{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:E6K2VRBDGJWCZ7OOJ4YZJCFY3B","short_pith_number":"pith:E6K2VRBD","schema_version":"1.0","canonical_sha256":"2795aac423326c2cfdce4f319488b8d87c02b5cf8bf7f7a0ca55d88bb1ff027b","source":{"kind":"arxiv","id":"2503.03730","version":2},"attestation_state":"computed","paper":{"title":"Towards Understanding Distilled Reasoning Models: A Representational Approach","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"David D. Baek, Max Tegmark","submitted_at":"2025-03-05T18:40:19Z","abstract_excerpt":"In this paper, we investigate how model distillation impacts the development of reasoning features in large language models (LLMs). To explore this, we train a crosscoder on Qwen-series models and their fine-tuned variants. Our results suggest that the crosscoder learns features corresponding to various types of reasoning, including self-reflection and computation verification. Moreover, we observe that distilled models contain unique reasoning feature directions, which could be used to steer the model into over-thinking or incisive-thinking mode. In particular, we perform analysis on four spe"},"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":"2503.03730","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-03-05T18:40:19Z","cross_cats_sorted":[],"title_canon_sha256":"0464a517e92795f86b74fa8bb3e59e7318280f8def712ae9eb998b7e2f3491e6","abstract_canon_sha256":"f6d4c106763ec3f15e0f3fe96167479b0b159143dae3fd52cbf5af6098be0125"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:38:43.640762Z","signature_b64":"SBg2WodDkBdBtNbwskwU3pKeMbPEJtMykjJBBV5aFowr6ilZBa4o/pbZfUYdEoh7ET2Ptn1c/TG7l1G1LNmzAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2795aac423326c2cfdce4f319488b8d87c02b5cf8bf7f7a0ca55d88bb1ff027b","last_reissued_at":"2026-07-05T10:38:43.640324Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:38:43.640324Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Towards Understanding Distilled Reasoning Models: A Representational Approach","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"David D. Baek, Max Tegmark","submitted_at":"2025-03-05T18:40:19Z","abstract_excerpt":"In this paper, we investigate how model distillation impacts the development of reasoning features in large language models (LLMs). To explore this, we train a crosscoder on Qwen-series models and their fine-tuned variants. Our results suggest that the crosscoder learns features corresponding to various types of reasoning, including self-reflection and computation verification. Moreover, we observe that distilled models contain unique reasoning feature directions, which could be used to steer the model into over-thinking or incisive-thinking mode. In particular, we perform analysis on four spe"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.03730","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/2503.03730/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":"2503.03730","created_at":"2026-07-05T10:38:43.640389+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.03730v2","created_at":"2026-07-05T10:38:43.640389+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.03730","created_at":"2026-07-05T10:38:43.640389+00:00"},{"alias_kind":"pith_short_12","alias_value":"E6K2VRBDGJWC","created_at":"2026-07-05T10:38:43.640389+00:00"},{"alias_kind":"pith_short_16","alias_value":"E6K2VRBDGJWCZ7OO","created_at":"2026-07-05T10:38:43.640389+00:00"},{"alias_kind":"pith_short_8","alias_value":"E6K2VRBD","created_at":"2026-07-05T10:38:43.640389+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.11922","citing_title":"StepCodeReasoner: Aligning Code Reasoning with Stepwise Execution Traces via Reinforcement Learning","ref_index":68,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09438","citing_title":"fmxcoders: Factorized Masked Crosscoders for Cross-Layer Feature Discovery","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2604.18936","citing_title":"Fine-Tuning Small Reasoning Models for Quantum Field Theory","ref_index":45,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/E6K2VRBDGJWCZ7OOJ4YZJCFY3B","json":"https://pith.science/pith/E6K2VRBDGJWCZ7OOJ4YZJCFY3B.json","graph_json":"https://pith.science/api/pith-number/E6K2VRBDGJWCZ7OOJ4YZJCFY3B/graph.json","events_json":"https://pith.science/api/pith-number/E6K2VRBDGJWCZ7OOJ4YZJCFY3B/events.json","paper":"https://pith.science/paper/E6K2VRBD"},"agent_actions":{"view_html":"https://pith.science/pith/E6K2VRBDGJWCZ7OOJ4YZJCFY3B","download_json":"https://pith.science/pith/E6K2VRBDGJWCZ7OOJ4YZJCFY3B.json","view_paper":"https://pith.science/paper/E6K2VRBD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.03730&json=true","fetch_graph":"https://pith.science/api/pith-number/E6K2VRBDGJWCZ7OOJ4YZJCFY3B/graph.json","fetch_events":"https://pith.science/api/pith-number/E6K2VRBDGJWCZ7OOJ4YZJCFY3B/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/E6K2VRBDGJWCZ7OOJ4YZJCFY3B/action/timestamp_anchor","attest_storage":"https://pith.science/pith/E6K2VRBDGJWCZ7OOJ4YZJCFY3B/action/storage_attestation","attest_author":"https://pith.science/pith/E6K2VRBDGJWCZ7OOJ4YZJCFY3B/action/author_attestation","sign_citation":"https://pith.science/pith/E6K2VRBDGJWCZ7OOJ4YZJCFY3B/action/citation_signature","submit_replication":"https://pith.science/pith/E6K2VRBDGJWCZ7OOJ4YZJCFY3B/action/replication_record"}},"created_at":"2026-07-05T10:38:43.640389+00:00","updated_at":"2026-07-05T10:38:43.640389+00:00"}