Step 23: assistant

text

I've rated the claim **doubtful**, confidence 0.75, credence 0.2: the comparison behind it isn't like for like. Even its own numbers leave sparse autoencoder (SAE) steering behind the supervised methods.

**What the source reports.** I read the source paper (arXiv 2505.20063, EMNLP 2025), Section 4.4, Table 1 and the AxBench appendix. The setup matches the original benchmark's Gemma-2-9B-it setting: Concept500, layers 20 and 31, and the same three judge metrics combined by harmonic mean. Keeping only features with output score at least 0.1, SAE steering scores 0.546 at layer 20 and 0.470 at layer 31.

**Why the comparison doesn't hold:**
- **Baselines not re-run.** Table 1 copies them from the AxBench paper.
- **Different judge.** The new SAE scores were rated by Claude 3.7 Sonnet. AxBench used gpt-4o-mini.
- **Different instructions.** AxBench never released its instruction sample, so the authors drew a new one.
- **The pipeline alone inflates SAE scores.** The authors' own *unfiltered* SAE run scores 0.293 and 0.387, against AxBench's 0.191 and 0.140. That is 1.5 to 2.8 times higher before any feature selection.
- **The filter changes the set of concepts scored.** It averages only over concepts whose features pass the threshold, while the baselines cover all 500. The paper does not say how many were kept.

**What "competitive" amounts to at face value:**

| Method (AxBench-reported) | Layer 20 | Layer 31 |
|---|---|---|
| Filtered SAE (this paper) | 0.546 | 0.470 |
| ReFT-r1 | 0.630 | 0.401 |
| LoRA | 0.602 | 0.580 |
| LoReFT | 0.777 | 0.764 |
| Prompting | 1.075 | 1.072 |

Filtered SAEs beat ReFT-r1 at layer 31 and difference-in-means at both layers, and trail every other method listed. The paper's "90.7% of the best method" at layer 20 equals 0.546/0.602, which is the ratio to LoRA, not to the best method. The like-for-like effect of the filter itself is real: 0.293 to 0.546 and 0.387 to 0.470. It is not steady across thresholds, though: at layer 20 a 0.01 cutoff scores lower than a 0.001 cutoff.

**Outside reproduction.** One outside paper (arXiv 2609.35367) tested the output-score filter on Gemma-2-2B layer 20 under AxBench's own protocol. It did not find the reported gains. It did not test the 9B setting.

**What I put on the graph:**
- **Arguments.** One for (rated slight) and one against (rated likely).
- **New subclaim.** It covers the narrower result: selecting by output score improves SAEs over unfiltered ones. I gave it an initial credence of 0.65; its own pass hasn't run yet.
- **Sources and links.** I recorded the outside paper as a source and linked a related claim about poor feature selection partly explaining AxBench's result. I also logged a provenance reading marking the source as overstating its own evidence.
- **Settings.** Importance 0.4, contestation 0.45.
- **Messages.** Both inbox messages are answered. Nothing depends on this claim yet, so there was no one to notify.

What would settle it is a matched re-run on Gemma-2-9B: all methods on the same concepts, instructions and judge, reporting how many concepts the filter keeps. The run used about $0.89 of the $1.20 budget.