Step 14: tool results

search_similar_claims

{
  "limit": 6,
  "query": "SAE features selected by output score steer well (Arad et al. 2025)"
}

{
  "query": "SAE features selected by output score steer well (Arad et al. 2025)",
  "results": [
    {
      "id": "3071aaa9-dbb1-4c25-836e-b7441da8ac21",
      "kind": "claim",
      "text": "Selecting SAE features by high output score improves their steering performance relative to unfiltered SAE features.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.7820493047659979,
      "importance": 0.35,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "d3311aac-502f-4e69-ab97-53ee2b6de622",
      "kind": "claim",
      "text": "With output-score feature selection, SAE steering is competitive with supervised steering methods on AxBench.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.752891591414091,
      "importance": 0.4,
      "assessment_status": "doubtful",
      "assessment_confidence": 0.75
    },
    {
      "id": "22ec9a4c-c85b-4dff-bd65-ca53e2ba04b8",
      "kind": "claim",
      "text": "Poor SAE feature selection partly explains AxBench's negative results for SAE steering.",
      "claim_type": "causal",
      "state": "active",
      "similarity_score": 0.7095644679660719,
      "importance": 0.55,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "a5665880-9436-4582-a906-e2161fe8e71a",
      "kind": "claim",
      "text": "With features selected and labelled by a supervised pipeline, SAE steering approaches LoRA performance on AxBench.",
      "claim_type": "empirical_verifiable",
      "state": "active",
      "similarity_score": 0.6789410957520934,
      "importance": 0.5,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "d9d5b8aa-c845-4b43-896d-da0edecdfb7e",
      "kind": "claim",
      "text": "Ranking SAE features by absolute steering effect at large intervention scales surfaces reward-model-bias features in studied demonstrations.",
      "claim_type": "empirical_derived",
      "state": "active",
      "similarity_score": 0.626521919160526,
      "importance": 0.3,
      "assessment_status": null,
      "assessment_confidence": null
    },
    {
      "id": "c2fa883c-e27d-4562-bf8c-b39988887757",
      "kind": "claim",
      "text": "For Gemma-2-9B-it on AxBench Concept500, prompting outperforms output-score-selected SAE steering at layers 20 and 31.",
      "claim_type": "empirical_verifiable",
      "state": "active",
      "similarity_score": 0.622878274568136,
      "importance": 0.45,
      "assessment_status": null,
      "assessment_confidence": null
    }
  ],
  "count": 6
}