A training-free dynamic steering framework that modulates intervention strength per token based on local hidden state geometry, eliminating text looping and scaling LLM test-time compute.
LaTense: Mitigating Latent Collapse in Activation Steering via Geometric Alignment LaTense: Mitigating Latent Collapse in Activation Steering via Geometric Alignment Nathan Egbuna University of Pennsylvania [email protected] Preprint, 2026 Latent space steering is a promising training-free paradigm for LLM control, yet current static interventions often trigger catastrophic text looping and reasoning collapse. We propose LaTense ( Lat ent S ense ), a dynamic steering framework that modulates intervention strength based on the local geometric alignment
A training-free dynamic steering framework that modulates intervention strength per token based on local hidden state geometry, eliminating text looping and scaling LLM test-time compute.
LaTense: Mitigating Latent Collapse in Activation Steering via Geometric Alignment LaTense: Mitigating Latent Collapse in Activation Steering via Geometric Alignment Nathan Egbuna University of Pennsylvania [email protected] Preprint, 2026 Latent space steering is a promising training-free paradigm for LLM control, yet current static interventions often trigger catastrophic text looping and reasoning collapse. We propose LaTense ( Lat ent S ense ), a dynamic steering framework that modulates intervention strength based on the local geometric alignment
