{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "arditi-et-al-refusal-direction-2024",
    "title": "Arditi and colleagues: a low-dimensional representation mediating model refusal",
    "summary": "Identified a direction in residual-stream activations whose controlled intervention substantially changes refusal behavior in the tested instruction-tuned language models, linking an activation measurement to a causal behavioral effect.",
    "boundary": "The result concerns selected prompts, intervention choices and thirteen tested models, not a theorem that every refusal in every model has one-dimensional cause. Refusal detection and harmful-response detection measure different outcomes.",
    "area": "Language-model interpretability",
    "topic": "Artificial intelligence",
    "year": 2024,
    "lane": "HUMAN",
    "result": "EMPIRICAL",
    "package_id": "arditi-et-al-refusal-direction-2024",
    "target": {
      "record_id": "urn:uuid:2524c4f3-37d5-4bb2-a8ad-caced5406f5d",
      "version": 1,
      "content_hash": "sha256:a11f3fba00d034ea8863fa404627642961755473f7cc7a375297f8c5fb7417f6"
    },
    "humans": [
      "Andy Arditi",
      "Oscar Obeso",
      "Aaquib Syed",
      "Daniel Paleka",
      "Nina Rimsky",
      "Wes Gurnee",
      "Neel Nanda"
    ],
    "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-2406-11717v1",
        "title": "Refusal in Language Models Is Mediated by a Single Direction",
        "url": "https://arxiv.org/abs/2406.11717v1",
        "date": "2024-06-17"
      }
    ],
    "points": "90000",
    "published_grade": {
      "id": "arditi-et-al-refusal-direction-2024",
      "entry_id": "arditi-et-al-refusal-direction-2024",
      "band": "A",
      "modifier": "-",
      "target": {
        "record_id": "urn:uuid:2524c4f3-37d5-4bb2-a8ad-caced5406f5d",
        "version": 1,
        "content_hash": "sha256:a11f3fba00d034ea8863fa404627642961755473f7cc7a375297f8c5fb7417f6"
      },
      "what_changed": "Identified a direction in residual-stream activations whose controlled intervention substantially changes refusal behavior in the tested instruction-tuned language models, linking an activation measurement to a causal behavioral effect.",
      "scope": "Identified a direction in residual-stream activations whose controlled intervention substantially changes refusal behavior in the tested instruction-tuned language models, linking an activation measurement to a causal behavioral effect. The result concerns selected prompts, intervention choices and thirteen tested models, not a theorem that every refusal in every model has one-dimensional cause. Refusal detection and harmful-response detection measure different outcomes.",
      "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:2524c4f3-37d5-4bb2-a8ad-caced5406f5d",
        "version": 1,
        "content_hash": "sha256:a11f3fba00d034ea8863fa404627642961755473f7cc7a375297f8c5fb7417f6"
      },
      "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": "Andy Arditi",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "discoverer"
          ],
          "source": {
            "citation": "Refusal in Language Models Is Mediated by a Single Direction",
            "url": "https://arxiv.org/abs/2406.11717v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Oscar Obeso",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "discoverer"
          ],
          "source": {
            "citation": "Refusal in Language Models Is Mediated by a Single Direction",
            "url": "https://arxiv.org/abs/2406.11717v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Aaquib Syed",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "discoverer"
          ],
          "source": {
            "citation": "Refusal in Language Models Is Mediated by a Single Direction",
            "url": "https://arxiv.org/abs/2406.11717v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Daniel Paleka",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "discoverer"
          ],
          "source": {
            "citation": "Refusal in Language Models Is Mediated by a Single Direction",
            "url": "https://arxiv.org/abs/2406.11717v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Nina Rimsky",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "discoverer"
          ],
          "source": {
            "citation": "Refusal in Language Models Is Mediated by a Single Direction",
            "url": "https://arxiv.org/abs/2406.11717v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Wes Gurnee",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "discoverer"
          ],
          "source": {
            "citation": "Refusal in Language Models Is Mediated by a Single Direction",
            "url": "https://arxiv.org/abs/2406.11717v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Neel Nanda",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "discoverer"
          ],
          "source": {
            "citation": "Refusal in Language Models Is Mediated by a Single Direction",
            "url": "https://arxiv.org/abs/2406.11717v1"
          }
        }
      ],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "Refusal in Language Models Is Mediated by a Single Direction",
            "url": "https://arxiv.org/abs/2406.11717v1"
          }
        }
      ],
      "origin": "HUMAN_LEGACY",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "The result concerns selected prompts, intervention choices and thirteen tested models, not a theorem that every refusal in every model has one-dimensional cause. Refusal detection and harmful-response detection measure different outcomes.",
        "domain": "Language-model interpretability"
      },
      "sources": [
        {
          "citation": "Refusal in Language Models Is Mediated by a Single Direction",
          "url": "https://arxiv.org/abs/2406.11717v1"
        }
      ],
      "statement": "# Arditi and colleagues: a low-dimensional representation mediating model refusal\n\n## Can a small internal change alter an answer?\n\nA language model processes information through long lists of numbers called activations. Researchers can sometimes identify a direction in those numbers that is associated with a behavior. This study goes beyond spotting a correlation: it changes activations and measures the resulting behavior. In thirteen tested instruction-tuned models, a compact direction strongly affected refusal behavior.\n\n## Correlation and intervention differ\n\nConsider an invented weather example. Umbrellas are associated with rain, but removing umbrellas does not stop the rain. To argue that a model feature helps cause a behavior, researchers need an intervention, suitable controls and measurements of what else changed. An activation direction associated with refusal is therefore a starting point, not by itself a complete causal explanation.\n\n## A simple vector picture\n\nImagine an arrow on a two-dimensional sheet. Moving a point along that arrow changes one combination of its coordinates. Real model activations have many more dimensions, but the geometric idea survives: a directional intervention can alter a selected component while leaving other components unchanged algebraically. That does not guarantee the model's other behaviors remain unchanged, because later computations combine those components. A useful comparison holds the questions and generation settings fixed while changing the tested intervention. Otherwise a different answer could reflect a different prompt or random sampling. Keep a separate set of harmless questions for checking unintended effects. This helps explain a causal experiment without assuming that one direction completely describes the model's behavior.\n\n## The boundary matters\n\nThe finding concerns the tested models, prompts and intervention procedures. It is not a theorem that all refusal mechanisms are one-dimensional, or that refusal and harmful-response detection are identical. The original discussion also limits broader generalization and some of its mechanism analyses. The human researchers receive credit for designing and testing the intervention; the experimental models are the objects studied.\n\n## Check the scientific argument\n\nRead the original results with three columns in mind: the behavior measured, the intervention made and the control used. Ask whether a change in refusal also changed unrelated capabilities. A source comparison or a diagram of these relationships is a bounded contribution that does not require bypassing an assistant's safeguards. Keep a causal claim restricted to the actual intervention evidence.\n\n## Scientific scope\n\nThe result concerns selected prompts, intervention choices and thirteen tested models, not a theorem that every refusal in every model has one-dimensional cause. Refusal detection and harmful-response detection measure different outcomes.\n\n## Sources for this explanation\n\n- [Refusal in Language Models Is Mediated by a Single Direction](https://arxiv.org/abs/2406.11717v1): Pinned original contribution and stated hypotheses or experimental limits.\n",
      "summary": "Identified a direction in residual-stream activations whose controlled intervention substantially changes refusal behavior in the tested instruction-tuned language models, linking an activation measurement to a causal behavioral effect.",
      "title": "Arditi and colleagues: a low-dimensional representation mediating model refusal"
    },
    "record_id": "urn:uuid:2524c4f3-37d5-4bb2-a8ad-caced5406f5d",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:2524c4f3-37d5-4bb2-a8ad-caced5406f5d",
      "version": 1,
      "content_hash": "sha256:a11f3fba00d034ea8863fa404627642961755473f7cc7a375297f8c5fb7417f6"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:75391e5a9021d93d82fec8c08d396faa9b38860ce560aac85078af1f1e261005"
  },
  "target": {
    "record_id": "urn:uuid:2524c4f3-37d5-4bb2-a8ad-caced5406f5d",
    "version": 1,
    "content_hash": "sha256:a11f3fba00d034ea8863fa404627642961755473f7cc7a375297f8c5fb7417f6"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "arditi-et-al-refusal-direction-2024",
    "entry_id": "arditi-et-al-refusal-direction-2024",
    "revision": 1,
    "title": "Arditi and colleagues: a low-dimensional representation mediating model refusal",
    "context": "Language-model interpretability",
    "track": "Human-designed model analysis or safeguard method",
    "band": "A",
    "modifier": "-",
    "status": "PROVISIONAL_EDITORIAL",
    "assessed_on": "2026-10-23",
    "evidence_cutoff": "2026-10-23",
    "achievement_state": "SUPPORTED_SCOPED_RESULT",
    "what_changed": "Identified a direction in residual-stream activations whose controlled intervention substantially changes refusal behavior in the tested instruction-tuned language models, linking an activation measurement to a causal behavioral effect.",
    "scope": "Identified a direction in residual-stream activations whose controlled intervention substantially changes refusal behavior in the tested instruction-tuned language models, linking an activation measurement to a causal behavioral effect. The result concerns selected prompts, intervention choices and thirteen tested models, not a theorem that every refusal in every model has one-dimensional cause. Refusal detection and harmful-response detection measure different outcomes.",
    "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. The result concerns selected prompts, intervention choices and thirteen tested models, not a theorem that every refusal in every model has one-dimensional cause. Refusal detection and harmful-response detection measure different outcomes.",
    "references": [
      {
        "title": "Refusal in Language Models Is Mediated by a Single Direction",
        "url": "https://arxiv.org/abs/2406.11717v1",
        "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:2524c4f3-37d5-4bb2-a8ad-caced5406f5d",
      "version": 1,
      "content_hash": "sha256:a11f3fba00d034ea8863fa404627642961755473f7cc7a375297f8c5fb7417f6"
    },
    "scientific_correctness_certified": false,
    "independent_expert_calibration": false,
    "method": "PROPRIETARY",
    "points": "90000"
  },
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": {
    "id": "arditi-et-al-refusal-direction-2024",
    "target": {
      "record_id": "urn:uuid:2524c4f3-37d5-4bb2-a8ad-caced5406f5d",
      "version": 1,
      "content_hash": "sha256:a11f3fba00d034ea8863fa404627642961755473f7cc7a375297f8c5fb7417f6"
    },
    "sections": [
      {
        "title": "Can a small internal change alter an answer?",
        "text": "A language model processes information through long lists of numbers called activations. Researchers can sometimes identify a direction in those numbers that is associated with a behavior. This study goes beyond spotting a correlation: it changes activations and measures the resulting behavior. In thirteen tested instruction-tuned models, a compact direction strongly affected refusal behavior."
      },
      {
        "title": "Correlation and intervention differ",
        "text": "Consider an invented weather example. Umbrellas are associated with rain, but removing umbrellas does not stop the rain. To argue that a model feature helps cause a behavior, researchers need an intervention, suitable controls and measurements of what else changed. An activation direction associated with refusal is therefore a starting point, not by itself a complete causal explanation."
      },
      {
        "title": "A simple vector picture",
        "text": "Imagine an arrow on a two-dimensional sheet. Moving a point along that arrow changes one combination of its coordinates. Real model activations have many more dimensions, but the geometric idea survives: a directional intervention can alter a selected component while leaving other components unchanged algebraically. That does not guarantee the model's other behaviors remain unchanged, because later computations combine those components. A useful comparison holds the questions and generation settings fixed while changing the tested intervention. Otherwise a different answer could reflect a different prompt or random sampling. Keep a separate set of harmless questions for checking unintended effects. This helps explain a causal experiment without assuming that one direction completely describes the model's behavior."
      },
      {
        "title": "The boundary matters",
        "text": "The finding concerns the tested models, prompts and intervention procedures. It is not a theorem that all refusal mechanisms are one-dimensional, or that refusal and harmful-response detection are identical. The original discussion also limits broader generalization and some of its mechanism analyses. The human researchers receive credit for designing and testing the intervention; the experimental models are the objects studied."
      },
      {
        "title": "Check the scientific argument",
        "text": "Read the original results with three columns in mind: the behavior measured, the intervention made and the control used. Ask whether a change in refusal also changed unrelated capabilities. A source comparison or a diagram of these relationships is a bounded contribution that does not require bypassing an assistant's safeguards. Keep a causal claim restricted to the actual intervention evidence."
      }
    ],
    "references": [
      {
        "title": "Refusal in Language Models Is Mediated by a Single Direction",
        "url": "https://arxiv.org/abs/2406.11717v1",
        "locator": "Pinned original contribution and stated hypotheses or experimental limits."
      }
    ]
  },
  "reading_status": "EXPLANATORY_GUIDE"
}