Awesome Activation Steering

A curated collection of papers on activation steering and representation engineering in large language models.

Since January 2026, I've been diving into this field as a beginner, and keeping up with the papers has been no small feat. There may be errors along the way, but if you're new here, too, I hope it saves you some of the confusion I went through.

What is Activation Steering?

• A technique for guiding model behavior by directly modifying internal activation values at inference time—typically by intervening in the residual stream.
• The term "Activation Engineering" was coined in the paper Steering Language Models With Activation Engineering (arXiv, August 2023).
• The central challenge is finding a meaningful steering direction within the LLM's high-dimensional latent space that reliably produces the desired behavioral change.

🤔 Does it really work?

From my own experimentation, I'd lean toward yes. That said, measuring the effectiveness of steering remains a non-trivial challenge, and in some domains, it demonstrably underperforms both prompting and fine-tuning.

Before You Start

The industry is moving fast — Anthropic is actively publishing updates in this space. The following resources are highly recommended:

• Transformer Circuits — Anthropic's research blog on mechanistic interpretability. Essential reading for understanding how language models work at a circuit level.
• Activation Engineering posts on LessWrong — A curated collection of posts exploring how to steer model behavior by intervening on internal activations.

---

📑 Table of Contents

• Benchmarks
• Activation Steering
• Representation Engineering
• Persona & Role-Play in LLMs
• Linear Representations in LLMs
• Personality Modelling in LLMs

---

🏆 Benchmarks

• AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders ICML 2025 Spotlight
• Steer-Bench: A Benchmark for Evaluating the Steerability of LLMs 2025-01 arXiv
• AISteer360 2025-01
• MIB: A Mechanistic Interpretability Benchmark ICML 2025

🎯 Activation Steering

• ActAdd: Steering language models with activation engineering 2023-08 arXiv
• ITI: Inference-Time Intervention: Eliciting Truthful Answers from a Language Model NIPS 2023 Spotlight
• CAA: Steering Llama 2 via Contrastive Activation Addition ACL 2024 🏆 Outstanding Paper
• Analyzing the Generalization and Reliability of Steering Vectors 2024-07 arXiv
• Reliability Challenges in Steering Language Models 2025-04 arXiv
• SAEs Are Good for Steering — If You Select the Right Features EMNLP 2025
• Improved Representation Steering for Language Models NIPS 2025
• HyperSteer: Activation Steering at Scale with Hypernetworks 2025-06 arXiv
• Steerable Chatbots: Personalizing LLMs with Preference-Based Activation Steering 2025-05 arXiv
• Mitigating Overthinking in Large Reasoning Models via Manifold Steering 2025-05 arXiv
• In-Distribution Steering: Balancing Control and Coherence in Language Model Generation 2025-10 arXiv
• BILLY: Steering LLMs via Merging Persona Vectors for Creative Generation 2025-10 arXiv
• CogSteer: Cognition-Inspired Selective Layer Intervention for Efficiently Steering LLMs ACL 2025 Findings
• Toward universal steering and monitoring of AI models 2026 Science
• Weight Updates as Activation Shifts: A Principled Framework for Steering 2026-02 arXiv
• Attention Residuals 2026-03 arXiv
• Fine-Grained Activation Steering: Steering Less, Achieving More ICLR 2026
• Steer Like the LLM: Activation Steering that Mimics Prompting ICML 2026
• Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control ICML 2026

AI Safety

• Refusal in language models is mediated by a single direction 2024-06 arXiv
• Programming Refusal with Conditional Activation Steering ICLR 2025 Spotlight

Personality Steering

• Identifying and Manipulating Personality Traits in LLMs Through Activation Engineering 2024-12 arXiv
• Activation-Space Personality Steering: Hybrid Layer Selection for Stable Trait Control in LLMs 2025-10 arXiv
• PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra 2026 ICLR

Dynamic Activation Steering

• Programming Refusal with Conditional Activation Steering ICLR 2025 Spotlight
• ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment ICLR 2026

Reasoning

• Improving Reasoning in LLMs via Representation Engineering 2025-04 arXiv
• Steering LLMs' Reasoning with activation state machines 2025-09 OpenReview

---

🔬 Representation Engineering

• RepE: Representation Engineering: A Top-Down Approach to AI Transparency 2023-10 arXiv
• Looking inward: Language models can learn about themselves by introspection 2024-10 arXiv
• LLM evaluators recognize and factor their own generations 2024-04 arXiv
• Inspection and control of self-generated-text recognition ability in Llama3-8B-Instruct ICLR 2025
• Tell me about yourself: LLMs are aware of their learned behaviors ICLR 2025
• Simple mechanistic explanations for out-of-context reasoning ICML 2025 Workshop
• Emergent introspective awareness in large language models 2026-01 arXiv

---

🎭 Persona & Role-Play in Language Models

Anthropic

• Persona Vectors: Monitoring and Controlling Character Traits in Language Models 2025-07 arXiv
• The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models 2026-01 arXiv

Role-Play

• Role Play with large language models Nature 2023-11
• RoleLLM: Benchmarking, Eliciting, and Enhancing Role-Playing Abilities 2023-10 arXiv
• Character-LLM: A Trainable Agent for Role-Playing EMNLP 2023
• Measuring and Controlling Instruction (In)Stability in Language Model Dialogs COLM 2024
• From Persona to Personalization: A Survey on Role-Playing Language Agents TMLR 2024

---

📐 Linear Representations in Language Models

Anthropic Blog Posts

• Towards Monosemanticity: Decomposing Language Models with Dictionary Learning 2023-10
• Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet 2024-05
• Circuit Tracing: Revealing Computational Graphs in Language Models 2025-03
• On the Biology of a LLM 2025-03
• Emotion Concepts and their Function in a Large Language Model 2026-04

SAEs

• Sparse Autoencoders Find Highly Interpretable Features in LMs 2023-09 arXiv
• Transcoders Find Interpretable LLM Feature Circuits NIPS 2024
• Auditing Language Models for Hidden Objectives 2025-03 arXiv

---

🧠 Personality Modelling in LLMs

• PersonaLLM: Investigating the ability of LLMs to express personality traits NAACL 2024 Findings
• Personality Traits in LLMs 2025-07 arXiv
• Big5-Chat: Shaping LLM personalities through training on human-grounded data 2024-10 arXiv
• Personality Vector: Modulating Personality of LLMs by Model Merging EMNLP 2025

Benchmarks

• LaMP: When Large Language Models Meet Personalization ACL 2024
• LongLaMP: A Benchmark for Personalized Long-form Text Generation 2025-06 arXiv