Local linearity of LLM layers enables LQR-based closed-loop activation steering with theoretical tracking guarantees.
International Conference on Learning Representations , year=
5 Pith papers cite this work. Polarity classification is still indexing.
5
Pith papers citing it
representative citing papers
Prefix-tuning matches or exceeds fine-tuning on NLG tasks by optimizing a continuous prefix using 0.1% of parameters while keeping the LM frozen.
Maximizing joint conditional mutual information I(Y; C_Z, W, Z | X) decomposes multi-objective LLM alignment into preference-specific DPO terms plus an I(Y;W|X) exploration term that reduces reward-distribution overlap.
Activation steering is cast as constrained optimization that minimizes collateral damage by weighting perturbations according to the empirical second-moment matrix of activations instead of assuming isotropy.
citing papers explorer
-
Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control
Local linearity of LLM layers enables LQR-based closed-loop activation steering with theoretical tracking guarantees.
-
Prefix-Tuning: Optimizing Continuous Prompts for Generation
Prefix-tuning matches or exceeds fine-tuning on NLG tasks by optimizing a continuous prefix using 0.1% of parameters while keeping the LM frozen.
-
Multi-Objective Exploration and Preference Optimization via Mutual Information
Maximizing joint conditional mutual information I(Y; C_Z, W, Z | X) decomposes multi-objective LLM alignment into preference-specific DPO terms plus an I(Y;W|X) exploration term that reduces reward-distribution overlap.
-
Minimizing Collateral Damage in Activation Steering
Activation steering is cast as constrained optimization that minimizes collateral damage by weighting perturbations according to the empirical second-moment matrix of activations instead of assuming isotropy.
- Lost in the Tower of Babel: The Adverse Effects of Incidental Multilingualism in LLMs