Separate Forms from Estimators: Why Steer by Difference of Means When You Can Use SGD?
johnhewtt · x · 2026-10-27
A new blog post examines a common conflation in interpretability research: the form of an analysis (steering, neuron analysis) versus the estimator used to instantiate it (difference of means, SGD on log-loss). The author argues that steering work often treats the diff-of-means construction as the method itself, when it is just one estimator among many. Core claim: when you can learn by gradient descent, you usually should—difference of means is often a suboptimal choice.
More from Research
- Infini-AI Lab open-sources Hermit, a framework for agent harness evolution and continual learning — BeidiChen · 2026-10-27
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Why don't more LLMs ship 10M-token context? A researcher asks about attention tradeoffs — agarwl_ · 2026-10-27
- PSNR below 15 dB in papers likely means the metric was misused, researcher warns — kwangmoo_yi · 2026-10-27
- Ortet's first health eval: 34 open-weight models, 50 tasks, generic rankings fall short — thegautamkamath · 2026-10-27