Helm activation steering: only 10 of 53 pushes held, author shares full data
leolambertini · reddit · 2026-10-29
The author of Helm, a ComfyUI activation-steering tool, published honest eval numbers: only 10 of 53 steering pushes held (7 with a stricter ceiling). Two candidate ideas, norm and clamp, didn't survive testing, and masks failed to localize changes. What did hold: benchmarked fairly against conditioningaverage on the same seeds, prompt mixing keeps the scene packed until it suddenly flips — a prompt is a switch, while steering inside activations behaves like a dial within its range. Methodology in one line: steer what the model knows, train what it doesn't. Full numbers including misses are in the repo; the tool is on the Comfy Registry.
More from Research
- EMNLP 2026 main conference accepts just 78 of 1,158 short papers, a 6.7% rate — windx0303 · 2026-10-29
- DeepMind paper: AI displaces jobs gradually within firms, suddenly when AI firms win — paulnovosad · 2026-10-29
- Evo 2 Team Credits a Researcher Spanning RL Post-Training, Inference Optimization, and Neuroscience — BrianHie · 2026-10-29
- Evo 2 Shows Emergent In-Context Learning: Few-Shot Prompts Steer Design of Enhancers, Antibodies, Phage Genomes — BrianHie · 2026-10-29
- Dev Open-Sources a 380M Data-Curation Decision Model That Filters 1M Web Pages for $5 — vanstriendaniel · 2026-10-29
- "$10M+ compute on Navier-Stokes, $100M+ on other unpublished discoveries" — ilkamoi · 2026-10-29