Step 13: assistant

text

Now record the counterpassage, the AF post instance, and the reading.

Calls record_argument_instance

{
  "url": "https://arxiv.org/pdf/2502.16681",
  "title": "Are Sparse Autoencoders Useful? A Case Study in Sparse Probing",
  "stance": "advances",
  "context": "Section 4.3.1, GLUE CoLA test against LLM-ensemble clean labels (Figure 9).",
  "reading": "premise_unstated",
  "speaker": "Kantamneni, Engels, Rajamanoharan, Tegmark, Nanda",
  "argument_id": "8395f9d1-9885-42bf-9754-d2c1bb76e1f2",
  "source_kind": "preprint",
  "verbatim_text": "Remarkably, when we test all classifiers on examples which were misclassified in CoLA (where the ensembled labels disagree with the original labels), we find that the latent 369585 classifier outperforms the dense SAE probe, which itself outperforms the baseline considerably."
}

Calls record_claim_instance

{
  "url": "https://www.alignmentforum.org/posts/osNKnwiJWHxDYvQTD/takeaways-from-our-recent-work-on-sae-probing",
  "title": "Takeaways From Our Recent Work on SAE Probing",
  "stance": "affirms",
  "context": "The authors' own summary of the paper; same voice as the preprint, covering the multi-token case.",
  "speaker": "Kantamneni and Engels et al.",
  "claim_id": "738d9786-0b87-4b75-9c33-cabe43c903d5",
  "confidence": 0.8,
  "publication": "AI Alignment Forum",
  "source_kind": "blog_post",
  "verbatim_text": "However, when we implemented an attention-pooled baseline that could also leverage information across tokens, the SAE probe win rate dropped to 8.7%."
}

Calls provenance_record_reading

{
  "note": "The body supports the abstract for the conclusions reached: dense baselines flag the same CoLA errors, point to the same punctuation cue, and largely close the pooled-probe gap. It slightly overstates on numbers: on the mislabeled CoLA subset the SAE latent (0.63) and dense SAE probe (0.56) beat logistic regression (0.46), pooled SAE probes still win 8.7% of datasets, and the baseline checks were run after SAEs had suggested the hypotheses. All comparisons are by one team on Gemma-2 SAEs.",
  "support": "overstates",
  "deployment": "Abstract summary of the paper's three interpretability case studies, used to argue SAEs offer no unique probing advantage.",
  "instance_id": "9b4840a6-77c5-448e-8caa-b6477b7c8149",
  "source_kind": "preprint",
  "source_read": true,
  "worth_reading": true,
  "worth_reading_reason": "Sections 4.3 and 5, with Figures 9 and 11, hold all the numbers the claim rests on."
}