{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:TAWQTPNYV3AMHTCHDSVN6WNT37","short_pith_number":"pith:TAWQTPNY","schema_version":"1.0","canonical_sha256":"982d09bdb8aec0c3cc471caadf59b3dfef91a9052a4c4837ee5a4bb5d5af54ea","source":{"kind":"arxiv","id":"2507.21141","version":1},"attestation_state":"computed","paper":{"title":"The Geometry of Harmfulness in LLMs through Subconcept Probing","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Adhitya Rajendra Kumar, Kevin Zhu, McNair Shah, Naitik Chheda, Saleena Angeline, Sean O'Brien, Vasu Sharma, Will Cai","submitted_at":"2025-07-23T07:56:05Z","abstract_excerpt":"Recent advances in large language models (LLMs) have intensified the need to understand and reliably curb their harmful behaviours. We introduce a multidimensional framework for probing and steering harmful content in model internals. For each of 55 distinct harmfulness subconcepts (e.g., racial hate, employment scams, weapons), we learn a linear probe, yielding 55 interpretable directions in activation space. Collectively, these directions span a harmfulness subspace that we show is strikingly low-rank. We then test ablation of the entire subspace from model internals, as well as steering and"},"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":"2507.21141","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-07-23T07:56:05Z","cross_cats_sorted":[],"title_canon_sha256":"2460229338a306cbb6e86695ea4031373304a42bbce55f597dee9a3a0ac6c727","abstract_canon_sha256":"79e00a97f82f0b38e2c27b11d25c228c033cf5b659de1245806df1d6efca7566"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:44:32.094951Z","signature_b64":"iJhDjNFDq8cp01u5t6tM3HeDBpykP8Tkhtg1Z7psUu/I9EG1G7kwsODDIizyK1jiMqQW724Y35NZLWU721gtCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"982d09bdb8aec0c3cc471caadf59b3dfef91a9052a4c4837ee5a4bb5d5af54ea","last_reissued_at":"2026-07-05T11:44:32.094454Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:44:32.094454Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Geometry of Harmfulness in LLMs through Subconcept Probing","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Adhitya Rajendra Kumar, Kevin Zhu, McNair Shah, Naitik Chheda, Saleena Angeline, Sean O'Brien, Vasu Sharma, Will Cai","submitted_at":"2025-07-23T07:56:05Z","abstract_excerpt":"Recent advances in large language models (LLMs) have intensified the need to understand and reliably curb their harmful behaviours. We introduce a multidimensional framework for probing and steering harmful content in model internals. For each of 55 distinct harmfulness subconcepts (e.g., racial hate, employment scams, weapons), we learn a linear probe, yielding 55 interpretable directions in activation space. Collectively, these directions span a harmfulness subspace that we show is strikingly low-rank. We then test ablation of the entire subspace from model internals, as well as steering and"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.21141","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/2507.21141/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":"2507.21141","created_at":"2026-07-05T11:44:32.094515+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.21141v1","created_at":"2026-07-05T11:44:32.094515+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.21141","created_at":"2026-07-05T11:44:32.094515+00:00"},{"alias_kind":"pith_short_12","alias_value":"TAWQTPNYV3AM","created_at":"2026-07-05T11:44:32.094515+00:00"},{"alias_kind":"pith_short_16","alias_value":"TAWQTPNYV3AMHTCH","created_at":"2026-07-05T11:44:32.094515+00:00"},{"alias_kind":"pith_short_8","alias_value":"TAWQTPNY","created_at":"2026-07-05T11:44:32.094515+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.12726","citing_title":"Before the Last Token: Diagnosing Final-Token Safety Probe Failures","ref_index":11,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/TAWQTPNYV3AMHTCHDSVN6WNT37","json":"https://pith.science/pith/TAWQTPNYV3AMHTCHDSVN6WNT37.json","graph_json":"https://pith.science/api/pith-number/TAWQTPNYV3AMHTCHDSVN6WNT37/graph.json","events_json":"https://pith.science/api/pith-number/TAWQTPNYV3AMHTCHDSVN6WNT37/events.json","paper":"https://pith.science/paper/TAWQTPNY"},"agent_actions":{"view_html":"https://pith.science/pith/TAWQTPNYV3AMHTCHDSVN6WNT37","download_json":"https://pith.science/pith/TAWQTPNYV3AMHTCHDSVN6WNT37.json","view_paper":"https://pith.science/paper/TAWQTPNY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.21141&json=true","fetch_graph":"https://pith.science/api/pith-number/TAWQTPNYV3AMHTCHDSVN6WNT37/graph.json","fetch_events":"https://pith.science/api/pith-number/TAWQTPNYV3AMHTCHDSVN6WNT37/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/TAWQTPNYV3AMHTCHDSVN6WNT37/action/timestamp_anchor","attest_storage":"https://pith.science/pith/TAWQTPNYV3AMHTCHDSVN6WNT37/action/storage_attestation","attest_author":"https://pith.science/pith/TAWQTPNYV3AMHTCHDSVN6WNT37/action/author_attestation","sign_citation":"https://pith.science/pith/TAWQTPNYV3AMHTCHDSVN6WNT37/action/citation_signature","submit_replication":"https://pith.science/pith/TAWQTPNYV3AMHTCHDSVN6WNT37/action/replication_record"}},"created_at":"2026-07-05T11:44:32.094515+00:00","updated_at":"2026-07-05T11:44:32.094515+00:00"}