Step 25: assistant

thinking

text

I've recorded the claim as **verified** (confidence 0.85, credence 0.92): AxBench's own steering table shows SAE steering below both prompting and difference-in-means, on average and at each of the four sites tested. The later papers don't dispute that measurement. They change how the SAE features are chosen.

**The original result**, read from the AxBench paper's TeX source (arXiv 2501.17148, Wu, Arora et al., ICML 2025):
- **Models and layers:** instruction-tuned Gemma-2-2B (layers 10 and 20) and Gemma-2-9B (layers 20 and 31). For the 2B model the SAEs were trained on the base model, because none existed for the instruction-tuned one.
- **SAEs and feature choice:** pretrained Gemma Scope SAEs. The 500 concepts were drawn from the published text descriptions of Gemma Scope SAE features. The SAE method steers with the feature that description belongs to. A second variant picks the feature with the best detection score (AUROC) on the 144 labelled training examples per concept.
- **How the steering is scored:** a language-model judge rates each steered answer 0 to 2 for concept presence, relevance to the instruction and fluency, and the score is the harmonic mean of the three. Steering strength is chosen per concept on half the instructions and scored on the other half.
- **Baselines:** prompting uses a steering prompt written by a language model and placed before the instruction. Difference-in-means is the normalised gap between average activations on positive and negative examples, added to the activations.
- **Average scores across the four sites:**

| Method | Score (0 to 2) |
|---|---|
| Prompting | 0.894 |
| Difference-in-means | 0.239 |
| SAE (description-matched feature) | 0.165 |
| SAE (best-AUROC feature) | 0.157 |

Difference-in-means beat the SAE on 58.7% of concepts, and the paper calls the gap significant. The gap is small at the last layer of the 9B model (0.158 against 0.140), which is one reason the credence is below 1. The other is that the whole result rests on one judge model and one protocol.

**The later papers:** Arad, Mueller and Belinkov (EMNLP 2025) keep only features with a strong effect on outputs and report two- to threefold gains. Jørgensen and Hansen (2026) choose and label features with a supervised pipeline and report results close to LoRA fine-tuning. I read only their abstracts, so the size of these gains is their own report. I found no paper that re-ran the original setup and got different numbers.

**Scope of the claim:** it is verified for the benchmark's standard SAE setup. It does not establish that every SAE steering method loses on AxBench, and the assessment says so.

**Graph changes:**
- I recorded the AxBench paper as the source asserting the claim, and the two later papers as sources that mention it.
- I linked the two related answers as see-alsos.
- I kept the claim as a single measurement with no subclaims.
- I set importance to 0.4 and contestation to 0.3, and logged the decision.
- I answered both inbox messages.

The parent question's assessment already uses these same figures, so I didn't notify its steward.