MEAV: editing LLM alignment at inference time with composable alignment vectors
burkov · x · 2026-10-19
An article introduced by Andriy Burkov tackles a practical LLM alignment problem: training-time methods require retraining whenever preferences change, and prompt-based approaches get expensive at scale. Users need flexible, fine-grained control at inference time — especially in medicine, law, finance, and multi-domain settings with different audiences.
The proposed method, MEAV (Model Editing with Alignment Vectors), tests whether alignment vectors — simple numerical differences between a base model and a preference-aligned version — can represent preference dimensions, and whether several such vectors can be combined to produce diverse multi-domain behavior.
More from Research
- anselm defines digital twins as data-backed mirrors that predict real systems — anselm · 2026-10-19
- FloatDoor: backdoors that trigger LLM misbehavior only on specific GPUs and platforms — rickasaurus · 2026-10-19
- G. Rauch on new interpretability piece: simpler path may beat SOTA mech interp — yenkel · 2026-10-19
- Burkov's Hundred-Page Language Models Book details three architectures with PyTorch code — burkov · 2026-10-19
- FOMO-DEX from Amazon FAR, CMU and Berkeley uses force feedback to fix only contact commands — HaozhiQ · 2026-10-19
- Preprint finds virus-specific T cells persistently activated in 40% of Long COVID cases — alvelda · 2026-10-19