Research map / Large language models / Language, safety and interpretability

Inverted Detection and Control in Steering Vectors

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Identifies inverted-steering vectors (ISVs) whose positive steering suppresses the concept despite high discriminability, proposes a forward-pass "representation response" to detect them, and uses it to flip sign in ITI, improving 27/30 experiments.

Problem

Steering vectors are assumed to promote a concept when steered along the direction that discriminates concept-positive representations; this paper shows highly discriminative vectors can instead consistently induce the opposite behavior, and that sign selection is needed for detection-based steering

Approach

Discriminative steering vectors aligned with positive examples can be ISVs whose effect is inverted in aggregate; a geometric measure of how steering shifts inner products in downstream discriminative heads (the representation response / spoof score) predicts whether to flip the steering sign, computed from forward passes without generation.

Key results

• ITI-RRF improves over ITI in 27/30 model-concept experiments, with gains from +0.9% to +138% (up to corrigibility in Gemma).
• Spoof score separates ISVs from RSVs with AUC 0.91 (accuracy 81% at threshold 0) over 90 mined vectors; AUC rises to 0.99 under a strict threshold on 36 vectors.
• Mined ISVs and RSVs across 15 model-concept pairs (45 each) in the multiple-choice setting, with many ISVs showing strongly negative monotonicity despite AUC above 0.9.

Surprising findings

• Some highly discriminative steering vectors (AUC around 0.97) aligned with concept-positive examples consistently induce the opposite behavior when steered positively.
• Under ITI-RRF, Qwen improved in only 7 of 10 experiments, so the sign-flip gain is not uniform.

Limitations

• Evaluated only translation-based (mean-difference) steering vectors on attention heads; affine and non-linear interventions not studied.
• Qwen showed no improvement in 3 of 10 ITI-RRF experiments, left uninvestigated.
• Open-ended ISV mining relies on a conservative alpha range and small candidate sets, so degradation thresholds were not determined per vector.

Experimental setup

Datasets: Model Written Evaluations (corrigibility, wealth-seeking, myopic reward) refusal dataset from Contrastive Activation Addition TruthfulQA

Benchmarks: Model Written Evaluations (corrigibility, wealth-seeking, myopia) refusal TruthfulQA

Metrics: AUC LLM-judge concept score (1-4 or 0-1) degradation-adjusted concept score Spearman monotonicity AUC of spoof score for ISV vs RSV classification

Baselines: ITI

Builds on: Inference Time Intervention linear representation hypothesis contrastive activation addition mean-difference steering

Closest papers by meaning

• Analysing the Generalisation and Reliability of Steering Vectors (NeurIPS 2024 · similarity 0.88)
• The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.87)
• DISCO: Disentangled Communication Steering for Large Language Models (NeurIPS 2025 · similarity 0.86)
• Iterative Vectors: In-Context Gradient Steering without Backpropagation (ICML 2025 · similarity 0.86)
• When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.86)
• Steer2Edit: From Activation Steering to Component-Level Editing (NeurIPS 2026 · similarity 0.85)