Step 33: assistant

text

I've recorded a first assessment of this claim: **inconclusive**, confidence 0.65, credence 0.45 that it is true as worded. The result is real, but the comparison behind "approaches LoRA" was never run like for like.

**What the paper reports.** The claim comes from one preprint (Jørgensen and Hansen, arXiv 2605.31183). Reading its main figure, SAE steering with the selected features scores between about 0.3 and 0.7 per forum on AxBench's combined rating. About half of the 11 forums sit near the LoRA figure of about 0.59 that the original benchmark (arXiv 2501.17148) reported for Gemma-2-9B. The unweighted mean is about 0.53, roughly 90% of that LoRA figure. The same evaluation code gives about 0.1 to 0.2 for randomly chosen Neuronpedia-labelled features, close to the original SAE result. So the evaluation code itself does not seem to raise scores, and the gain comes from how features are chosen. These per-forum numbers are my estimates read off the plot, accurate to about ±0.03.

**How the setup differs from the original benchmark:**
- **Model:** the same, Gemma-2-9b-it.
- **Layers:** 17 and 32 here, against 20 and 31 in the original.
- **SAEs:** the sparsest Gemma Scope SAEs here; the original used the denser versions that Neuronpedia annotates.
- **Concepts:** Stack Exchange user tags, kept because some feature matched them well (the top 50 feature–tag pairs per forum). The original used its own fixed concept set (Concept500).
- **Metric:** the same three-part judged score, combined by harmonic mean. The paper names GPT-4o as the judge in one place and GPT-4o-mini in another; the original used GPT-4o-mini.
- **Steering strength:** tuned separately for each feature, where the original used one strength for all features.
- **Baselines:** LoRA, LoReFT, ReFT-r1, prompting and SAE figures are quoted from the original paper, not re-run. Only a random-feature SAE baseline and a prompting baseline were re-run.

**Is it still the unsupervised method the benchmark tested?** No. Labelled data decide which features are used and what each is said to mean, and four pipeline settings were tuned by hand on concept scores. That turns the benchmark's task around: it finds concepts that fit good features instead of steering toward concepts it was given. The paper's own ablation fits this. Features picked by the pipeline steer better even with their original Neuronpedia labels, and dropping the output-score filter costs only a little.

**Why not "supported".** Prompting on these tags scores about 1.2, above its original figure of about 1.07, which suggests the tags are easier concepts. A LoRA trained on them would plausibly score higher too. A LoRA run on the same tags, layers, instructions and judge would settle the claim. I didn't find any independent replication.

**Against the sibling answers.** The original benchmark result stays verified; this claim doesn't contradict it, because the setup is different. The output-score claim was rated doubtful for the same problem of comparing against quoted, not re-run, baselines. This claim is weaker, comparing against LoRA only and saying "approaches", and its numbers are closer, so I rated it inconclusive rather than doubtful.

**What I recorded on the page:**
- One argument for (the reported scores), rated slight support even if its premises hold. One argument against (the unmatched comparison), rated likely.
- The source passage that had no argument is now placed on the "for" argument.
- The output-score claim is linked as related.
- A reading of the source and its link to the original benchmark, plus a short source summary for readers.
- Importance 0.35, contestation 0.5. I added no subclaims; the facts about the setup are stated in the assessment text.
- Both inbox messages are answered, and I left a note for the Steward of the question this claim answers.

About $0.87 of the $1.20 budget was used.