{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:OH4WFENFACW37YY43ITAF4AG4H","short_pith_number":"pith:OH4WFENF","schema_version":"1.0","canonical_sha256":"71f96291a500adbfe31cda2602f006e1cce8ab9dfa051eb587115678e9f3ff46","source":{"kind":"arxiv","id":"2409.15268","version":3},"attestation_state":"computed","paper":{"title":"Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Benjamin Feuer, John P. Dickerson, Max Cembalest, Micah Goldblum, Raz Besaleli, Samuel Dooley, Sanjana Nambiar, Teresa Datta","submitted_at":"2024-09-23T17:58:07Z","abstract_excerpt":"The release of ChatGPT in November 2022 sparked an explosion of interest in post-training and an avalanche of new preference optimization (PO) methods. These methods claim superior alignment by virtue of better correspondence with human pairwise preferences, often measured by LLM-judges. In this work, we attempt to answer the following question -- do LLM-judge preferences translate to progress on other, more concrete metrics for alignment, and if not, why not? We define a concrete metric for alignment, and introduce SOS-Bench (Substance Outweighs Style Benchmark), which is to the best of our k"},"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":"2409.15268","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-09-23T17:58:07Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"dbbbe7741708e0d085370e17cd4f89e68b9afc32bf00c2523e9c6284958e74c3","abstract_canon_sha256":"167bcd80e6a65439adfabaa8e10924cafa19ac5ba6f4a789883ea233202fc565"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:05:44.998868Z","signature_b64":"iFPdXLbQ/QC4R9gj6t0eqpXhZMjvcGo52s9WJ2WdewLF2VYkWHCr3GmhUDADgWwdzXycOc6RsGd4K3A/rUGBBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"71f96291a500adbfe31cda2602f006e1cce8ab9dfa051eb587115678e9f3ff46","last_reissued_at":"2026-07-05T10:05:44.998367Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:05:44.998367Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Benjamin Feuer, John P. Dickerson, Max Cembalest, Micah Goldblum, Raz Besaleli, Samuel Dooley, Sanjana Nambiar, Teresa Datta","submitted_at":"2024-09-23T17:58:07Z","abstract_excerpt":"The release of ChatGPT in November 2022 sparked an explosion of interest in post-training and an avalanche of new preference optimization (PO) methods. These methods claim superior alignment by virtue of better correspondence with human pairwise preferences, often measured by LLM-judges. In this work, we attempt to answer the following question -- do LLM-judge preferences translate to progress on other, more concrete metrics for alignment, and if not, why not? We define a concrete metric for alignment, and introduce SOS-Bench (Substance Outweighs Style Benchmark), which is to the best of our k"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2409.15268","kind":"arxiv","version":3},"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/2409.15268/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":"2409.15268","created_at":"2026-07-05T10:05:44.998428+00:00"},{"alias_kind":"arxiv_version","alias_value":"2409.15268v3","created_at":"2026-07-05T10:05:44.998428+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2409.15268","created_at":"2026-07-05T10:05:44.998428+00:00"},{"alias_kind":"pith_short_12","alias_value":"OH4WFENFACW3","created_at":"2026-07-05T10:05:44.998428+00:00"},{"alias_kind":"pith_short_16","alias_value":"OH4WFENFACW37YY4","created_at":"2026-07-05T10:05:44.998428+00:00"},{"alias_kind":"pith_short_8","alias_value":"OH4WFENF","created_at":"2026-07-05T10:05:44.998428+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2604.10784","citing_title":"TorchUMM: A Unified Multimodal Model Codebase for Evaluation, Analysis, and Post-training","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10784","citing_title":"TorchUMM: A Unified Multimodal Model Codebase for Evaluation, Analysis, and Post-training","ref_index":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OH4WFENFACW37YY43ITAF4AG4H","json":"https://pith.science/pith/OH4WFENFACW37YY43ITAF4AG4H.json","graph_json":"https://pith.science/api/pith-number/OH4WFENFACW37YY43ITAF4AG4H/graph.json","events_json":"https://pith.science/api/pith-number/OH4WFENFACW37YY43ITAF4AG4H/events.json","paper":"https://pith.science/paper/OH4WFENF"},"agent_actions":{"view_html":"https://pith.science/pith/OH4WFENFACW37YY43ITAF4AG4H","download_json":"https://pith.science/pith/OH4WFENFACW37YY43ITAF4AG4H.json","view_paper":"https://pith.science/paper/OH4WFENF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2409.15268&json=true","fetch_graph":"https://pith.science/api/pith-number/OH4WFENFACW37YY43ITAF4AG4H/graph.json","fetch_events":"https://pith.science/api/pith-number/OH4WFENFACW37YY43ITAF4AG4H/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OH4WFENFACW37YY43ITAF4AG4H/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OH4WFENFACW37YY43ITAF4AG4H/action/storage_attestation","attest_author":"https://pith.science/pith/OH4WFENFACW37YY43ITAF4AG4H/action/author_attestation","sign_citation":"https://pith.science/pith/OH4WFENFACW37YY43ITAF4AG4H/action/citation_signature","submit_replication":"https://pith.science/pith/OH4WFENFACW37YY43ITAF4AG4H/action/replication_record"}},"created_at":"2026-07-05T10:05:44.998428+00:00","updated_at":"2026-07-05T10:05:44.998428+00:00"}