Spherical Steering Overcomes Vector Addition Limits in LLMs
hanjie_chen · x · 2026-07-06
Researchers have introduced "Spherical Steering," questioning the sufficiency of relying solely on vector addition for current LLM activation steering. By incorporating a geometry-aware mechanism, they redesign activation control. Modeled from the perspective of spherical geometry, this approach aims to provide more precise model behavior modulation than traditional linear vector addition. The paper suggests that existing vector addition methods have geometric limitations, and modeling in spherical space may better capture the intrinsic structure of the activation manifold.
More from Models
- Kimi K3.1 Pro tops MathArena open models, closing frontier gap to about 6 months — teortaxesTex · 2026-10-26
- Ruff creator Charlie Marsh: frontier models find infinite bugs in any software — charliermarsh · 2026-10-26
- Dev testing one skill across Opus 5.5 and Sol 6.1 finds instruction following 'all over the map' — lucasmeijer · 2026-10-26
- supermemory unveils learner-1, claiming continual learning for any model — julianweisser · 2026-10-26
- Anthropic says Opus 5.5 cuts cost per task by ~40% vs Opus 5 — taufiqintech · 2026-10-26
- AI search APIs ranked: top 12 within 7 points, yet prices differ nearly 7x — FinanceYF5 · 2026-10-26