To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Models

2025/10/15 by Hedström, Anna, Salim I. Amoukou, Amoukou, Salim I. +6 · 3 citations
Computer Science · #Artificial Intelligence (cs.AI) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Natural Language Processing Techniques #Text Readability and Simplification #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2510.13290

openalex publication_date 2025/10/15 · openalex created_date 2025/10/17 · openalex updated_date 2026/07/28

Abstract

We introduce Mechanistic Error Reduction with Abstention (MERA), a principled framework for steering language models (LMs) to mitigate errors through selective, adaptive interventions. Unlike existing methods that rely on fixed, manually tuned steering strengths, often resulting in under or oversteering, MERA addresses these limitations by (i) optimising the intervention direction, and (ii) calibrating when, and how much to steer, thereby provably improving performance or abstaining when no confident correction is possible. Experiments across diverse datasets, and LM families demonstrate safe, effective, non-degrading error correction, and that MERA outperforms existing baselines. Moreover, MERA can be applied on top of existing steering techniques to further enhance their performance, establishing it as a general-purpose, and efficient approach to mechanistic activation steering.

Citations

• Are Sparse Autoencoders Useful? A Case Study in Sparse Probing
• AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
• Can sparse autoencoders be used to decompose and interpret steering vectors?
• Towards Reliable Evaluation of Behavior Steering Interventions in LLMs
• Steering Knowledge Selection Behaviours in LLMs via SAE-Based Representation Engineering
• Steering Large Language Models using Conceptors: Improving Addition-Based Activation Engineering
• Towards Inference-time Category-wise Safety Steering for Large Language Models
• Larger and more instructable language models become less reliable
• Programming Refusal with Conditional Activation Steering
• The Llama 3 Herd of Models
• Gemma 2: Improving Open Language Models at a Practical Size
• Planning in a recurrent neural network that plays Sokoban
• Analyzing the Generalization and Reliability of Steering Vectors
• External Invariants: A Cryptographic Trust Architecture for Institutional AI Inference
• LoFiT: Localized Fine-tuning on LLM Representations
• Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization
• LiSeCo: Linear Semantic Control for Language Generation
• Spectral Editing of Activations for Large Language Model Alignment
• ReFT: Representation Finetuning for Language Models
• A Language Model's Guide Through Latent Space
• Challenges with unsupervised LLM knowledge discovery
• Steering Llama 2 via Contrastive Activation Addition
• Improving Activation Steering in Language Models with Mean-Centring
• Eliciting Latent Knowledge from Quirky Language Models
• The Linear Representation Hypothesis and the Geometry of Large Language Models
• Function Vectors in Large Language Models
• The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets
• Representation Engineering: A Top-Down Approach to AI Transparency
• Inference-Time Intervention: Eliciting Truthful Answers from a Language Model
• LEACE: Perfect linear concept erasure in closed form
• Toy Models of Superposition
• Large Language Models are Zero-Shot Reasoners
• Do Prompt-Based Models Really Understand the Meaning of their Prompts?
• DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts
• Measuring Mathematical Problem Solving With the MATH Dataset
• Of Non-Linearity and Commutativity in BERT
• Measuring Massive Multitask Language Understanding

Cited by

• Towards Understanding Steering Strength
• In-Distribution Steering: Balancing Control and Coherence in Language Model Generation
• Aligning Recommendations with User Popularity Preferences
• NeuroFaith: Evaluating LLM Self-Explanation Faithfulness via Internal Representation Alignment
• Steered LLM Activations are Non-Surjective

Related

• Steering off Course: Reliability Challenges in Steering Language Models
• Shifting Perspectives: Steering Vectors for Robust Bias Mitigation in LLMs
• Steering When Necessary: Flexible Steering Large Language Models with Backtracking
• Painless Activation Steering: An Automated, Lightweight Approach for Post-Training Large Language Models
• Improving Activation Steering in Language Models with Mean-Centring