{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:36KIXTG34G7FL3KYVYQOIGXCZQ","short_pith_number":"pith:36KIXTG3","schema_version":"1.0","canonical_sha256":"df948bccdbe1be55ed58ae20e41ae2cc27570addc024ea02f64b11d7b227f66f","source":{"kind":"arxiv","id":"2507.04742","version":2},"attestation_state":"computed","paper":{"title":"Activation Steering for Chain-of-Thought Compression","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Erfan Baghaei Potraghloo, Massoud Pedram, Seyedarmin Azizi","submitted_at":"2025-07-07T08:16:54Z","abstract_excerpt":"Large language models (LLMs) excel at complex reasoning when they include intermediate steps, known as \"chains of thought\" (CoTs). However, these rationales are often overly verbose, even for simple problems, leading to wasted context, increased latency, and higher energy consumption. We observe that verbose, English-heavy CoTs and concise, math-centric CoTs occupy distinct regions in the model's residual-stream activation space. By extracting and injecting a \"steering vector\" to transition between these modes, we can reliably shift generation toward more concise reasoning, effectively compres"},"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":"2507.04742","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-07-07T08:16:54Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"cfd62a083c71e48db77d07f79aeca3cd6dc929e6c7706aeeabf5d380637c7f59","abstract_canon_sha256":"3e4625ac791b02e05455714a701f5cd50f80c39c29c7d8b31a1287d5ec437904"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:33:26.735549Z","signature_b64":"Tgr1XMc984q/Z3aESYkKfHDR38T9jqvPsFQ5YZK/3aDRJNZFjLOOOxjYpi/COi0Jq1DcLFToMQEeRdM9+bkSBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"df948bccdbe1be55ed58ae20e41ae2cc27570addc024ea02f64b11d7b227f66f","last_reissued_at":"2026-07-05T11:33:26.735067Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:33:26.735067Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Activation Steering for Chain-of-Thought Compression","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Erfan Baghaei Potraghloo, Massoud Pedram, Seyedarmin Azizi","submitted_at":"2025-07-07T08:16:54Z","abstract_excerpt":"Large language models (LLMs) excel at complex reasoning when they include intermediate steps, known as \"chains of thought\" (CoTs). However, these rationales are often overly verbose, even for simple problems, leading to wasted context, increased latency, and higher energy consumption. We observe that verbose, English-heavy CoTs and concise, math-centric CoTs occupy distinct regions in the model's residual-stream activation space. By extracting and injecting a \"steering vector\" to transition between these modes, we can reliably shift generation toward more concise reasoning, effectively compres"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.04742","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/2507.04742/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":"2507.04742","created_at":"2026-07-05T11:33:26.735127+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.04742v2","created_at":"2026-07-05T11:33:26.735127+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.04742","created_at":"2026-07-05T11:33:26.735127+00:00"},{"alias_kind":"pith_short_12","alias_value":"36KIXTG34G7F","created_at":"2026-07-05T11:33:26.735127+00:00"},{"alias_kind":"pith_short_16","alias_value":"36KIXTG34G7FL3KY","created_at":"2026-07-05T11:33:26.735127+00:00"},{"alias_kind":"pith_short_8","alias_value":"36KIXTG3","created_at":"2026-07-05T11:33:26.735127+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.11599","citing_title":"When is Your LLM Steerable?","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2509.22739","citing_title":"Painless Activation Steering: An Automated, Lightweight Approach for Post-Training Large Language Models","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20745","citing_title":"The Hidden Signal of Verifier Strictness: Controlling and Improving Step-Wise Verification via Selective Latent Steering","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24693","citing_title":"Contextual Linear Activation Steering of Language Models","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06377","citing_title":"The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment","ref_index":7,"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":7,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/36KIXTG34G7FL3KYVYQOIGXCZQ","json":"https://pith.science/pith/36KIXTG34G7FL3KYVYQOIGXCZQ.json","graph_json":"https://pith.science/api/pith-number/36KIXTG34G7FL3KYVYQOIGXCZQ/graph.json","events_json":"https://pith.science/api/pith-number/36KIXTG34G7FL3KYVYQOIGXCZQ/events.json","paper":"https://pith.science/paper/36KIXTG3"},"agent_actions":{"view_html":"https://pith.science/pith/36KIXTG34G7FL3KYVYQOIGXCZQ","download_json":"https://pith.science/pith/36KIXTG34G7FL3KYVYQOIGXCZQ.json","view_paper":"https://pith.science/paper/36KIXTG3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.04742&json=true","fetch_graph":"https://pith.science/api/pith-number/36KIXTG34G7FL3KYVYQOIGXCZQ/graph.json","fetch_events":"https://pith.science/api/pith-number/36KIXTG34G7FL3KYVYQOIGXCZQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/36KIXTG34G7FL3KYVYQOIGXCZQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/36KIXTG34G7FL3KYVYQOIGXCZQ/action/storage_attestation","attest_author":"https://pith.science/pith/36KIXTG34G7FL3KYVYQOIGXCZQ/action/author_attestation","sign_citation":"https://pith.science/pith/36KIXTG34G7FL3KYVYQOIGXCZQ/action/citation_signature","submit_replication":"https://pith.science/pith/36KIXTG34G7FL3KYVYQOIGXCZQ/action/replication_record"}},"created_at":"2026-07-05T11:33:26.735127+00:00","updated_at":"2026-07-05T11:33:26.735127+00:00"}