Step 7: assistant

Calls update_claim_assessment

{
  "status": "supported",
  "claim_id": "759728cb-0411-4140-925b-d97d8022e57f",
  "assessment": "The evidence favours the claim for the task families where designed comparisons against tuned baselines exist, which are mainly concept detection (probing) and steering, with a smaller body on unlearning. It is thinner and more favourable to sparse autoencoders (SAEs) on exploratory tasks such as auditing, model diffing and hypothesis generation, where clear baselines are harder to define.\n\n**Probing.** The largest comparison, a 113-dataset study on Gemma-2-9B, found that adding SAE probes to a toolkit of tuned baselines did not improve results under scarce data, class imbalance, label noise or covariate shift; earlier apparent wins disappeared against the stronger toolkit ([[claim:9fbf1b17-4bf8-42bc-a5e1-420054bf7463|SAE probes show no consistent advantage over baseline probes]]). Pooled SAE probes beat single-token raw probes on multi-token text, but most of that advantage vanishes once the raw-activation probe is also pooled, leaving a small residual of a few datasets, mostly within noise. Google DeepMind found SAE probes worse out of distribution on harmful-intent detection. Reported detection wins (an OpenAI moderation study, a Goodfire and Rakuten personal-data report) mostly compare against unmatched or weaker baselines, or come from groups that build SAE tooling and are unreplicated.\n\n**Steering.** On AxBench, [[claim:7c1050dc-7ca9-492e-8d3f-ade8d793de4c|description-matched SAE features steer worse than difference-of-means and far worse than prompting]]. Follow-up work recovers rough parity by choosing features with labels or by measured output effect, but does not show a general win. The one direct win located is [[claim:b4a2eb5a-7524-401b-937b-65697fe8d491|SAE-space mean-shift steering beating dense mean-shift for jailbreak mitigation]], especially on held-out attacks: a single unreplicated 2026 study.\n\n**Unlearning.** Work cited in the Google DeepMind update (Farrell et al.) found SAE-based unlearning less effective than established fine-tuning methods; some later SAE-based guardrail methods claim gains, but these have not been independently confirmed.\n\n**Auditing, model diffing, hypothesis generation.** Here SAEs have their most credible practical uses: auditors with SAE access found a planted hidden objective in Anthropic's auditing game, crosscoder and SAE diffing has surfaced chat-tuning-specific features, and SAE-based hypothesis generation has been reported to beat prompting-based baselines on text datasets. These tasks seldom have a tuned simple baseline run alongside, and in the auditing game simpler search over training data also worked, so they show usefulness more than a measured advantage.\n\nTaken together, losses or ties against well-tuned baselines dominate where the comparison has been run, and the documented wins are few, narrow, mostly from interested groups and largely unreplicated. That fits \"rarely\". The claim would weaken if matched, independently replicated wins accumulated, for example label-selected SAE features under distribution shift, or exploratory tasks such as auditing with a proper baseline arm.",
  "confidence": 0.7,
  "claim_credence": 0.72,
  "marginal_yield": 0.35,
  "reasoning_trace": "Scope: the words \"practical downstream interpretability tasks\" are read as tasks with a defined objective where a non-SAE method could be used instead: probing/detection, steering, unlearning, auditing, model diffing. \"Simple baselines\" read as linear probes, difference-of-means, prompting, and fine-tuning methods such as LoRA or RMU.\n\nTally by task family (who ran it; baseline tuning):\n- Probing: Kantamneni et al. 2025 (arXiv 2502.16681, MIT/GDM, not SAE proponents): strong tuned toolkit, SAE adds nothing; multi-token per-dataset table read by an earlier researcher run: best-of-six SAE beats best dense on at most 8 of 46 datasets by test-oracle, only one by more than 0.011 AUC; validation-selected 4/46. GDM progress update: SAE probes worse OOD on harmful intent. Apollo deception probes worse on SAE features. Wins: OpenAI GPT-4o moderation study (pooled SAE against last-token raw, unmatched pooling), Gemma 2 jailbreak/toxicity study (unmatched pooling), Goodfire/Rakuten PII (interested party, different classifier heads). Net: loss/tie when matched.\n- Steering: AxBench (Wu et al. 2025, authors proposing rival method ReFT-r1, so interested against): SAE 0.165 vs DiffMean 0.239 vs prompting 0.894. Rebuttals (arXiv 2505.20063, 2605.31183) reach parity with LoRA on chosen concepts, still below prompting. Win: arXiv 2602.12418 CC-Delta jailbreak mitigation, four models, held-out attacks; authors include a dense-baseline author; unreplicated; tables not fully checked.\n- Unlearning: Farrell et al. 2024 negative vs RMU (cited in GDM update). Positive SAE guardrail claims exist from memory but not verified this pass.\n- Auditing/diffing/hypothesis generation: stated from background knowledge, not read this pass (Anthropic auditing game 2025; crosscoder chat-diffing work; HypotheSAEs). Baselines weak or absent.\nSources on the page: GDM negative-results update (authors deprioritised SAE research on this basis; a group that had invested in SAEs reporting against them is an admission against interest, which adds weight) and the 2026 random-baseline sanity-check paper, which restates the claim citing AxBench and GDM (derivative, not independent).\nCredence 0.72 rather than higher because \"rarely\" is vague, exploratory-task evidence is under-examined, and recent label-guided steering results narrow the gap. Leaned on the sibling question's assessment and its researcher reports for the probing and steering details; unlearning and exploratory-task items were not read this pass."
}

Calls log_stewardship_decision

{
  "claim_id": "759728cb-0411-4140-925b-d97d8022e57f",
  "reasoning": "First pass on small budget. Linked sibling answers as basis: probing null (supports), AxBench steering loss (supports), SAE-space jailbreak steering win (contradicts). No named arguments needed. Dismissed one unplaced passage (covered by basis). Importance 0.55, contestation 0.55. Assessed supported, credence 0.72, tallied by task family per the plan; unlearning and exploratory-task evidence from background knowledge, flagged for a later pass.",
  "action_taken": "structured_and_assessed"
}

Calls answer_messages

{
  "answers": [
    {
      "note": "Structured with three basis subclaims and assessed as supported.",
      "message": "m1",
      "outcome": "changed"
    },
    {
      "note": "Assessment tallies probing, steering, unlearning, auditing/diffing by who ran the comparison and baseline tuning; documented wins (jailbreak steering, pooled detection, auditing game, hypothesis generation) are listed alongside losses. Unlearning and exploratory-task items rest on background knowledge, not read sources, this pass.",
      "message": "m2",
      "outcome": "changed"
    }
  ]
}