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

Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior

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.

Proposes INNSteer, which learns a per-layer RealNVP-style invertible network so that LLM activation steering becomes a mean-difference translation in latent space mapped back to an input-dependent update in activation space.

Problem

Existing activation steering methods apply a single global linear offset in activation space, which is inadequate when behavioral features lie on curved or anisotropic manifolds, so steering strength and consistency vary across inputs.

Approach

Learn a layer-specific invertible affine-coupling network with a Gaussian NLL, a directional class-mean separation loss, and a log-determinant regularizer; steer by translating along the latent mean difference and map back via the exact inverse, yielding nonlinear input-dependent edits.

Key results

• On LLaMA3-3B, INNSteer reaches 93.41% average alignment probability, within 3.64 points of LoRA fine-tuning and about 393x faster at inference.
• On refusal multiple-choice, INNSteer raises refusal-aligned probability from 69.06% (best baseline) to 84.48%.
• INNSteer improves over the strongest steering baseline by 15.9% on LLaMA and 13.42% on Qwen in average alignment probability while keeping perplexity largely stable.

Surprising findings

• Several baselines (e.g., PoD, PoE, ITI on some traits) reduce alignment probability below base on certain tasks, indicating steering can be inconsistent across behaviors.

Limitations

• Requires training a separate INN for each target behavior and intervention layer.
• Invertibility does not guarantee semantic preservation, fluency, or optimal controllability.
• Evaluation limited to next-token probability, single-turn settings and some open-ended judges.

Experimental setup

Datasets: Persona Dataset refusal benchmark hallucination benchmark POPE

Benchmarks: Persona alignment (six traits) refusal hallucination POPE

Metrics: alignment probability (Prob %) perplexity (PPL) LLM-judge score accuracy F1

Baselines: CAA RepE ITI LAcT HPR REC TruthFlow ODESteer PoD PoE TF LoRA

Builds on: RealNVP ActNorm Representation Engineering Contrastive Activation Addition ODESteer

Closest papers by meaning

• Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.90)
• Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors (ICLR 2025 · similarity 0.90)
• The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.90)
• Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.90)
• Steer2Edit: From Activation Steering to Component-Level Editing (NeurIPS 2026 · similarity 0.89)
• Controlling Language and Diffusion Models by Transporting Activations (ICLR 2025 · similarity 0.89)