analiza aiqbench
This paper is squarely about AI: it studies activation steering of large language models using sparse autoencoders, asking whether steering every token is necessary. The proposed Stochastic Token Steering and Stochastic Block Steering methods show that intervening on only a fraction of tokens can preserve much of the behavioral effect while reducing fluency loss, a practical result for anyone building controllable LLMs. The findings are limited to two model families and two behavioral tasks, so readers should treat the trade-off between intervention ratio and steering magnitude as a promising lead rather than a settled rule.
A Coin Flip Per Token: Bernoulli Sparse Steering of Large Language Models
arXiv:2607.05615v2 Announce Type: replace Abstract: Activation steering via sparse autoencoders (SAEs) enables behavioral control of large language models without task-specific fine-tuning, but standard methods apply the steering signal at every generated token, incurring constant per-token perturbation that risks degrading fluency. We ask: is dense intervention necessary? We introduce Stochastic Token Steering (STS), which gates each token independently with probability $p$, and Stochastic Block Steering (SBS), which gates a leading window once per sequence; neither requires a reward model or learned gating policy. Across two model families and two behavioral tasks, steering only a fraction of tokens recovers most of the behavioral effect of dense steering while incurring less fluency degradation. Across our experimental settings, we further find that the best-performing steering magnitude consistently rises as the intervention ratio falls.
Abstract:Activation steering via sparse autoencoders (SAEs) enables behavioral control of large language models without task-specific fine-tuning, but standard methods apply the steering signal at every generated token, incurring constant per-token perturbation that risks degrading fluency. We ask: is dense intervention necessary? We introduce Stochastic Token Steering (STS), which gates each token independently with probability $p$, and Stochastic Block Steering (SBS), which gates a leading window once per sequence; neither requires a reward model or learned gating policy. Across two model families and two behavioral tasks, steering only a fraction of tokens recovers most of the behavioral effect of dense steering while incurring less fluency degradation. Across our experimental settings, we further find that the best-performing steering magnitude consistently rises as the intervention ratio falls.
Machine Learning (cs.LG)
arXiv:2607.05615 [cs.LG]
(or arXiv:2607.05615v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2607.05615 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 216 cuvinte
Subiecte conexe
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · 8 oct. 2026
Open Distillation of Hereditary Traits
AI Alignment Forum · 14 iul. 2026
Heads, Tails, and AI Fails: LLMs, Randomness, and Human Judgments
arXiv · 6 oct. 2026
LangFIR: Discovering Sparse Language-Specific Features from Monolingual Data for Language Steering
arXiv · 5 aug. 2026