Step 10: tool results

search_similar_claims

{
  "limit": 8,
  "query": "SAE-based model diffing suits low-level details such as tokens and formatting, LLM-based methods suit abstract behavioral differences"
}

{
  "query": "SAE-based model diffing suits low-level details such as tokens and formatting, LLM-based methods suit abstract behavioral differences",
  "results": [
    {
      "id": "653e653c-cda0-4e4c-bfe2-de7f99acee34",
      "kind": "claim",
      "text": "SAE methods suit low-level model details, while LLM methods surface abstract behavioral differences.",
      "claim_type": "evaluative",
      "state": "active",
      "similarity_score": 0.845659140850616,
      "importance": 0.45,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "846a4d41-3752-45ed-9210-b4177d75b5e9",
      "kind": "claim",
      "text": "In API-only model diffing, an improved LLM baseline performs comparably to SAE-based methods and surfaces more abstract behavioral differences.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.7613676904002139,
      "importance": 0.5,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "5a169f7a-a36c-4a9e-b730-1708978afe7d",
      "kind": "claim",
      "text": "LLM-based model diffing yields higher hypothesis acceptance rates than SAE-based model diffing.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.717386842479661,
      "importance": 0.3,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "d40577aa-06f7-469e-9dc5-d0e07ea3fcdd",
      "kind": "claim",
      "text": "Token-level model-diffing methods inherently struggle to generate abstract behavioral hypotheses.",
      "claim_type": "causal",
      "state": "active",
      "similarity_score": 0.6517810361367715,
      "importance": 0.45,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "ba9643d1-d067-43c9-8bb2-32e15e929fec",
      "kind": "claim",
      "text": "API-only model diffing cannot surface behavioral differences that do not manifest in model outputs.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.5835391796773453,
      "importance": 0.35,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "62551d05-8af2-4a43-93f1-e5b06f0f60f4",
      "kind": "claim",
      "text": "SAE-based and LLM-based API-only diffing both identify harmful advice and actions in risky-finance-finetuned Qwen-2.5-7B-Instruct.",
      "claim_type": "empirical_verifiable",
      "state": "active",
      "similarity_score": 0.5615931619448109,
      "importance": 0.3,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "e6e4fb20-be4f-4113-895f-877cd28ec6ca",
      "kind": "claim",
      "text": "Delta-Crosscoder identifies fine-tuning-induced directions with less feature inspection than SAE model diffing.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.5500655651791577,
      "importance": 0.3,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "fbdaf39c-c588-44f6-a759-42045a479af6",
      "kind": "claim",
      "text": "At L0=160, fully trained SAEs significantly outperform frozen variants in most sparse-probing and causal-editing comparisons.",
      "claim_type": "empirical_verifiable",
      "state": "active",
      "similarity_score": 0.5404630761197159,
      "importance": 0.4,
      "assessment_status": null,
      "assessment_confidence": null
    }
  ],
  "count": 8
}