← News

ResearchSources: arXiv · 16 Nov 2026

Researchers propose budget-based token selection for activation steering in language models

An arXiv paper argues that steering language model behavior works better when interventions target only some tokens under a fixed budget, rather than every token a model generates.

In plain words

Activation steering is a way to guide how a language model behaves without retraining it. The authors say applying this guidance at every single word can push the model off its trained patterns and hurt output quality. They propose choosing which words to steer, spending a limited steering budget where it helps most.

What was announced

The paper, posted as a cross-listed arXiv abstract, frames steering as an allocation problem: a fixed budget of steerable tokens is set, and methods differ only in which tokens they pick. Within that setup the authors introduce two gates. Contrastive Feature Projection (CFP) is memoryless and selects tokens least aligned with the target behavior using a sparse, interpretable feature projection. Closed-Loop Adaptive Bernoulli Steering (CLABS) is trajectory aware, treating allocation as feedback control that tracks how much alignment a generation has accumulated and spends budget where the trajectory falls behind.

The authors report that on a linear steering-response model they derive an explicit gain region guaranteeing closed-loop stability of CLABS, and that its integral term reduces the steady-state tracking offset. Empirically, they state that CFP and CLABS produced the largest behavioral shift among budgeted gates at matched budgets, with CLABS strongest under tight budgets while best preserving fluency. These results are the authors' own claims from the abstract; the excerpt does not include benchmark names, model details, or independent evaluation.

What it does not show

  • The source is an abstract only; full experimental setup, benchmarks and model details are not available in the supplied excerpt.
  • Reported performance and fluency results are the authors' own claims and have not been independently confirmed.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.