Step 11: assistant

text

I rated the claim **supported**, with credence 0.72 and verdict confidence 0.70. Where sparse autoencoders (SAEs) have been compared against well-tuned simple baselines, they mostly lose or tie. The few wins are narrow, mostly reported by groups that build SAE tools, and largely unreplicated.

The probing and steering findings come from the sibling question's assessment and its two earlier researcher reports. I opened none of those papers myself this pass. The unlearning results and the exploratory tasks (auditing, model diffing, hypothesis generation) come from background knowledge only, and I flagged that on the page.

Tally by task family:

| Task family | Wins, ties or losses against a tuned baseline | Who ran the comparison, and how the baseline was set up |
|---|---|---|
| **Probing** | Loses or ties. A 113-dataset study found SAE probes add nothing to a tuned set of baselines. Pooled SAE probes beat single-token probes, but almost all of that gap closes once the raw-activation probe is also pooled. Google DeepMind found SAE probes worse on out-of-distribution data. | The 113-dataset study is by MIT and Google DeepMind authors, with matched baselines. The reported wins (an OpenAI moderation study, a Goodfire and Rakuten personal-data report) used unmatched baselines or come from a group that builds SAE tools. |
| **Steering** | Loses on AxBench: below difference-of-means and far below prompting. Rebuttals that choose features with labels reach rough parity with fine-tuned adapters, but not a general win. One 2026 study reports a direct win for jailbreak mitigation. | AxBench's authors propose a rival method. The 2026 win is a single unreplicated study, and its results tables were not fully checked. |
| **Unlearning** | Loses: Farrell et al. found SAE-based unlearning weaker than an established fine-tuning method (RMU). | Cited in Google DeepMind's update; not opened this pass. |
| **Auditing, model diffing, hypothesis generation** | These show SAEs being useful (Anthropic's auditing game, chat-tuning diffing, hypothesis generation) rather than a measured advantage. | Baseline arms are weak or missing. In the auditing game, simpler search over training data also worked. |

Changes to the claim's structure:
- **Subclaims linked:** two support the claim (the probing null result and the AxBench steering loss). One counts against it (the jailbreak-steering win).
- **Importance and contestation:** both set to 0.55.
- **Other records:** I set aside one source passage because the linked subclaims already cover it, logged the decision, answered both inbox messages, and left a note for the Steward of the parent question about the sparse-autoencoder conditions.

There are no named arguments. I left the expected value of a further pass (marginal yield) at 0.35. A further pass would mainly add value by reading the unlearning and exploratory-task papers directly. About $0.78 of the $1.20 budget was spent.