Step 11: assistant

text

I assessed the claim as **supported**, with a credence of 0.85 and a verdict confidence of 0.78. I left it as a single claim with no subclaims, and I only read the abstracts of the three newer papers, not the papers themselves.

The claim says probes on sparse-autoencoder (SAE) features do not reliably beat ordinary probes on the model's raw activations. It is weakly worded ("not consistently", not "never"), so occasional wins don't count against it.

Study by study:
- **arXiv 2502.16681 (Kantamneni et al., 2025; authors include Nanda and Rajamanoharan).** This is the main evidence. It used Gemma 2 with Gemma Scope SAEs, about 113 datasets, and tuned baselines. It tested scarce data, class imbalance, label noise and distribution shift, and found no meaningful improvement in any of them. "Consistently" was measured two ways: how often SAE probes win, and whether adding them to a set of candidate probes improves test scores. SAE probes did win on some individual datasets. A multi-token advantage (19.6% win rate) fell to 8.7% once the baselines also got attention pooling. The model, SAE and baseline details come partly from my background knowledge of the paper, not from this pass.
- **DeepMind's "Negative results for SAEs on downstream tasks" post.** SAE probes for harmful intent did worse than linear probes on out-of-distribution data. It shares authors with the paper above, so it is not an independent line. I did not open it this pass.
- **arXiv 2602.14111 (Korznikov et al., 2026; a separate group).** It found SAEs with randomised features nearly match trained ones on sparse probing (0.69 vs 0.72). This compares SAEs with random SAEs, not with raw-activation probes, so it supports the claim only indirectly. I recorded it on the page as a mention.
- **arXiv 2505.20063 (Arad, Mueller, Belinkov) and 2605.31183 (Jørgensen, Hansen).** These are about steering, not probing, so they don't bear directly on this claim. They are the main pro-SAE results. I told the parent question's Steward they belong on the steering side of its answer set.

The main caveat is independence: most of the controlled probing evidence comes from one group and one model family.

Open for a later pass (marginal yield set to 0.35):
- Full-text reading of the three newer papers.
- The seven unplaced argument passages.
- Named arguments.

I also set importance to 0.5 and contestation to 0.45, logged both passes, and answered both inbox messages.