steer2adapt

name	steer2adapt
title	Steer2Adapt: Dynamically Composing Steering Vectors for Efficient LLM Adaptation
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2602.07276
keywords	["Activation Steering","Domain Adaptation","Bayesian Optimization","Semantic Subspace","Inference-Time Adaptation"]
description	Adapt LLMs efficiently by composing multiple pre-existing semantic steering vectors via Bayesian optimization, balancing adaptation gain and safety without retraining model parameters.

Steer2Adapt: Dynamically Composing Steering Vectors for Efficient LLM Adaptation

Problem Context

Existing activation steering approaches suffer from inflexibility. Single static directions cannot adapt across task variations; complex tasks requiring multiple coordinated capabilities cannot be captured by one vector; and vectors optimized for one task may harm performance on related tasks.

Core Concept

Steer2Adapt shifts steering from discovering individual task-specific directions to dynamically composing multiple reusable semantic vectors. Rather than learning new steering vectors from scratch for each task, the method finds optimal linear combinations of pre-existing concept vectors within a domain-specific subspace.

Architecture Overview

• Semantic Subspace Construction: Identify k behavioral concepts and extract corresponding steering vectors using representation engineering
• Composed Vector Search: Use Bayesian Optimization to find optimal coefficients balancing adaptation gain and safety
• Inference-Time Application: Inject composite steering vector into model activations without parameter updates

Implementation

Phase 1: Semantic Subspace Construction

def construct_semantic_subspace(domain, concepts=None):
    # Define behavioral concepts for domain
    # (e.g., Big Five traits for reasoning)
    if concepts is None:
        concepts = ['Openness', 'Conscientiousness', 'Extroversion',
                    'Agreeableness', 'Neuroticism']

    # Extract steering vectors via representation engineering
    vectors = []

    for concept in concepts:
        # Generate positive/negative exemplars
        positive = generate_examples(concept, valence='positive')
        negative = generate_examples(concept, valence='negative')

        # Compute activations
        pos_acts = get_activations(model, positive, layer=target_layer)
        neg_acts = get_activations(model, negative, layer=target_layer)

        
        steering_vec = np.mean(pos_acts) - np.mean(neg_acts)
        vectors.append(steering_vec)

    
    V = np.column_stack(vectors)  

     V, concepts