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

Values as Style: Disentangling Values from Semantics with One-Way Mixing for Low-Damage LLM Steering

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 a frozen-LLM steering interface that splits residual-stream states into semantic and value codes via one-way semantic-to-value mixing, then edits only the value code by injecting a residual delta to reduce collateral semantic damage.

Problem

Activation steering for value or stance change often also shifts topic, phrasing, and task content, causing collateral semantic damage and benign refusals.

Approach

A lightweight dual-encoder with a stop-gradient one-way semantic-to-value mixing path, trained with swap consistency, adversarial topic suppression, and decorrelation, yields an editable value code; inference edits the value code with a residual-space delta (optionally null-space projected and gated) while holding the semantic code fixed.

Key results

• Full one-way + GatedNSI reaches Align 0.750 and SemSim 0.873 with FRR 0.043 on LLaMA-3.1-8B, versus LinearAdd Align 0.770, SemSim 0.792, FRR 0.119.
• Human preservation score 4.27 for the full interface vs 3.72 for LinearAdd on a blinded 320-item comparison.
• Leakage probes: Topic from z_v drops to 0.19 and Value from z_s to 0.21, versus 0.421/0.603 for random split.

Surprising findings

• Direct replacement edit (no-delta) raises alignment to 0.762 but sharply lowers semantic similarity to 0.778 and raises FRR to 0.123.
• GPT-4o text-only judge achieves higher ValueBench accuracy (0.914) than the learned z_v head (0.873), so the interface is not the best value reader.

Limitations

• Factorization is underconstrained; claims are operational rather than identifiable, and residual cross-factor information remains.
• Evaluated only at a single mid-layer and last-token site on 7-8B backbones, with limited longer-interaction testing.
• Human agreement is only moderate (Krippendorff alpha 0.55-0.68) and varies by criterion.

Experimental setup

Datasets: SVQ 630 quadruples SVQ-EQ-10K ValueBench

Benchmarks: ValueBench SVQ-Test MMLU

Metrics: accuracy macro-F1 Align SemSim FRR ΔPPL ΔMMLU BERTScore NLI contradiction entity recall

Baselines: LinearAdd RepE NSI GatedNSI SWAI SAE-steering YaPO Circuit-breaker rerouting Prompt P2 GPT-4o ValueLlama-3-8B Kaleido

Builds on: RepE Activation addition MUNIT/DRIT/Swapping Autoencoder gradient reversal Barlow Twins-style decorrelation

Closest papers by meaning

• Search for Truth from Reasoning: A Dynamic Representation Editing Framework for Steering LLM Trajectories (ICML 2026 · similarity 0.84)
• Language Models can Self-Improve at State-Value Estimation for Better Search (NeurIPS 2025 · similarity 0.84)
• Steering Information Utility in Key-Value Memory for Language Model Post-Training (NeurIPS 2025 · similarity 0.84)
• Steer2Edit: From Activation Steering to Component-Level Editing (NeurIPS 2026 · similarity 0.84)
• When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.83)
• Don't Lose Focus: Activation Steering via Key-Orthogonal Projections (NeurIPS 2026 · similarity 0.83)