{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "rimsky-et-al-contrastive-activation-addition-2023",
    "title": "Rimsky and colleagues: contrastive activation addition for model steering",
    "summary": "Constructed steering directions by averaging activation differences from contrasting response pairs and applied them during inference to shift selected behaviors of Llama 2 models without updating the base model's weights.",
    "boundary": "This uses contrastive examples, activation access and selected intervention parameters. Behavioral changes and side effects depend on the model, task and steering strength; pair-averaged construction is distinct from a single-prompt activation difference.",
    "area": "Language-model representation control",
    "topic": "Artificial intelligence",
    "year": 2023,
    "lane": "HUMAN",
    "result": "ALGORITHM",
    "package_id": "rimsky-et-al-contrastive-activation-addition-2023",
    "target": {
      "record_id": "urn:uuid:3f35adca-bdc0-4e5c-a752-be97a4cc2eb6",
      "version": 1,
      "content_hash": "sha256:41490337080b2337c760fdc702b761656e6e25a64debe9a093f0edce0f125c7b"
    },
    "humans": [
      "Nina Rimsky",
      "Nick Gabrieli",
      "Julian Schulz",
      "Meg Tong",
      "Evan Hubinger",
      "Alexander Matt Turner"
    ],
    "human_identities": [
      {
        "author_position": 5,
        "source_name": "Alexander Matt 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-2312-06681v1",
        "title": "Steering Llama 2 via Contrastive Activation Addition",
        "url": "https://arxiv.org/abs/2312.06681v1",
        "date": "2023-12-09"
      }
    ],
    "points": "35000",
    "published_grade": {
      "id": "rimsky-et-al-contrastive-activation-addition-2023",
      "entry_id": "rimsky-et-al-contrastive-activation-addition-2023",
      "band": "B",
      "modifier": "+",
      "target": {
        "record_id": "urn:uuid:3f35adca-bdc0-4e5c-a752-be97a4cc2eb6",
        "version": 1,
        "content_hash": "sha256:41490337080b2337c760fdc702b761656e6e25a64debe9a093f0edce0f125c7b"
      },
      "what_changed": "Constructed steering directions by averaging activation differences from contrasting response pairs and applied them during inference to shift selected behaviors of Llama 2 models without updating the base model's weights.",
      "scope": "Constructed steering directions by averaging activation differences from contrasting response pairs and applied them during inference to shift selected behaviors of Llama 2 models without updating the base model's weights. This uses contrastive examples, activation access and selected intervention parameters. Behavioral changes and side effects depend on the model, task and steering strength; pair-averaged construction is distinct from a single-prompt activation difference.",
      "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:3f35adca-bdc0-4e5c-a752-be97a4cc2eb6",
        "version": 1,
        "content_hash": "sha256:41490337080b2337c760fdc702b761656e6e25a64debe9a093f0edce0f125c7b"
      },
      "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": "Nina Rimsky",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Steering Llama 2 via Contrastive Activation Addition",
            "url": "https://arxiv.org/abs/2312.06681v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Nick Gabrieli",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Steering Llama 2 via Contrastive Activation Addition",
            "url": "https://arxiv.org/abs/2312.06681v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Julian Schulz",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Steering Llama 2 via Contrastive Activation Addition",
            "url": "https://arxiv.org/abs/2312.06681v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Meg Tong",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Steering Llama 2 via Contrastive Activation Addition",
            "url": "https://arxiv.org/abs/2312.06681v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Evan Hubinger",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Steering Llama 2 via Contrastive Activation Addition",
            "url": "https://arxiv.org/abs/2312.06681v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Alexander Matt Turner",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Steering Llama 2 via Contrastive Activation Addition",
            "url": "https://arxiv.org/abs/2312.06681v1"
          }
        }
      ],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "Steering Llama 2 via Contrastive Activation Addition",
            "url": "https://arxiv.org/abs/2312.06681v1"
          }
        }
      ],
      "origin": "HUMAN_LEGACY",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "This uses contrastive examples, activation access and selected intervention parameters. Behavioral changes and side effects depend on the model, task and steering strength; pair-averaged construction is distinct from a single-prompt activation difference.",
        "domain": "Language-model representation control"
      },
      "sources": [
        {
          "citation": "Steering Llama 2 via Contrastive Activation Addition",
          "url": "https://arxiv.org/abs/2312.06681v1"
        }
      ],
      "statement": "# Rimsky and colleagues: contrastive activation addition for model steering\n\n## Average examples instead of trusting one pair\n\nContrastive Activation Addition builds steering directions from many paired examples of a behavior. It averages differences in internal activations and applies the result while a model generates text. The original tests Llama 2 models. It is closely related to earlier prompt-pair Activation Addition, with a distinct dataset-based construction and evaluation.\n\n## A small average shows the idea\n\nFor an invented arithmetic illustration, suppose three example differences are (2, 0), (4, 2) and (3, 1). Their average is (3, 1). Combining examples can reduce some example-specific variation. It cannot eliminate a bias shared by all three examples, and averaging alone does not prove that a direction captures the intended behavior.\n\n## The comparison is about the increment\n\nThe paper contrasts its many-example construction with single-pair steering and other intervention methods. Source-reported benefits concern specified behavioral evaluations. The existing ideas of internal activations, linear directions and inference-time steering are not counted again as entirely new discoveries. The useful question is what the new construction makes more reliable or easier to study. For an invented averaging example, suppose two prompt pairs give differences (2, 0) and (4, 2). Their mean is (3, 1). Averaging can reduce variation specific to one example, but cannot remove a bias shared by both pairs. A separate evaluation set matters because reusing the construction examples would make the intervention look more general than the check establishes.\n\n## An evaluator is part of the experiment\n\nSome behavior measurements use another model as a judge. Its rating is a measurement under a particular rubric, not independent proof of a human psychological trait. Effects can depend on the examples, the model, intervention strength and the task. Side effects and transfer to new prompts therefore need their own evidence.\n\n## What to check next\n\nTrace one reported behavioral comparison to the examples used to form the direction and to the evaluation procedure. Keep construction data and test data separate in your notes. Explain which result was measured and which broader conclusion would need a new experiment. This is a useful source-checking task without claiming that you reproduced the model run.\n\n## Scientific scope\n\nThis uses contrastive examples, activation access and selected intervention parameters. Behavioral changes and side effects depend on the model, task and steering strength; pair-averaged construction is distinct from a single-prompt activation difference.\n\n## Sources for this explanation\n\n- [Steering Llama 2 via Contrastive Activation Addition](https://arxiv.org/abs/2312.06681v1): Pinned original contribution and stated hypotheses or experimental limits.\n",
      "summary": "Constructed steering directions by averaging activation differences from contrasting response pairs and applied them during inference to shift selected behaviors of Llama 2 models without updating the base model's weights.",
      "title": "Rimsky and colleagues: contrastive activation addition for model steering"
    },
    "record_id": "urn:uuid:3f35adca-bdc0-4e5c-a752-be97a4cc2eb6",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:3f35adca-bdc0-4e5c-a752-be97a4cc2eb6",
      "version": 1,
      "content_hash": "sha256:41490337080b2337c760fdc702b761656e6e25a64debe9a093f0edce0f125c7b"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:e20ba0e5b2388823f6d7b1ef3ecbaef7a17ba965f510f79e08da9e31dd6814a3"
  },
  "target": {
    "record_id": "urn:uuid:3f35adca-bdc0-4e5c-a752-be97a4cc2eb6",
    "version": 1,
    "content_hash": "sha256:41490337080b2337c760fdc702b761656e6e25a64debe9a093f0edce0f125c7b"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "rimsky-et-al-contrastive-activation-addition-2023",
    "entry_id": "rimsky-et-al-contrastive-activation-addition-2023",
    "revision": 1,
    "title": "Rimsky and colleagues: contrastive activation addition for model steering",
    "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": "Constructed steering directions by averaging activation differences from contrasting response pairs and applied them during inference to shift selected behaviors of Llama 2 models without updating the base model's weights.",
    "scope": "Constructed steering directions by averaging activation differences from contrasting response pairs and applied them during inference to shift selected behaviors of Llama 2 models without updating the base model's weights. This uses contrastive examples, activation access and selected intervention parameters. Behavioral changes and side effects depend on the model, task and steering strength; pair-averaged construction is distinct from a single-prompt activation difference.",
    "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. This uses contrastive examples, activation access and selected intervention parameters. Behavioral changes and side effects depend on the model, task and steering strength; pair-averaged construction is distinct from a single-prompt activation difference.",
    "references": [
      {
        "title": "Steering Llama 2 via Contrastive Activation Addition",
        "url": "https://arxiv.org/abs/2312.06681v1",
        "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:3f35adca-bdc0-4e5c-a752-be97a4cc2eb6",
      "version": 1,
      "content_hash": "sha256:41490337080b2337c760fdc702b761656e6e25a64debe9a093f0edce0f125c7b"
    },
    "scientific_correctness_certified": false,
    "independent_expert_calibration": false,
    "method": "PROPRIETARY",
    "points": "35000"
  },
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": {
    "id": "rimsky-et-al-contrastive-activation-addition-2023",
    "target": {
      "record_id": "urn:uuid:3f35adca-bdc0-4e5c-a752-be97a4cc2eb6",
      "version": 1,
      "content_hash": "sha256:41490337080b2337c760fdc702b761656e6e25a64debe9a093f0edce0f125c7b"
    },
    "sections": [
      {
        "title": "Average examples instead of trusting one pair",
        "text": "Contrastive Activation Addition builds steering directions from many paired examples of a behavior. It averages differences in internal activations and applies the result while a model generates text. The original tests Llama 2 models. It is closely related to earlier prompt-pair Activation Addition, with a distinct dataset-based construction and evaluation."
      },
      {
        "title": "A small average shows the idea",
        "text": "For an invented arithmetic illustration, suppose three example differences are (2, 0), (4, 2) and (3, 1). Their average is (3, 1). Combining examples can reduce some example-specific variation. It cannot eliminate a bias shared by all three examples, and averaging alone does not prove that a direction captures the intended behavior."
      },
      {
        "title": "The comparison is about the increment",
        "text": "The paper contrasts its many-example construction with single-pair steering and other intervention methods. Source-reported benefits concern specified behavioral evaluations. The existing ideas of internal activations, linear directions and inference-time steering are not counted again as entirely new discoveries. The useful question is what the new construction makes more reliable or easier to study. For an invented averaging example, suppose two prompt pairs give differences (2, 0) and (4, 2). Their mean is (3, 1). Averaging can reduce variation specific to one example, but cannot remove a bias shared by both pairs. A separate evaluation set matters because reusing the construction examples would make the intervention look more general than the check establishes."
      },
      {
        "title": "An evaluator is part of the experiment",
        "text": "Some behavior measurements use another model as a judge. Its rating is a measurement under a particular rubric, not independent proof of a human psychological trait. Effects can depend on the examples, the model, intervention strength and the task. Side effects and transfer to new prompts therefore need their own evidence."
      },
      {
        "title": "What to check next",
        "text": "Trace one reported behavioral comparison to the examples used to form the direction and to the evaluation procedure. Keep construction data and test data separate in your notes. Explain which result was measured and which broader conclusion would need a new experiment. This is a useful source-checking task without claiming that you reproduced the model run."
      }
    ],
    "references": [
      {
        "title": "Steering Llama 2 via Contrastive Activation Addition",
        "url": "https://arxiv.org/abs/2312.06681v1",
        "locator": "Pinned original contribution and stated hypotheses or experimental limits."
      }
    ]
  },
  "reading_status": "EXPLANATORY_GUIDE"
}