{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:3GRGBG2MBRAVMQCPHH57NZI2QA","short_pith_number":"pith:3GRGBG2M","schema_version":"1.0","canonical_sha256":"d9a2609b4c0c4156404f39fbf6e51a800819df044ef4b5d661fa6f17cfc79102","source":{"kind":"arxiv","id":"2310.18512","version":2},"attestation_state":"computed","paper":{"title":"Preventing Language Models From Hiding Their Reasoning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Fabien Roger, Ryan Greenblatt","submitted_at":"2023-10-27T22:02:29Z","abstract_excerpt":"Large language models (LLMs) often benefit from intermediate steps of reasoning to generate answers to complex problems. When these intermediate steps of reasoning are used to monitor the activity of the model, it is essential that this explicit reasoning is faithful, i.e. that it reflects what the model is actually reasoning about. In this work, we focus on one potential way intermediate steps of reasoning could be unfaithful: encoded reasoning, where an LLM could encode intermediate steps of reasoning in the generated text in a way that is not understandable to human readers. We show that la"},"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":"2310.18512","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2023-10-27T22:02:29Z","cross_cats_sorted":[],"title_canon_sha256":"59d0c8f3daf91c66baa3c2201d83d63d2a9f62e52b9eb1d23d07f90ea88f8810","abstract_canon_sha256":"cb6a9831e895e3a02b78daa9ae61d223a1f19c6749bcd5dfd6be37ddc94dcb1f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:07:40.290159Z","signature_b64":"xmEij8+zXAIH51zdRGmaZKwc1yDqGPOyZm8Xum9/x27aAG9YMofpaTUW3H0JaMEh3Ot1OIqYyWzAWxfccbJHCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d9a2609b4c0c4156404f39fbf6e51a800819df044ef4b5d661fa6f17cfc79102","last_reissued_at":"2026-07-05T07:07:40.289697Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:07:40.289697Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Preventing Language Models From Hiding Their Reasoning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Fabien Roger, Ryan Greenblatt","submitted_at":"2023-10-27T22:02:29Z","abstract_excerpt":"Large language models (LLMs) often benefit from intermediate steps of reasoning to generate answers to complex problems. When these intermediate steps of reasoning are used to monitor the activity of the model, it is essential that this explicit reasoning is faithful, i.e. that it reflects what the model is actually reasoning about. In this work, we focus on one potential way intermediate steps of reasoning could be unfaithful: encoded reasoning, where an LLM could encode intermediate steps of reasoning in the generated text in a way that is not understandable to human readers. We show that la"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2310.18512","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/2310.18512/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":"2310.18512","created_at":"2026-07-05T07:07:40.289756+00:00"},{"alias_kind":"arxiv_version","alias_value":"2310.18512v2","created_at":"2026-07-05T07:07:40.289756+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2310.18512","created_at":"2026-07-05T07:07:40.289756+00:00"},{"alias_kind":"pith_short_12","alias_value":"3GRGBG2MBRAV","created_at":"2026-07-05T07:07:40.289756+00:00"},{"alias_kind":"pith_short_16","alias_value":"3GRGBG2MBRAVMQCP","created_at":"2026-07-05T07:07:40.289756+00:00"},{"alias_kind":"pith_short_8","alias_value":"3GRGBG2M","created_at":"2026-07-05T07:07:40.289756+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":10,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.19603","citing_title":"Comparing Linear Probes with Mahalanobis Cosine Similarity","ref_index":65,"is_internal_anchor":false},{"citing_arxiv_id":"2606.11063","citing_title":"CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28615","citing_title":"What LLMs explain is not what they believe: Evaluating explanation sufficiency under models' own input beliefs","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2605.24396","citing_title":"Understanding and Mitigating Premature Confidence for Better LLM Reasoning","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26537","citing_title":"Conceptual Steganography","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09931","citing_title":"A Note on the Strategic Confinement Problem","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2507.11473","citing_title":"Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11746","citing_title":"When Reasoning Traces Become Performative: Step-Level Evidence that Chain-of-Thought Is an Imperfect Oversight Channel","ref_index":44,"is_internal_anchor":false},{"citing_arxiv_id":"2412.14093","citing_title":"Alignment faking in large language models","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2604.16242","citing_title":"Detecting and Suppressing Reward Hacking with Gradient Fingerprints","ref_index":4,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3GRGBG2MBRAVMQCPHH57NZI2QA","json":"https://pith.science/pith/3GRGBG2MBRAVMQCPHH57NZI2QA.json","graph_json":"https://pith.science/api/pith-number/3GRGBG2MBRAVMQCPHH57NZI2QA/graph.json","events_json":"https://pith.science/api/pith-number/3GRGBG2MBRAVMQCPHH57NZI2QA/events.json","paper":"https://pith.science/paper/3GRGBG2M"},"agent_actions":{"view_html":"https://pith.science/pith/3GRGBG2MBRAVMQCPHH57NZI2QA","download_json":"https://pith.science/pith/3GRGBG2MBRAVMQCPHH57NZI2QA.json","view_paper":"https://pith.science/paper/3GRGBG2M","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2310.18512&json=true","fetch_graph":"https://pith.science/api/pith-number/3GRGBG2MBRAVMQCPHH57NZI2QA/graph.json","fetch_events":"https://pith.science/api/pith-number/3GRGBG2MBRAVMQCPHH57NZI2QA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3GRGBG2MBRAVMQCPHH57NZI2QA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3GRGBG2MBRAVMQCPHH57NZI2QA/action/storage_attestation","attest_author":"https://pith.science/pith/3GRGBG2MBRAVMQCPHH57NZI2QA/action/author_attestation","sign_citation":"https://pith.science/pith/3GRGBG2MBRAVMQCPHH57NZI2QA/action/citation_signature","submit_replication":"https://pith.science/pith/3GRGBG2MBRAVMQCPHH57NZI2QA/action/replication_record"}},"created_at":"2026-07-05T07:07:40.289756+00:00","updated_at":"2026-07-05T07:07:40.289756+00:00"}