ResearchOctober 30, 2026
-58
🧊
Steering quietly rewrites your text
Researchers show that a second steering vector can reverse about 78 percent of unintended discourse shifts in language models.
#Activation Steering#LLM#Interpretability#Discourse#arXiv

🔥 What happened
Researchers at the University of Copenhagen show that activation steering doesn't just nudge the behavior you target — it silently shifts formality, evidence use, and ambiguity in the output. Their fix: a second, complementary vector that restores 78% of the unintended discourse drift.
💡 Why it matters
Until now, side effects were measured with coarse proxies like perplexity. That one steering vector bends multiple discourse dimensions at once explains why safety steering often "works" yet reads oddly. The corrective vector keeps 93% of the intended behavior intact.
⚡ Our take
Anyone shipping steering vectors in production is flying blind at the discourse level. This paper offers the first usable correction — and exposes how immature the toolchain still is.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions — when in doubt, read the linked original source.
Deep Dives & Similar Intelligence

GPT-6 Reads Neural Nets Like Plain Text
ResearchNovember 02, 2026

GPT-6 Astra rewrites other models' weights
ResearchNovember 04, 2026

Looping Only Pays Off for Tiny Models
ResearchOctober 30, 2026

Claude audits neural nets by weight
ResearchNovember 03, 2026

AI Agents Run Lab, Reject Own Results
ResearchOctober 31, 2026