Steering Qwen3-8B with the <|im_end|> vector cuts verbose replies dramatically

Sauers_ · x · 2026-11-19

A neat activation-steering demo: adding the unembedding vector of the end-of-turn token <|imend|> to layer 18 of Qwen3-8B's residual stream drastically curbs the model's filler output. Per the shared chart, an unsteered 100-token reply drops to 15 tokens at strength 1, and 29 of 30 replies end immediately at strength 4 (p=0.02), while random-direction steering has the opposite effect, inflating replies to a 200-token cap. A lightweight, tuning-free way to control response verbosity.

Original post →

More from coding & agent

coding & agent channel →