{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:3GPC2WLSMZ65AUDBUSS2MK2ZZ6","short_pith_number":"pith:3GPC2WLS","schema_version":"1.0","canonical_sha256":"d99e2d5972667dd05061a4a5a62b59cfafe3b0a1e3e01f9af2a916098eea6e1b","source":{"kind":"arxiv","id":"2505.02273","version":2},"attestation_state":"computed","paper":{"title":"Demystifying optimized prompts in language models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"H. Howie Huang, Lucas H. McCabe, Rimon Melamed","submitted_at":"2025-05-04T22:04:14Z","abstract_excerpt":"Modern language models (LMs) are not robust to out-of-distribution inputs. Machine generated (``optimized'') prompts can be used to modulate LM outputs and induce specific behaviors while appearing completely uninterpretable. In this work, we investigate the composition of optimized prompts, as well as the mechanisms by which LMs parse and build predictions from optimized prompts. We find that optimized prompts primarily consist of punctuation and noun tokens which are more rare in the training data. Internally, optimized prompts are clearly distinguishable from natural language counterparts b"},"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":"2505.02273","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-05-04T22:04:14Z","cross_cats_sorted":[],"title_canon_sha256":"29f1dc733836a89374dee17377714927c9411341e35d8917de5d429de2291ac2","abstract_canon_sha256":"e4fdaa648caa3ce9d13c0b6b6e6daafe07d5f5f3d0951ed05a93381b1125a411"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T12:04:14.946072Z","signature_b64":"PD/UGDRnTRHbTyJS6zcqRTkI0goFzY38E1/Y/oIhHnVtXOy/unmPN256dIEaHNVbafyc1K5HvFTCyJkOiM/2AQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d99e2d5972667dd05061a4a5a62b59cfafe3b0a1e3e01f9af2a916098eea6e1b","last_reissued_at":"2026-07-05T12:04:14.945506Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T12:04:14.945506Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Demystifying optimized prompts in language models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"H. Howie Huang, Lucas H. McCabe, Rimon Melamed","submitted_at":"2025-05-04T22:04:14Z","abstract_excerpt":"Modern language models (LMs) are not robust to out-of-distribution inputs. Machine generated (``optimized'') prompts can be used to modulate LM outputs and induce specific behaviors while appearing completely uninterpretable. In this work, we investigate the composition of optimized prompts, as well as the mechanisms by which LMs parse and build predictions from optimized prompts. We find that optimized prompts primarily consist of punctuation and noun tokens which are more rare in the training data. Internally, optimized prompts are clearly distinguishable from natural language counterparts b"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.02273","kind":"arxiv","version":2},"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/2505.02273/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":"2505.02273","created_at":"2026-07-05T12:04:14.945566+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.02273v2","created_at":"2026-07-05T12:04:14.945566+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.02273","created_at":"2026-07-05T12:04:14.945566+00:00"},{"alias_kind":"pith_short_12","alias_value":"3GPC2WLSMZ65","created_at":"2026-07-05T12:04:14.945566+00:00"},{"alias_kind":"pith_short_16","alias_value":"3GPC2WLSMZ65AUDB","created_at":"2026-07-05T12:04:14.945566+00:00"},{"alias_kind":"pith_short_8","alias_value":"3GPC2WLS","created_at":"2026-07-05T12:04:14.945566+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/3GPC2WLSMZ65AUDBUSS2MK2ZZ6","json":"https://pith.science/pith/3GPC2WLSMZ65AUDBUSS2MK2ZZ6.json","graph_json":"https://pith.science/api/pith-number/3GPC2WLSMZ65AUDBUSS2MK2ZZ6/graph.json","events_json":"https://pith.science/api/pith-number/3GPC2WLSMZ65AUDBUSS2MK2ZZ6/events.json","paper":"https://pith.science/paper/3GPC2WLS"},"agent_actions":{"view_html":"https://pith.science/pith/3GPC2WLSMZ65AUDBUSS2MK2ZZ6","download_json":"https://pith.science/pith/3GPC2WLSMZ65AUDBUSS2MK2ZZ6.json","view_paper":"https://pith.science/paper/3GPC2WLS","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.02273&json=true","fetch_graph":"https://pith.science/api/pith-number/3GPC2WLSMZ65AUDBUSS2MK2ZZ6/graph.json","fetch_events":"https://pith.science/api/pith-number/3GPC2WLSMZ65AUDBUSS2MK2ZZ6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3GPC2WLSMZ65AUDBUSS2MK2ZZ6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3GPC2WLSMZ65AUDBUSS2MK2ZZ6/action/storage_attestation","attest_author":"https://pith.science/pith/3GPC2WLSMZ65AUDBUSS2MK2ZZ6/action/author_attestation","sign_citation":"https://pith.science/pith/3GPC2WLSMZ65AUDBUSS2MK2ZZ6/action/citation_signature","submit_replication":"https://pith.science/pith/3GPC2WLSMZ65AUDBUSS2MK2ZZ6/action/replication_record"}},"created_at":"2026-07-05T12:04:14.945566+00:00","updated_at":"2026-07-05T12:04:14.945566+00:00"}