{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:OYG7BYHP6SQ2WEQXS6V3KZLGQ7","short_pith_number":"pith:OYG7BYHP","schema_version":"1.0","canonical_sha256":"760df0e0eff4a1ab121797abb5656687e095b997fdf574cfdf61ddbfb796d9d7","source":{"kind":"arxiv","id":"2505.17697","version":1},"attestation_state":"computed","paper":{"title":"Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Biwei Huang, Kun Zhou, Qi Liu, Yifei Shao, Zekai Zhao, Zhiting Hu, Zihan Liu","submitted_at":"2025-05-23T10:07:18Z","abstract_excerpt":"Despite the remarkable reasoning performance, eliciting the long chain-of-thought (CoT) ability in large language models (LLMs) typically requires costly reinforcement learning or supervised fine-tuning on high-quality distilled data. We investigate the internal mechanisms behind this capability and show that a small set of high-impact activations in the last few layers largely governs long-form reasoning attributes, such as output length and self-reflection. By simply amplifying these activations and inserting \"wait\" tokens, we can invoke the long CoT ability without any training, resulting i"},"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":"2505.17697","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-05-23T10:07:18Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"e9a8a910ea266fdc32ed6e5ff72a90b59211f1a7c107c666e744fd9b77773374","abstract_canon_sha256":"21e021d22d532ed42ebc49ff3b7647ab41f6280d9e7c9f8532e07ab735d9a94e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:08:30.362401Z","signature_b64":"D0Yv6v+hRwtxKKob5s0F6CLhmdyoz50YNlqmFmwY6OVd+IHwcGnsdQJ1dWrpt6vuqDwXciCZbYmRveTIBR2RBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"760df0e0eff4a1ab121797abb5656687e095b997fdf574cfdf61ddbfb796d9d7","last_reissued_at":"2026-07-05T11:08:30.361938Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:08:30.361938Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Biwei Huang, Kun Zhou, Qi Liu, Yifei Shao, Zekai Zhao, Zhiting Hu, Zihan Liu","submitted_at":"2025-05-23T10:07:18Z","abstract_excerpt":"Despite the remarkable reasoning performance, eliciting the long chain-of-thought (CoT) ability in large language models (LLMs) typically requires costly reinforcement learning or supervised fine-tuning on high-quality distilled data. We investigate the internal mechanisms behind this capability and show that a small set of high-impact activations in the last few layers largely governs long-form reasoning attributes, such as output length and self-reflection. By simply amplifying these activations and inserting \"wait\" tokens, we can invoke the long CoT ability without any training, resulting i"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.17697","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/2505.17697/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":"2505.17697","created_at":"2026-07-05T11:08:30.362000+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.17697v1","created_at":"2026-07-05T11:08:30.362000+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.17697","created_at":"2026-07-05T11:08:30.362000+00:00"},{"alias_kind":"pith_short_12","alias_value":"OYG7BYHP6SQ2","created_at":"2026-07-05T11:08:30.362000+00:00"},{"alias_kind":"pith_short_16","alias_value":"OYG7BYHP6SQ2WEQX","created_at":"2026-07-05T11:08:30.362000+00:00"},{"alias_kind":"pith_short_8","alias_value":"OYG7BYHP","created_at":"2026-07-05T11:08:30.362000+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.18089","citing_title":"From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning","ref_index":73,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06188","citing_title":"The Tell-Tale Norm: $\\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07021","citing_title":"Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through Oversight","ref_index":51,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07021","citing_title":"Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through Oversight","ref_index":51,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19149","citing_title":"How Do Answer Tokens Read Reasoning Traces? Self-Reading Patterns in Thinking LLMs for Quantitative Reasoning","ref_index":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OYG7BYHP6SQ2WEQXS6V3KZLGQ7","json":"https://pith.science/pith/OYG7BYHP6SQ2WEQXS6V3KZLGQ7.json","graph_json":"https://pith.science/api/pith-number/OYG7BYHP6SQ2WEQXS6V3KZLGQ7/graph.json","events_json":"https://pith.science/api/pith-number/OYG7BYHP6SQ2WEQXS6V3KZLGQ7/events.json","paper":"https://pith.science/paper/OYG7BYHP"},"agent_actions":{"view_html":"https://pith.science/pith/OYG7BYHP6SQ2WEQXS6V3KZLGQ7","download_json":"https://pith.science/pith/OYG7BYHP6SQ2WEQXS6V3KZLGQ7.json","view_paper":"https://pith.science/paper/OYG7BYHP","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.17697&json=true","fetch_graph":"https://pith.science/api/pith-number/OYG7BYHP6SQ2WEQXS6V3KZLGQ7/graph.json","fetch_events":"https://pith.science/api/pith-number/OYG7BYHP6SQ2WEQXS6V3KZLGQ7/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OYG7BYHP6SQ2WEQXS6V3KZLGQ7/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OYG7BYHP6SQ2WEQXS6V3KZLGQ7/action/storage_attestation","attest_author":"https://pith.science/pith/OYG7BYHP6SQ2WEQXS6V3KZLGQ7/action/author_attestation","sign_citation":"https://pith.science/pith/OYG7BYHP6SQ2WEQXS6V3KZLGQ7/action/citation_signature","submit_replication":"https://pith.science/pith/OYG7BYHP6SQ2WEQXS6V3KZLGQ7/action/replication_record"}},"created_at":"2026-07-05T11:08:30.362000+00:00","updated_at":"2026-07-05T11:08:30.362000+00:00"}