Showing 2026Show all
3 papers · 1 filter
cs.LG2026
Steering Interference Reflects the Model's Defaults, Not the Behavior Directions
Srikanth Malla, Chiho Choi, Joon Hee Choi
Activation steering promises modular control of language model behavior: a behavior such as politeness corresponds to a direction in a model's activations, and adding that directio…
cs.LG2026
The Geometry of Refusal: Why Post-Hoc Safety Is Fragile and Pretraining-Time Safety Persists
Srikanth Malla, Chiho Choi, Joon Hee Choi
Post-hoc safety training (RLHF, DPO) is the dominant way to align large language models, yet jailbreaks (Zou et al., 2023), fine-tuning attacks (Qi et al., 2024), and activation-sp…
cs.CL2026
ADEPT: Adaptive Dynamic Early-Exit Process for Transformers
Sangmin Yoo, Srikanth Malla, Chiho Choi +2
The inference of large language models imposes significant computational workloads, often requiring the processing of billions of parameters. Although early-exit strategies have pr…