{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "turner-et-al-activation-addition-2023",
    "title": "Turner and colleagues: prompt-pair activation addition",
    "summary": "Used activation differences elicited by contrasting prompts as inference-time steering vectors, showing selected changes to generated text without gradient-based fitting of the steering direction.",
    "boundary": "No optimization of the vector does not mean no parameter selection or computational cost: layer and strength choices still matter and internal activation access is required. The demonstrated controls are not universal concept disentanglement.",
    "area": "Language-model representation control",
    "topic": "Artificial intelligence",
    "year": 2023,
    "lane": "HUMAN",
    "result": "ALGORITHM",
    "package_id": "turner-et-al-activation-addition-2023",
    "target": {
      "record_id": "urn:uuid:898c8062-a12e-41eb-a0da-65e3d577b272",
      "version": 1,
      "content_hash": "sha256:e095fb8bbe2be20d4e3e8def7b842c519e50e630dcfc6929118a1217e47df10d"
    },
    "humans": [
      "Alex Turner",
      "Lisa Thiergart",
      "David Udell",
      "Gavin Leech",
      "Ulisse Mini",
      "Monte MacDiarmid"
    ],
    "human_identities": [
      {
        "author_position": 0,
        "source_name": "Alex Turner",
        "person_id": "alexander-matt-turner-actadd-caa",
        "display_name": "Alexander Matt Turner",
        "status": "SOURCE_LINKED_IDENTITY",
        "sources": [
          {
            "title": "Author explicit name identification",
            "url": "https://turntrout.com/about"
          },
          {
            "title": "Author research and publication list",
            "url": "https://turntrout.com/research"
          }
        ]
      }
    ],
    "authors_complete": true,
    "human_role_note": "Source-listed authors for the exact contribution; experimental subjects are not autonomous scientific authors. Source support and predecessor work retain separate roles. Individual effort shares and unresolved identities are not inferred.",
    "agents": [],
    "related": [],
    "subresults": [],
    "scientific_assessment": {
      "source_review": "SOURCE_CURATED",
      "scientific_review": "NOT_REVIEWED_BY_WRS",
      "formal_replay": "NOT_RUN",
      "provider_attestation": "NONE"
    },
    "sources": [
      {
        "id": "cur20261017-arxiv-2308-10248v1",
        "title": "Activation Addition: Steering Language Models Without Optimization",
        "url": "https://arxiv.org/abs/2308.10248v1",
        "date": "2023-08-20"
      }
    ],
    "points": "50000",
    "published_grade": {
      "id": "turner-et-al-activation-addition-2023",
      "entry_id": "turner-et-al-activation-addition-2023",
      "band": "B",
      "modifier": "+",
      "target": {
        "record_id": "urn:uuid:898c8062-a12e-41eb-a0da-65e3d577b272",
        "version": 1,
        "content_hash": "sha256:e095fb8bbe2be20d4e3e8def7b842c519e50e630dcfc6929118a1217e47df10d"
      },
      "what_changed": "Used activation differences elicited by contrasting prompts as inference-time steering vectors, showing selected changes to generated text without gradient-based fitting of the steering direction.",
      "scope": "Used activation differences elicited by contrasting prompts as inference-time steering vectors, showing selected changes to generated text without gradient-based fitting of the steering direction. No optimization of the vector does not mean no parameter selection or computational cost: layer and strength choices still matter and internal activation access is required. The demonstrated controls are not universal concept disentanglement.",
      "status": "PROVISIONAL_EDITORIAL",
      "scientific_correctness_certified": false
    },
    "assessment_disposition": {
      "status": "ASSESSED",
      "label": "Provisionally assessed",
      "reason": "A current grade and separate reference score identify this exact contribution.",
      "target": {
        "record_id": "urn:uuid:898c8062-a12e-41eb-a0da-65e3d577b272",
        "version": 1,
        "content_hash": "sha256:e095fb8bbe2be20d4e3e8def7b842c519e50e630dcfc6929118a1217e47df10d"
      },
      "edition": "ratings-e039106f30dec177"
    }
  },
  "version": {
    "kind": "PUBLIC_RECORD_VIEW",
    "previous": null,
    "published_at": "2026-10-23T00:00:00Z",
    "publisher": {
      "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33",
      "role": "WRS_EDITORIAL_PUBLICATION"
    },
    "record": {
      "contributors": [
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Alex Turner",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Activation Addition: Steering Language Models Without Optimization",
            "url": "https://arxiv.org/abs/2308.10248v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Lisa Thiergart",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Activation Addition: Steering Language Models Without Optimization",
            "url": "https://arxiv.org/abs/2308.10248v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "David Udell",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Activation Addition: Steering Language Models Without Optimization",
            "url": "https://arxiv.org/abs/2308.10248v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Gavin Leech",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Activation Addition: Steering Language Models Without Optimization",
            "url": "https://arxiv.org/abs/2308.10248v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Ulisse Mini",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Activation Addition: Steering Language Models Without Optimization",
            "url": "https://arxiv.org/abs/2308.10248v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Monte MacDiarmid",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Activation Addition: Steering Language Models Without Optimization",
            "url": "https://arxiv.org/abs/2308.10248v1"
          }
        }
      ],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "Activation Addition: Steering Language Models Without Optimization",
            "url": "https://arxiv.org/abs/2308.10248v1"
          }
        }
      ],
      "origin": "HUMAN_LEGACY",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "No optimization of the vector does not mean no parameter selection or computational cost: layer and strength choices still matter and internal activation access is required. The demonstrated controls are not universal concept disentanglement.",
        "domain": "Language-model representation control"
      },
      "sources": [
        {
          "citation": "Activation Addition: Steering Language Models Without Optimization",
          "url": "https://arxiv.org/abs/2308.10248v1"
        }
      ],
      "statement": "# Turner and colleagues: prompt-pair activation addition\n\n## Change an internal direction without fitting new weights\n\nActivation Addition constructs an internal steering direction from the difference between two prompts' activations. It then uses that direction during another generation. The pinned 2023 version studies GPT-2. Later versions expanded the work, so later models, experiments and additional authors should not silently be inserted into this original contribution.\n\n## Try the arithmetic on paper\n\nIn this invented two-coordinate example, one prompt produces (4, 3) and another produces (1, 2). Their difference is (3, 1). Adding half of that difference to a third point (2, 5) gives (3.5, 5.5). The calculation illustrates a vector operation only. It does not show that either coordinate represents a real concept or that a language model will behave as intended.\n\n## No vector training does not mean no choices\n\nThe method avoids gradient-based fitting of the steering vector, but researchers still choose where and how strongly to intervene. The original experiments select layer and strength settings. Access to internal activations is also required. This differs from simply pasting a new instruction into a closed chat interface. In an invented two-coordinate example, start at (2, 5) and add half of the direction (3, 1). The result is (3.5, 5.5): each coordinate receives its own scaled addition. This arithmetic explains what intervention strength means. Actual hidden states have many more coordinates, and a useful numerical direction at one layer need not behave the same way at another.\n\n## Measure the intended and unintended effects\n\nThe original evaluates selected topic and sentiment changes, with tests of performance away from the target behavior. Those tests support bounded claims about the tested settings. They do not establish universal disentangled concepts or zero side effects. Earlier learned steering methods remain predecessor contributions; this record concerns the prompt-pair construction and its particular demonstrations.\n\n## Make an honest comparison\n\nCompare the original method with dataset-averaged steering and supervised attention-head intervention. Record how each obtains its direction and which parameters still require selection. A clear source table can be useful even if you cannot run the models. Preserve the original version and report an untested configuration as untested.\n\n## Scientific scope\n\nNo optimization of the vector does not mean no parameter selection or computational cost: layer and strength choices still matter and internal activation access is required. The demonstrated controls are not universal concept disentanglement.\n\n## Sources for this explanation\n\n- [Activation Addition: Steering Language Models Without Optimization](https://arxiv.org/abs/2308.10248v1): Pinned original contribution and stated hypotheses or experimental limits.\n",
      "summary": "Used activation differences elicited by contrasting prompts as inference-time steering vectors, showing selected changes to generated text without gradient-based fitting of the steering direction.",
      "title": "Turner and colleagues: prompt-pair activation addition"
    },
    "record_id": "urn:uuid:898c8062-a12e-41eb-a0da-65e3d577b272",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:898c8062-a12e-41eb-a0da-65e3d577b272",
      "version": 1,
      "content_hash": "sha256:e095fb8bbe2be20d4e3e8def7b842c519e50e630dcfc6929118a1217e47df10d"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:1b1a4959a973221d8e5a6d2f8b03912f40d9deaa2270b18d2d5059d6a6398d34"
  },
  "target": {
    "record_id": "urn:uuid:898c8062-a12e-41eb-a0da-65e3d577b272",
    "version": 1,
    "content_hash": "sha256:e095fb8bbe2be20d4e3e8def7b842c519e50e630dcfc6929118a1217e47df10d"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "turner-et-al-activation-addition-2023",
    "entry_id": "turner-et-al-activation-addition-2023",
    "revision": 1,
    "title": "Turner and colleagues: prompt-pair activation addition",
    "context": "Language-model representation control",
    "track": "Human-designed model analysis or safeguard method",
    "band": "B",
    "modifier": "+",
    "status": "PROVISIONAL_EDITORIAL",
    "assessed_on": "2026-10-23",
    "evidence_cutoff": "2026-10-23",
    "achievement_state": "SUPPORTED_SCOPED_RESULT",
    "what_changed": "Used activation differences elicited by contrasting prompts as inference-time steering vectors, showing selected changes to generated text without gradient-based fitting of the steering direction.",
    "scope": "Used activation differences elicited by contrasting prompts as inference-time steering vectors, showing selected changes to generated text without gradient-based fitting of the steering direction. No optimization of the vector does not mean no parameter selection or computational cost: layer and strength choices still matter and internal activation access is required. The demonstrated controls are not universal concept disentanglement.",
    "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. No optimization of the vector does not mean no parameter selection or computational cost: layer and strength choices still matter and internal activation access is required. The demonstrated controls are not universal concept disentanglement.",
    "references": [
      {
        "title": "Activation Addition: Steering Language Models Without Optimization",
        "url": "https://arxiv.org/abs/2308.10248v1",
        "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained."
      }
    ],
    "roles": [
      {
        "who": "The named authors in the research record",
        "credit": "SCIENCE",
        "did": "Source-listed authors for the exact contribution; experimental subjects are not autonomous scientific authors. Source support and predecessor work retain separate roles. Individual effort shares and unresolved identities are not inferred."
      }
    ],
    "target": {
      "record_id": "urn:uuid:898c8062-a12e-41eb-a0da-65e3d577b272",
      "version": 1,
      "content_hash": "sha256:e095fb8bbe2be20d4e3e8def7b842c519e50e630dcfc6929118a1217e47df10d"
    },
    "scientific_correctness_certified": false,
    "independent_expert_calibration": false,
    "method": "PROPRIETARY",
    "points": "50000"
  },
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": {
    "id": "turner-et-al-activation-addition-2023",
    "target": {
      "record_id": "urn:uuid:898c8062-a12e-41eb-a0da-65e3d577b272",
      "version": 1,
      "content_hash": "sha256:e095fb8bbe2be20d4e3e8def7b842c519e50e630dcfc6929118a1217e47df10d"
    },
    "sections": [
      {
        "title": "Change an internal direction without fitting new weights",
        "text": "Activation Addition constructs an internal steering direction from the difference between two prompts' activations. It then uses that direction during another generation. The pinned 2023 version studies GPT-2. Later versions expanded the work, so later models, experiments and additional authors should not silently be inserted into this original contribution."
      },
      {
        "title": "Try the arithmetic on paper",
        "text": "In this invented two-coordinate example, one prompt produces (4, 3) and another produces (1, 2). Their difference is (3, 1). Adding half of that difference to a third point (2, 5) gives (3.5, 5.5). The calculation illustrates a vector operation only. It does not show that either coordinate represents a real concept or that a language model will behave as intended."
      },
      {
        "title": "No vector training does not mean no choices",
        "text": "The method avoids gradient-based fitting of the steering vector, but researchers still choose where and how strongly to intervene. The original experiments select layer and strength settings. Access to internal activations is also required. This differs from simply pasting a new instruction into a closed chat interface. In an invented two-coordinate example, start at (2, 5) and add half of the direction (3, 1). The result is (3.5, 5.5): each coordinate receives its own scaled addition. This arithmetic explains what intervention strength means. Actual hidden states have many more coordinates, and a useful numerical direction at one layer need not behave the same way at another."
      },
      {
        "title": "Measure the intended and unintended effects",
        "text": "The original evaluates selected topic and sentiment changes, with tests of performance away from the target behavior. Those tests support bounded claims about the tested settings. They do not establish universal disentangled concepts or zero side effects. Earlier learned steering methods remain predecessor contributions; this record concerns the prompt-pair construction and its particular demonstrations."
      },
      {
        "title": "Make an honest comparison",
        "text": "Compare the original method with dataset-averaged steering and supervised attention-head intervention. Record how each obtains its direction and which parameters still require selection. A clear source table can be useful even if you cannot run the models. Preserve the original version and report an untested configuration as untested."
      }
    ],
    "references": [
      {
        "title": "Activation Addition: Steering Language Models Without Optimization",
        "url": "https://arxiv.org/abs/2308.10248v1",
        "locator": "Pinned original contribution and stated hypotheses or experimental limits."
      }
    ]
  },
  "reading_status": "EXPLANATORY_GUIDE"
}