{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:OUDWHWEUIP57XMCUMOAYY3MKL7","short_pith_number":"pith:OUDWHWEU","schema_version":"1.0","canonical_sha256":"750763d89443fbfbb05463818c6d8a5fc7b6f0892d3b00df9f89237a8b4050c9","source":{"kind":"arxiv","id":"2605.03052","version":2},"attestation_state":"computed","paper":{"title":"How Language Models Process Negation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Language models process negation by constructing representations of negative phrases more than by suppressing positives.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Jonathan May, Robin Jia, Tianyi Zhou, Zhejian Zhou","submitted_at":"2026-05-04T18:17:05Z","abstract_excerpt":"We study how Large Language Models (LLMs) process negation mechanistically. First, we establish that even though open-weight models often provide wrong answers to questions involving negation, they do possess internal components that process negation correctly. Their poor accuracy is due to late-layer attention behavior that promotes simple shortcuts; ablating those attention modules greatly improves accuracy on negation-related questions. Second, we uncover how models process negation. We consider two hypotheses: models could use attention heads that attend to the phrase being negated and sup"},"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":true},"canonical_record":{"source":{"id":"2605.03052","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-04T18:17:05Z","cross_cats_sorted":[],"title_canon_sha256":"d3761b2174bae54d4638480ef8a504bd600dacbb661389e838b37b4ae07ecbaa","abstract_canon_sha256":"3f7d4817f55118373e90a8f549904e01dab05f60d99547f0ba79e90118fd7d0c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-02T01:03:16.146511Z","signature_b64":"Ra4qtZxbvxQRhmDsC/lEoELUXmiDCQfNiGm7kOH7HBwNFiQ56xNfgBe7OgJ4mewwDPmpwnkdxONfN3DV6/CDBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"750763d89443fbfbb05463818c6d8a5fc7b6f0892d3b00df9f89237a8b4050c9","last_reissued_at":"2026-06-02T01:03:16.146064Z","signature_status":"signed_v1","first_computed_at":"2026-06-02T01:03:16.146064Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"How Language Models Process Negation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Language models process negation by constructing representations of negative phrases more than by suppressing positives.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Jonathan May, Robin Jia, Tianyi Zhou, Zhejian Zhou","submitted_at":"2026-05-04T18:17:05Z","abstract_excerpt":"We study how Large Language Models (LLMs) process negation mechanistically. First, we establish that even though open-weight models often provide wrong answers to questions involving negation, they do possess internal components that process negation correctly. Their poor accuracy is due to late-layer attention behavior that promotes simple shortcuts; ablating those attention modules greatly improves accuracy on negation-related questions. Second, we uncover how models process negation. We consider two hypotheses: models could use attention heads that attend to the phrase being negated and sup"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We apply a range of observational and causal interpretability techniques on Mistral-7B and Llama-3.1-8B to show that models implement both mechanisms, with the 'constructive' mechanism being more prominent.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the observational and causal interpretability techniques (such as attention ablation and activation analysis) accurately isolate and identify the specific mechanisms responsible for negation processing without confounding effects from other model components.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LLMs implement both attention-based suppression and constructive representations for negation, with construction dominant, despite poor accuracy from late-layer attention shortcuts.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Language models process negation by constructing representations of negative phrases more than by suppressing positives.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"4b4d986a153b4e4622dc8d6ba904d39aca1febc493f9916a8c23ff313813d177"},"source":{"id":"2605.03052","kind":"arxiv","version":2},"verdict":{"id":"25b7f7ce-73a9-48bb-a34a-6ca0d57f8800","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T18:25:01.470535Z","strongest_claim":"We apply a range of observational and causal interpretability techniques on Mistral-7B and Llama-3.1-8B to show that models implement both mechanisms, with the 'constructive' mechanism being more prominent.","one_line_summary":"LLMs implement both attention-based suppression and constructive representations for negation, with construction dominant, despite poor accuracy from late-layer attention shortcuts.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the observational and causal interpretability techniques (such as attention ablation and activation analysis) accurately isolate and identify the specific mechanisms responsible for negation processing without confounding effects from other model components.","pith_extraction_headline":"Language models process negation by constructing representations of negative phrases more than by suppressing positives."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.03052/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T14:38:55.614335Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-20T02:01:22.071726Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T15:45:24.319750Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"234feb544cad5e17319c2461e599326dd4a45404d68059479eae4fe58ca2c0c0"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"620e41cb351eb0f0ae9a0d56671e7dc6e556bb64be9a46408d0d25b851018262"},"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":"2605.03052","created_at":"2026-06-02T01:03:16.146110+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.03052v2","created_at":"2026-06-02T01:03:16.146110+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.03052","created_at":"2026-06-02T01:03:16.146110+00:00"},{"alias_kind":"pith_short_12","alias_value":"OUDWHWEUIP57","created_at":"2026-06-02T01:03:16.146110+00:00"},{"alias_kind":"pith_short_16","alias_value":"OUDWHWEUIP57XMCU","created_at":"2026-06-02T01:03:16.146110+00:00"},{"alias_kind":"pith_short_8","alias_value":"OUDWHWEU","created_at":"2026-06-02T01:03:16.146110+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OUDWHWEUIP57XMCUMOAYY3MKL7","json":"https://pith.science/pith/OUDWHWEUIP57XMCUMOAYY3MKL7.json","graph_json":"https://pith.science/api/pith-number/OUDWHWEUIP57XMCUMOAYY3MKL7/graph.json","events_json":"https://pith.science/api/pith-number/OUDWHWEUIP57XMCUMOAYY3MKL7/events.json","paper":"https://pith.science/paper/OUDWHWEU"},"agent_actions":{"view_html":"https://pith.science/pith/OUDWHWEUIP57XMCUMOAYY3MKL7","download_json":"https://pith.science/pith/OUDWHWEUIP57XMCUMOAYY3MKL7.json","view_paper":"https://pith.science/paper/OUDWHWEU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.03052&json=true","fetch_graph":"https://pith.science/api/pith-number/OUDWHWEUIP57XMCUMOAYY3MKL7/graph.json","fetch_events":"https://pith.science/api/pith-number/OUDWHWEUIP57XMCUMOAYY3MKL7/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OUDWHWEUIP57XMCUMOAYY3MKL7/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OUDWHWEUIP57XMCUMOAYY3MKL7/action/storage_attestation","attest_author":"https://pith.science/pith/OUDWHWEUIP57XMCUMOAYY3MKL7/action/author_attestation","sign_citation":"https://pith.science/pith/OUDWHWEUIP57XMCUMOAYY3MKL7/action/citation_signature","submit_replication":"https://pith.science/pith/OUDWHWEUIP57XMCUMOAYY3MKL7/action/replication_record"}},"created_at":"2026-06-02T01:03:16.146110+00:00","updated_at":"2026-06-02T01:03:16.146110+00:00"}