sorted":[],"title_canon_sha256":"edec682b53e8891fd1c9845d245eb96895c1371bc3bda8f1b3f39a8246f4a235","abstract_canon_sha256":"d02ad61dc1bbd72532a753ddf5608cb472eef4c2f1e29255034555cdc0d3532f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-23T02:13:02.528965Z","signature_b64":"WaGdvxn1eqZwi3Tvaff4bt1uCY+9lumgyCIkUZ3IyCEai+6jEGGAmyB3KVw7tz3aCBpcfDQuVWo+AhImhn1GAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b948ca7c69b57476b6a7060cabb4b1de335261c2e720089c36947842b60d90e8","last_reissued_at":"2026-06-23T02:13:02.528565Z","signature_status":"signed_v1","first_computed_at":"2026-06-23T02:13:02.528565Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"An Yang, Chujie Zheng, Dayiheng Liu, Fei Huang, Gao Huang, Jingren Zhou, Sean Du, Sining Zhoubian, Tianyi Tang, Xuanming Zhang, Yuxuan Chen","submitted_at":"2026-06-20T07:03:26Z","abstract_excerpt":"Autoregressive generation in large language models (LLMs) conventionally decodes from the final layer, assuming that deeper representations yield more reliable next-token predictions. We revisit this assumption by revealing a recurring Guess-Refine-Perturb dynamic: early layers form coarse guesses, intermediate layers refine reasoning-relevant semantics, and final layers can perturb these refined predictions toward generic or alignment-preferred tokens. We introduce Confident Decoding, a training-free decoding strategy that dynamically selects the most reliable near-final layer through entropy"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.21906","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2606.21906/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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2606.21906","created_at":"2026-06-23T02:13:02.528624+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.21906v1","created_at":"2026-06-23T02:13:02.528624+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.21906","created_at":"2026-06-23T02:13:02.528624+00:00"},{"alias_kind":"pith_short_12","alias_value":"XFEMU7DJWV2H","created_at":"2026-06-23T02:13:02.528624+00:00"},{"alias_kind":"pith_short_16","alias_value":"XFEMU7DJWV2HNNVH","created_at":"2026-06-23T02:13:02.528624+00:00"},{"alias_kind":"pith_short_8","alias_value":"XFEMU7DJ","created_at":"2026-06-23T02:13:02.528624+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/XFEMU7DJWV2HNNVHAYGKXNFR3Y","json":"https://pith.science/pith/XFEMU7DJWV2HNNVHAYGKXNFR3Y.json","graph_json":"https://pith.science/api/pith-number/XFEMU7DJWV2HNNVHAYGKXNFR3Y/graph.json","events_json":"https://pith.science/api/pith-number/XFEMU7DJWV2HNNVHAYGKXNFR3Y/events.json","paper":"https://pith.science/paper/XFEMU7DJ"},"agent_actions":{"view_html":"https://pith.science/pith/XFEMU7DJWV2HNNVHAYGKXNFR3Y","download_json":"https://pith.science/pith/XFEMU7DJWV2HNNVHAYGKXNFR3Y.json","view_paper":"https://pith.science/paper/XFEMU7DJ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.21906&json=true","fetch_graph":"https://pith.science/api/pith-number/XFEMU7DJWV2HNNVHAYGKXNFR3Y/graph.json","fetch_events":"https://pith.science/api/pith-number/XFEMU7DJWV2HNNVHAYGKXNFR3Y/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XFEMU7DJWV2HNNVHAYGKXNFR3Y/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XFEMU7DJWV2HNNVHAYGKXNFR3Y/action/storage_attestation","attest_author":"https://pith.science/pith/XFEMU7DJWV2HNNVHAYGKXNFR3Y/action/author_attestation","sign_citation":"https://pith.science/pith/XFEMU7DJWV2HNNVHAYGKXNFR3Y/action/citation_signature","submit_replication":"https://pith.science/pith/XFEMU7DJWV2HNNVHAYGKXNFR3Y/action/replication_record"}},"created_at":"2026-06-23T02:13:02.528624+00:00","updated_at":"2026-06-23T02:13:02.528624+00:00"}