Step 16: tool results

{
  "content": "A sparse dictionary enumerates many interpretable features without being told what to look for, and a researcher can then select those that satisfy a criterion and test them by other means. Used this way, sparse dictionaries have surfaced mechanisms that researchers had not specified in advance, such as [[claim:5c397b68-66c7-4e8e-b2b4-22d77ac1ec06|a model choosing its rhyme word before writing the line]], and in one audit [[claim:391dac24-440f-438b-8cff-8933dfac302c|SAE tools raised an automated agent's success at finding a hidden objective]]. Methods that need a concept specified in advance, such as probes and steering vectors, cannot carry out this enumeration. So SAEs are powerful tools for discovering unknown concepts in language models.",
  "argument_id": "09caf9a4-6968-4c4c-9fd3-4eaa6b5adda6"
}

{
  "content": "That a tool can find something shows little about whether it is powerful; what shows that is how it does against the alternatives on the same task. When SAE-based discovery has been run beside simpler methods on discovery tasks, the simpler methods matched or beat it: on AuditBench [[claim:461a798c-19b2-4b7a-9f26-aa1dfed627fb|SAE tools gave little benefit]] while prompting-based elicitation gained much; in output-based model diffing [[claim:846a4d41-3752-45ed-9210-b4177d75b5e9|a strengthened LLM baseline performed comparably]]; and the fastest SAE success in the blind audit [[claim:5393957c-bef5-4262-a152-84c5fe63ece2|could have been replicated by embedding search]]. So the evidence does not show that SAEs are powerful discovery tools.",
  "argument_id": "237dd1f1-d6f3-41b7-a81d-f132bb376640"
}

{
  "stance": "advances",
  "reading": "premise_unstated",
  "argument_id": "09caf9a4-6968-4c4c-9fd3-4eaa6b5adda6",
  "instance_id": "6aedce88-e38e-4a16-9b0b-fb637fb6def9"
}

{
  "stance": "advances",
  "reading": "premise_unstated",
  "argument_id": "09caf9a4-6968-4c4c-9fd3-4eaa6b5adda6",
  "instance_id": "1c248dc7-4cd3-4340-9b76-318a3c8b7188"
}