Step 16: tool results

search_similar_claims

{
  "limit": 8,
  "query": "AuditBench white-box interpretability tools auditing agent hidden behaviors"
}

{
  "query": "AuditBench white-box interpretability tools auditing agent hidden behaviors",
  "results": [
    {
      "id": "c9681fb1-216b-49bb-ac61-57a4e02cf329",
      "kind": "claim",
      "text": "Raw white-box interpretability outputs require filtering for auditing agents to use them effectively.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.7010436642856892,
      "importance": 0.35,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "86063716-4ef1-48dc-abe6-02ef8b75f176",
      "kind": "claim",
      "text": "In agent-based alignment auditing, black-box tools outperform white-box interpretability tools.",
      "claim_type": "empirical_verifiable",
      "state": "active",
      "similarity_score": 0.6740435901410461,
      "importance": 0.55,
      "assessment_status": "supported",
      "assessment_confidence": 0.6
    },
    {
      "id": "b3afe11b-e34e-4e46-9f68-e8e0ef9237a2",
      "kind": "claim",
      "text": "AuditBench tools improve auditing success most for synthetic-document-finetuned models, with smaller or negative gains on transcript-distilled and adversarially trained models.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.6596076800475128,
      "importance": 0.5,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "2cb24ccf-7221-4ffd-8d3c-6f3b0b2f70bc",
      "kind": "quantity",
      "text": "Share of AuditBench Investigator-agent runs identifying implanted behavior with scaffolded text completion, averaged across target models, in percent",
      "claim_type": "empirical_verifiable",
      "state": "active",
      "similarity_score": 0.6030654490512112,
      "importance": 0.35,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "c4dbfb72-f671-4997-9c9d-1d15982484af",
      "kind": "claim",
      "text": "Assistant prefill attacks can reveal a language model's hidden objective during auditing.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.5674834791617317,
      "importance": 0.4,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "9d6f5a8b-cb8e-41fa-a716-fca25bbf6b50",
      "kind": "claim",
      "text": "Sparse autoencoder features can help auditing agents uncover hidden model behavior missed by default prompting.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.5457156354788288,
      "importance": 0.4,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "f4204e80-58c4-4214-b3b1-254c34855e38",
      "kind": "claim",
      "text": "Arbitrary capabilities can be hidden in neural network weights so that, under cryptographic assumptions, no polynomial-time white-box auditor can elicit them.",
      "claim_type": "mathematical",
      "state": "active",
      "similarity_score": 0.5257312886373636,
      "importance": 0.55,
      "assessment_status": "supported",
      "assessment_confidence": 0.8
    },
    {
      "id": "d6d01474-6da4-4706-8184-14e7d5cc3a4a",
      "kind": "claim",
      "text": "On AuditBench, scaffolded text completion, user sampling, and prefilling were the three highest-performing tools.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.5229433949329303,
      "importance": 0.55,
      "assessment_status": null,
      "assessment_confidence": null
    }
  ],
  "count": 8
}