With features selected and labelled by a supervised pipeline, SAE steering approaches LoRA performance on AxBench.
First assessment
Read arXiv 2605.31183 in full including Figure 2 from TeX source; compared setup with AxBench (2501.17148) and the sibling assessment of output-score selection. Assessed inconclusive (0.65, credence 0.45): numbers near the quoted LoRA reference for about half of forums, but LoRA not re-run, concepts are forum tags chosen to fit features, different layers, per-feature strength tuning, and prompting rises on these labels. Added for/against arguments (slight / likely), placed the unplaced passage, linked the output-score sibling as related, recorded reading and provenance edge, set importance 0.35, contestation 0.5. No new subclaims: premises are settled facts about the paper's setup, stated in prose.
Assessed Inconclusive
verdict confidence 0.65 · credence 0.45
The claim comes from a single 2026 preprint by Jørgensen and Hansen, a partial rebuttal of AxBench (Wu et al., 2025), which had found sparse autoencoder (SAE) steering far behind prompting, fine-tuning and simple representation baselines. The rebuttal labels Gemma Scope SAE features on Gemma-2-9b-it using Stack Exchange posts and their tags: each feature is treated as a probe, matched to a tag by a calibrated F1 score, filtered by an output score that proxies steering ability, and the top 50 feature-label pairs per forum are kept. These pairs are then steered and rated with AxBench's three-part, LLM-judged harmonic-mean score. The reported numbers are real as far as they go. Read from the paper's main figure, forum means for the selected features range from about 0.3 to 0.7, and about half of the eleven forums sit near the LoRA reference of about 0.59 that AxBench reported for this model, against about 0.1 to 0.2 for randomly chosen Neuronpedia-labelled features under the same evaluation code. Selection and labelling clearly turn SAE features into much better steering vectors than AxBench's setup did. Whether this means SAE steering approaches LoRA on AxBench is not established. The LoRA figure was quoted, not re-run. It was measured on AxBench's own Concept500 concepts at layers 20 and 31, while the SAE scores are for different concepts, broad forum tags that survived because a feature matched them, at layers 17 and 32, with the steering strength tuned per feature. Prompting on the same tags scores about 1.2, above its AxBench figure of about 1.07, so these concepts appear easier, and a LoRA trained on them would plausibly score higher too. The method is also no longer the unsupervised use of SAEs that AxBench tested: labelled data choose which features are used and what they are said to mean, so the result reverses the benchmark's task, finding concepts for good features rather than steering toward given concepts. Prompting stays roughly twice as strong, as the authors acknowledge. A LoRA run on the same tags, layers, instructions and judge would settle the comparison.
Updated claim instance
Instance 6912e235-b510-4be2-9a4c-5eb12f2c7841 (https://arxiv.org/pdf/2605.31183): was stance=affirms, confidence=0.85; set confidence=0.95, speaker="Mikkel Godsk Jørgensen and Lars Kai Hansen", publication="arXiv preprint 2605.31183", sourceDate="2026-05-29". Filled authors and date from the paper's first page; the abstract passage states the claim directly.
Claim entered the graph