{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:2ZTXNN37CZTH3F3SJPXCVLZYVX","short_pith_number":"pith:2ZTXNN37","schema_version":"1.0","canonical_sha256":"d66776b77f16667d97724bee2aaf38adf49c7891030c7e816588ec5b41311245","source":{"kind":"arxiv","id":"2606.23443","version":1},"attestation_state":"computed","paper":{"title":"What Does a Chemical Language Model Know About Molecules?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","physics.chem-ph"],"primary_cat":"cs.LG","authors_text":"Christian Kenneth, Etowah Adams, Gerard JP van Westen, Liam Bai","submitted_at":"2026-06-22T14:59:57Z","abstract_excerpt":"Chemical language models (cLMs) are widely assumed to learn surface-level syntactic patterns rather than learning meaningful molecular semantics. Here, we apply sparse autoencoders (SAEs) to MolFormer, an encoder-only cLM, to mechanistically examine how molecular representations are built across layers. We discover that early layers rely on position-tracking latents to parse molecular grammar, while later layers encode atom-in-substructure and pharmacologically relevant features. Additionally, we show that non-canonical SMILES produce more disruptive representation shifts than invalid SMILES, "},"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":"2606.23443","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-06-22T14:59:57Z","cross_cats_sorted":["cs.AI","physics.chem-ph"],"title_canon_sha256":"0d621d68b1f8e5f7e796828fc59e728a9f463f83952dd1adeb535291d3f3060d","abstract_canon_sha256":"2c90c7a6a1ddf27384f9db2466b9fc587dd0b8c7f41ca7180e30dab053a3059f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-23T03:14:20.104153Z","signature_b64":"RJrT9MeMXyLxsJkBqSycD92hBuZHOWwCkfD/advBfJX7pyYe38DXt1xMN4CJDVJrfvd4lt4GU6/yimRI+nqmDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d66776b77f16667d97724bee2aaf38adf49c7891030c7e816588ec5b41311245","last_reissued_at":"2026-06-23T03:14:20.103748Z","signature_status":"signed_v1","first_computed_at":"2026-06-23T03:14:20.103748Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"What Does a Chemical Language Model Know About Molecules?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","physics.chem-ph"],"primary_cat":"cs.LG","authors_text":"Christian Kenneth, Etowah Adams, Gerard JP van Westen, Liam Bai","submitted_at":"2026-06-22T14:59:57Z","abstract_excerpt":"Chemical language models (cLMs) are widely assumed to learn surface-level syntactic patterns rather than learning meaningful molecular semantics. Here, we apply sparse autoencoders (SAEs) to MolFormer, an encoder-only cLM, to mechanistically examine how molecular representations are built across layers. We discover that early layers rely on position-tracking latents to parse molecular grammar, while later layers encode atom-in-substructure and pharmacologically relevant features. Additionally, we show that non-canonical SMILES produce more disruptive representation shifts than invalid SMILES, "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.23443","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/2606.23443/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":"2606.23443","created_at":"2026-06-23T03:14:20.103815+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.23443v1","created_at":"2026-06-23T03:14:20.103815+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.23443","created_at":"2026-06-23T03:14:20.103815+00:00"},{"alias_kind":"pith_short_12","alias_value":"2ZTXNN37CZTH","created_at":"2026-06-23T03:14:20.103815+00:00"},{"alias_kind":"pith_short_16","alias_value":"2ZTXNN37CZTH3F3S","created_at":"2026-06-23T03:14:20.103815+00:00"},{"alias_kind":"pith_short_8","alias_value":"2ZTXNN37","created_at":"2026-06-23T03:14:20.103815+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/2ZTXNN37CZTH3F3SJPXCVLZYVX","json":"https://pith.science/pith/2ZTXNN37CZTH3F3SJPXCVLZYVX.json","graph_json":"https://pith.science/api/pith-number/2ZTXNN37CZTH3F3SJPXCVLZYVX/graph.json","events_json":"https://pith.science/api/pith-number/2ZTXNN37CZTH3F3SJPXCVLZYVX/events.json","paper":"https://pith.science/paper/2ZTXNN37"},"agent_actions":{"view_html":"https://pith.science/pith/2ZTXNN37CZTH3F3SJPXCVLZYVX","download_json":"https://pith.science/pith/2ZTXNN37CZTH3F3SJPXCVLZYVX.json","view_paper":"https://pith.science/paper/2ZTXNN37","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.23443&json=true","fetch_graph":"https://pith.science/api/pith-number/2ZTXNN37CZTH3F3SJPXCVLZYVX/graph.json","fetch_events":"https://pith.science/api/pith-number/2ZTXNN37CZTH3F3SJPXCVLZYVX/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/2ZTXNN37CZTH3F3SJPXCVLZYVX/action/timestamp_anchor","attest_storage":"https://pith.science/pith/2ZTXNN37CZTH3F3SJPXCVLZYVX/action/storage_attestation","attest_author":"https://pith.science/pith/2ZTXNN37CZTH3F3SJPXCVLZYVX/action/author_attestation","sign_citation":"https://pith.science/pith/2ZTXNN37CZTH3F3SJPXCVLZYVX/action/citation_signature","submit_replication":"https://pith.science/pith/2ZTXNN37CZTH3F3SJPXCVLZYVX/action/replication_record"}},"created_at":"2026-06-23T03:14:20.103815+00:00","updated_at":"2026-06-23T03:14:20.103815+00:00"}